{
"measurement": {
"model.layers.0.parallel_decoder": {
"attn": [
{
"accuracy": 0.9303842345742803,
"total_bits": 569401344,
"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": 0.9347852895918646,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9427289143204689,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9635975023633555,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9667245029030662,
"total_bits": 844115968,
"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": 0.9669895519159342,
"total_bits": 845691904,
"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": 0.9768258622406345,
"total_bits": 1083191296,
"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": 0.9771570884377548,
"total_bits": 1085291520,
"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": 0.979096496820842,
"total_bits": 1095792640,
"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": 0.9803530230726066,
"total_bits": 1108393984,
"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": 0.9833227916571655,
"total_bits": 1114127360,
"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": 0.9843777033843493,
"total_bits": 1123055616,
"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": 0.9852898445372519,
"total_bits": 1179148288,
"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": 0.986554389476384,
"total_bits": 1195950080,
"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": 0.9918731623574307,
"total_bits": 1440755712,
"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": 0.9931551611148998,
"total_bits": 1464385536,
"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": 0.9938113806083014,
"total_bits": 1620062208,
"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": 0.9958307487326429,
"total_bits": 1779482624,
"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": 0.9982804666689941,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.9013213344702595,
"total_bits": 1238922592,
"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": 0.9044951154409271,
"total_bits": 1283618144,
"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": 0.9252728422809589,
"total_bits": 1433121792,
"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": 0.934573331455651,
"total_bits": 1608233984,
"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": 0.9559546972772008,
"total_bits": 1812774416,
"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": 0.9592476198449731,
"total_bits": 1862132736,
"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": 0.9700819502927756,
"total_bits": 2002191888,
"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": 0.9755036541211762,
"total_bits": 2289099280,
"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": 0.9781854922245992,
"total_bits": 2322850816,
"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": 0.9783917933604435,
"total_bits": 2356985360,
"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": 0.9808909041120818,
"total_bits": 2406343680,
"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": 0.9890399953889611,
"total_bits": 2901196304,
"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": 0.9903700370586624,
"total_bits": 2950554624,
"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": 0.9935116498535009,
"total_bits": 3360210448,
"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": 0.9944661014872652,
"total_bits": 3477813504,
"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": 0.9969439248885273,
"total_bits": 3790289152,
"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": 0.9977952129495526,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.789222569057816,
"total_bits": 569401344,
"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": 0.8031664890678305,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.839575784771066,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8937715848809794,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8977077932734239,
"total_bits": 844115968,
"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": 0.903477250745422,
"total_bits": 845691904,
"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": 0.9226165125636678,
"total_bits": 1083191296,
"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": 0.9292781598081714,
"total_bits": 1085291520,
"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": 0.9394573935944783,
"total_bits": 1095792640,
"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": 0.9418931975960732,
"total_bits": 1108393984,
"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": 0.9497488203801607,
"total_bits": 1114127360,
"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": 0.9559735923417305,
"total_bits": 1123055616,
"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": 0.9539593970893245,
"total_bits": 1179148288,
"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": 0.9607334929077249,
"total_bits": 1195950080,
"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": 0.9733124170452356,
"total_bits": 1440755712,
"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": 0.9797718209753695,
"total_bits": 1464385536,
"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": 0.978172190095249,
"total_bits": 1620062208,
"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": 0.9894466305788803,
"total_bits": 1779482624,
"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": 0.9942019592088304,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.9821041786160908,
"total_bits": 1238922592,
"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": 0.9836893297584826,
"total_bits": 1283618144,
"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": 0.9837581939554136,
"total_bits": 1433121792,
"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": 0.9838904618334613,
"total_bits": 1608233984,
"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": 0.9951160707339448,
"total_bits": 1812774416,
"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": 0.9958619444851616,
"total_bits": 1862132736,
"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": 0.9960689754776755,
"total_bits": 2002191888,
"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": 0.995537420777653,
"total_bits": 2289099280,
"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": 0.9974437449280614,
"total_bits": 2322850816,
"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": 0.9980259773928035,
"total_bits": 2356985360,
"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": 0.9972622429149373,
"total_bits": 2406343680,
"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": 0.9990591999268968,
"total_bits": 2901196304,
"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": 0.9989926667476539,
"total_bits": 2950554624,
"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": 0.9992290542163813,
"total_bits": 3360210448,
"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": 0.9992603220016463,
"total_bits": 3477813504,
"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": 0.999302385100715,
"total_bits": 3790289152,
"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": 0.9993915057918792,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7697779771528745,
"total_bits": 569401344,
"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": 0.77609941049626,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7998516998792949,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8681403598503062,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8823128072054762,
"total_bits": 844115968,
"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": 0.8860459343383187,
"total_bits": 845691904,
"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": 0.9202371288679148,
"total_bits": 1083191296,
"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": 0.925126322790196,
"total_bits": 1085291520,
"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": 0.9326278741814589,
"total_bits": 1095792640,
"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": 0.9351774112958657,
"total_bits": 1108393984,
"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": 0.9423959717938775,
"total_bits": 1114127360,
"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": 0.9466294710965533,
"total_bits": 1123055616,
"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": 0.9476316332032806,
"total_bits": 1179148288,
"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": 0.9524680923082327,
"total_bits": 1195950080,
"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": 0.9710500623638693,
"total_bits": 1440755712,
"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": 0.9759477490657255,
"total_bits": 1464385536,
"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": 0.9790856321587375,
"total_bits": 1620062208,
"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": 0.9882383543605867,
"total_bits": 1779482624,
"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": 0.9945326560471011,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.7673931992367694,
"total_bits": 1238922592,
"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": 0.7727242465081968,
"total_bits": 1283618144,
"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": 0.8172001132839605,
"total_bits": 1433121792,
"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": 0.8306846202988374,
"total_bits": 1608233984,
"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": 0.8859227429095068,
"total_bits": 1812774416,
"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": 0.894504975331457,
"total_bits": 1862132736,
"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": 0.9132850683833423,
"total_bits": 2002191888,
"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": 0.9412267692387104,
"total_bits": 2289099280,
"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": 0.946603602092517,
"total_bits": 2322850816,
"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": 0.9423464977819669,
"total_bits": 2356985360,
"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": 0.9493834037137658,
"total_bits": 2406343680,
"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": 0.9707334265112877,
"total_bits": 2901196304,
"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": 0.974831157903138,
"total_bits": 2950554624,
"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": 0.984319271813882,
"total_bits": 3360210448,
"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": 0.9851419845301854,
"total_bits": 3477813504,
"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": 0.9888988728880098,
"total_bits": 3790289152,
"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": 0.995813385767274,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7316183298826218,
"total_bits": 569401344,
"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": 0.7417925941316705,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7748563022989976,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8521358692332318,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8644231749992621,
"total_bits": 844115968,
"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": 0.8686524578615239,
"total_bits": 845691904,
"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": 0.9096135771588275,
"total_bits": 1083191296,
"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": 0.9150119086629466,
"total_bits": 1085291520,
"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": 0.9228952487832621,
"total_bits": 1095792640,
"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": 0.9256139621138573,
"total_bits": 1108393984,
"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": 0.9329717374945942,
"total_bits": 1114127360,
"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": 0.9380678846255729,
"total_bits": 1123055616,
"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": 0.9401963088465364,
"total_bits": 1179148288,
"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": 0.9459135891183427,
"total_bits": 1195950080,
"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": 0.9671060367438354,
"total_bits": 1440755712,
"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": 0.9728677619836832,
"total_bits": 1464385536,
"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": 0.9759551680793888,
"total_bits": 1620062208,
"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": 0.9868984410147134,
"total_bits": 1779482624,
"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": 0.9936109059490263,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.7084821650856419,
"total_bits": 1238922592,
"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": 0.7155458158568332,
"total_bits": 1283618144,
"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": 0.7616819802083468,
"total_bits": 1433121792,
"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": 0.7755139830865359,
"total_bits": 1608233984,
"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": 0.855720838041682,
"total_bits": 1812774416,
"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": 0.8666727009572481,
"total_bits": 1862132736,
"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": 0.8862912250976813,
"total_bits": 2002191888,
"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": 0.9260177282910598,
"total_bits": 2289099280,
"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": 0.9325160621420333,
"total_bits": 2322850816,
"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": 0.927121980802009,
"total_bits": 2356985360,
"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": 0.9359541286371256,
"total_bits": 2406343680,
"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": 0.9630529564854345,
"total_bits": 2901196304,
"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": 0.9681834888301397,
"total_bits": 2950554624,
"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": 0.9802628968303141,
"total_bits": 3360210448,
"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": 0.9812434518611745,
"total_bits": 3477813504,
"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": 0.9850834817380497,
"total_bits": 3790289152,
"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": 0.994818658665999,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.750744522402161,
"total_bits": 569401344,
"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": 0.7602384137479883,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7938525049309981,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8619606475296774,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8748256603353902,
"total_bits": 844115968,
"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": 0.8777512331542215,
"total_bits": 845691904,
"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": 0.9157319872787124,
"total_bits": 1083191296,
"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": 0.9193719005898425,
"total_bits": 1085291520,
"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": 0.9240473061800003,
"total_bits": 1095792640,
"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": 0.9264032946605432,
"total_bits": 1108393984,
"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": 0.9383166048087572,
"total_bits": 1114127360,
"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": 0.9413787791211354,
"total_bits": 1123055616,
"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": 0.9450969433313922,
"total_bits": 1179148288,
"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": 0.9484741928938188,
"total_bits": 1195950080,
"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": 0.9698161725935183,
"total_bits": 1440755712,
"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": 0.974512132080762,
"total_bits": 1464385536,
"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": 0.9777489860395068,
"total_bits": 1620062208,
"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": 0.9878527664726502,
"total_bits": 1779482624,
"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": 0.9941838240790132,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6892777697036141,
"total_bits": 1238922592,
"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": 0.6964539854150069,
"total_bits": 1283618144,
"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": 0.7464700686304193,
"total_bits": 1433121792,
"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": 0.7617328190489819,
"total_bits": 1608233984,
"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": 0.8452819091709036,
"total_bits": 1812774416,
"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": 0.857101209461689,
"total_bits": 1862132736,
"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": 0.878613113572723,
"total_bits": 2002191888,
"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": 0.9206322019821719,
"total_bits": 2289099280,
"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": 0.9277063855215123,
"total_bits": 2322850816,
"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": 0.9218036810818472,
"total_bits": 2356985360,
"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": 0.9313480297201558,
"total_bits": 2406343680,
"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": 0.9603720489693315,
"total_bits": 2901196304,
"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": 0.9659078275098613,
"total_bits": 2950554624,
"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": 0.9788482003894291,
"total_bits": 3360210448,
"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": 0.9798768551922158,
"total_bits": 3477813504,
"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": 0.9841200111826,
"total_bits": 3790289152,
"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": 0.9944652818437469,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7497934204967398,
"total_bits": 569401344,
"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": 0.7585005987631648,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7910812724577754,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8504390065607271,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8678363647900129,
"total_bits": 844115968,
"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": 0.873850089939017,
"total_bits": 845691904,
"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": 0.9035571652807688,
"total_bits": 1083191296,
"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": 0.9111967161297798,
"total_bits": 1085291520,
"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": 0.9169337671054036,
"total_bits": 1095792640,
"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": 0.9188358073956088,
"total_bits": 1108393984,
"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": 0.9359885016946417,
"total_bits": 1114127360,
"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": 0.9391960497749479,
"total_bits": 1123055616,
"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": 0.9421609789133072,
"total_bits": 1179148288,
"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": 0.9455530865019873,
"total_bits": 1195950080,
"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": 0.9672100725338647,
"total_bits": 1440755712,
"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": 0.9727046039739722,
"total_bits": 1464385536,
"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": 0.9742121076897571,
"total_bits": 1620062208,
"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": 0.9872502906149939,
"total_bits": 1779482624,
"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": 0.9932240618049706,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6777241559405076,
"total_bits": 1238922592,
"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": 0.6848174427684985,
"total_bits": 1283618144,
"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": 0.7405145780036324,
"total_bits": 1433121792,
"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": 0.7572572003853948,
"total_bits": 1608233984,
"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": 0.8387426067339747,
"total_bits": 1812774416,
"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": 0.8511045159477937,
"total_bits": 1862132736,
"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": 0.8749901652336121,
"total_bits": 2002191888,
"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": 0.9172578293241953,
"total_bits": 2289099280,
"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": 0.924734395585562,
"total_bits": 2322850816,
"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": 0.9184356304375749,
"total_bits": 2356985360,
"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": 0.9283909699634502,
"total_bits": 2406343680,
"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": 0.9586578425216048,
"total_bits": 2901196304,
"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": 0.9644293217478614,
"total_bits": 2950554624,
"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": 0.9779828962330755,
"total_bits": 3360210448,
"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": 0.979020337427133,
"total_bits": 3477813504,
"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": 0.9837888305712688,
"total_bits": 3790289152,
"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": 0.9942649806800642,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7517806150411305,
"total_bits": 569401344,
"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": 0.7599675498510662,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7896987160569743,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8557778898822634,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.872961010195707,
"total_bits": 844115968,
"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": 0.8768085355037137,
"total_bits": 845691904,
"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": 0.9134991306223368,
"total_bits": 1083191296,
"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": 0.9182219266107208,
"total_bits": 1085291520,
"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": 0.9242247252872116,
"total_bits": 1095792640,
"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": 0.9271961685858274,
"total_bits": 1108393984,
"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": 0.9372414530892121,
"total_bits": 1114127360,
"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": 0.9407453931083805,
"total_bits": 1123055616,
"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": 0.944403310747523,
"total_bits": 1179148288,
"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": 0.9481906722250738,
"total_bits": 1195950080,
"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": 0.9693588743868627,
"total_bits": 1440755712,
"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": 0.9741137648295415,
"total_bits": 1464385536,
"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": 0.9771067703044728,
"total_bits": 1620062208,
"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": 0.9866838542450416,
"total_bits": 1779482624,
"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": 0.994027871466977,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6719344822983992,
"total_bits": 1238922592,
"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": 0.6792134391634088,
"total_bits": 1283618144,
"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": 0.7360603621131496,
"total_bits": 1433121792,
"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": 0.7529835850000381,
"total_bits": 1608233984,
"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": 0.8346382991263741,
"total_bits": 1812774416,
"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": 0.8474542278992503,
"total_bits": 1862132736,
"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": 0.8716175587553727,
"total_bits": 2002191888,
"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": 0.9151293572626615,
"total_bits": 2289099280,
"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": 0.9226715650997663,
"total_bits": 2322850816,
"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": 0.91622065909599,
"total_bits": 2356985360,
"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": 0.9264723689932572,
"total_bits": 2406343680,
"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": 0.9574970744158092,
"total_bits": 2901196304,
"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": 0.9634383873719918,
"total_bits": 2950554624,
"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": 0.9773437686656651,
"total_bits": 3360210448,
"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": 0.9784100540962658,
"total_bits": 3477813504,
"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": 0.9832555247764838,
"total_bits": 3790289152,
"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": 0.9940952051087821,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7471208556702262,
"total_bits": 569401344,
"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": 0.7538793588939466,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7887233346700668,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8443632286630178,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8717821412964871,
"total_bits": 844115968,
"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": 0.8757424064372715,
"total_bits": 845691904,
"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": 0.911762542042293,
"total_bits": 1083191296,
"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": 0.916518220187802,
"total_bits": 1085291520,
"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": 0.9254803247749805,
"total_bits": 1095792640,
"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": 0.9284224918014125,
"total_bits": 1108393984,
"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": 0.936424653388952,
"total_bits": 1114127360,
"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": 0.9416626427126558,
"total_bits": 1123055616,
"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": 0.9425656297489217,
"total_bits": 1179148288,
"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": 0.9483996972833809,
"total_bits": 1195950080,
"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": 0.9686941024112075,
"total_bits": 1440755712,
"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": 0.9739087387723359,
"total_bits": 1464385536,
"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": 0.976624695584178,
"total_bits": 1620062208,
"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": 0.9875646461977771,
"total_bits": 1779482624,
"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": 0.9937027993268872,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.663761389882941,
"total_bits": 1238922592,
"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": 0.6711357267279374,
"total_bits": 1283618144,
"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": 0.7292955094262173,
"total_bits": 1433121792,
"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": 0.7465345365436453,
"total_bits": 1608233984,
"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": 0.8308745606949455,
"total_bits": 1812774416,
"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": 0.8439458961549559,
"total_bits": 1862132736,
"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": 0.8686464664183164,
"total_bits": 2002191888,
"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": 0.913304333231951,
"total_bits": 2289099280,
"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": 0.9209021949454358,
"total_bits": 2322850816,
"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": 0.9143190101573342,
"total_bits": 2356985360,
"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": 0.9247986803713598,
"total_bits": 2406343680,
"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": 0.9565238787939674,
"total_bits": 2901196304,
"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": 0.9626098657517057,
"total_bits": 2950554624,
"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": 0.9768191052502707,
"total_bits": 3360210448,
"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": 0.9778832708926577,
"total_bits": 3477813504,
"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": 0.9828270857075327,
"total_bits": 3790289152,
"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": 0.9939104075494566,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7416311306388754,
"total_bits": 569401344,
"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": 0.7504795930887523,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7823857312139711,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8435015176471911,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8702234335635838,
"total_bits": 844115968,
"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": 0.8728399802195399,
"total_bits": 845691904,
"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": 0.914300709570709,
"total_bits": 1083191296,
"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": 0.9173027049554021,
"total_bits": 1085291520,
"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": 0.9239216548831839,
"total_bits": 1095792640,
"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": 0.9271637730692562,
"total_bits": 1108393984,
"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": 0.9358892599610906,
"total_bits": 1114127360,
"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": 0.93980039773803,
"total_bits": 1123055616,
"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": 0.9435502651490664,
"total_bits": 1179148288,
"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": 0.9478243618811432,
"total_bits": 1195950080,
"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": 0.9690238544227261,
"total_bits": 1440755712,
"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": 0.9740064350986167,
"total_bits": 1464385536,
"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": 0.9777446134310019,
"total_bits": 1620062208,
"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": 0.9867200987707627,
"total_bits": 1779482624,
"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": 0.9940355235563689,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6608438491821289,
"total_bits": 1238922592,
"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": 0.6682142411407672,
"total_bits": 1283618144,
"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": 0.727003477905926,
"total_bits": 1433121792,
"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": 0.7443362874420065,
"total_bits": 1608233984,
"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": 0.828904147210874,
"total_bits": 1812774416,
"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": 0.8422447380266691,
"total_bits": 1862132736,
"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": 0.8671176229652605,
"total_bits": 2002191888,
"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": 0.9121512315775219,
"total_bits": 2289099280,
"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": 0.9199591173153174,
"total_bits": 2322850816,
"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": 0.9132788718531006,
"total_bits": 2356985360,
"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": 0.9239349604437226,
"total_bits": 2406343680,
"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": 0.9560167956116953,
"total_bits": 2901196304,
"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": 0.9621855336192408,
"total_bits": 2950554624,
"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": 0.9765427119441723,
"total_bits": 3360210448,
"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": 0.9776457125615132,
"total_bits": 3477813504,
"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": 0.9826568366077385,
"total_bits": 3790289152,
"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": 0.9938950193800816,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7074635875852484,
"total_bits": 569401344,
"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": 0.7152771259609021,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.749971833668257,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8143850627698397,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8499537039744227,
"total_bits": 844115968,
"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": 0.8548754085051387,
"total_bits": 845691904,
"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": 0.8985369919162047,
"total_bits": 1083191296,
"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": 0.9050459787249565,
"total_bits": 1085291520,
"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": 0.9102737468324209,
"total_bits": 1095792640,
"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": 0.9131755550440989,
"total_bits": 1108393984,
"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": 0.9262886151279274,
"total_bits": 1114127360,
"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": 0.9306571701247441,
"total_bits": 1123055616,
"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": 0.9338722993668757,
"total_bits": 1179148288,
"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": 0.9387424231359833,
"total_bits": 1195950080,
"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": 0.9632627219942055,
"total_bits": 1440755712,
"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": 0.9693134530006271,
"total_bits": 1464385536,
"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": 0.9729687459766865,
"total_bits": 1620062208,
"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": 0.9852403771426332,
"total_bits": 1779482624,
"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": 0.9930287168214196,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.656695704711111,
"total_bits": 1238922592,
"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": 0.6643124426666058,
"total_bits": 1283618144,
"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": 0.7227729119752583,
"total_bits": 1433121792,
"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": 0.7398799391169297,
"total_bits": 1608233984,
"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": 0.8270297638679805,
"total_bits": 1812774416,
"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": 0.8405336925857946,
"total_bits": 1862132736,
"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": 0.864972778448933,
"total_bits": 2002191888,
"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": 0.9112836710716549,
"total_bits": 2289099280,
"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": 0.9189844092256144,
"total_bits": 2322850816,
"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": 0.9123104791107931,
"total_bits": 2356985360,
"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": 0.9230402413951724,
"total_bits": 2406343680,
"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": 0.9554938862198278,
"total_bits": 2901196304,
"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": 0.9617415505804514,
"total_bits": 2950554624,
"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": 0.9762329322923171,
"total_bits": 3360210448,
"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": 0.9773633854561731,
"total_bits": 3477813504,
"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": 0.9822656991646478,
"total_bits": 3790289152,
"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": 0.9938020746125594,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7227069233593189,
"total_bits": 569401344,
"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": 0.7300998231298046,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7641117086536006,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8288007888354754,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8573986809504659,
"total_bits": 844115968,
"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": 0.8620470264240315,
"total_bits": 845691904,
"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": 0.905801068403219,
"total_bits": 1083191296,
"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": 0.9121392545731444,
"total_bits": 1085291520,
"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": 0.9171513091576726,
"total_bits": 1095792640,
"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": 0.9205105069436526,
"total_bits": 1108393984,
"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": 0.930188994070417,
"total_bits": 1114127360,
"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": 0.9345270118823177,
"total_bits": 1123055616,
"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": 0.9381768162313261,
"total_bits": 1179148288,
"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": 0.9429381699546387,
"total_bits": 1195950080,
"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": 0.9661307816246623,
"total_bits": 1440755712,
"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": 0.9712771923330269,
"total_bits": 1464385536,
"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": 0.975770266726613,
"total_bits": 1620062208,
"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": 0.9859326277908526,
"total_bits": 1779482624,
"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": 0.9935938970577952,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6510948209386123,
"total_bits": 1238922592,
"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": 0.658737805328871,
"total_bits": 1283618144,
"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": 0.7187614817368357,
"total_bits": 1433121792,
"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": 0.7363917388414082,
"total_bits": 1608233984,
"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": 0.8237378785484716,
"total_bits": 1812774416,
"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": 0.8374799364491513,
"total_bits": 1862132736,
"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": 0.8627987939276194,
"total_bits": 2002191888,
"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": 0.909450214552252,
"total_bits": 2289099280,
"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": 0.9174262351895633,
"total_bits": 2322850816,
"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": 0.9105898428120112,
"total_bits": 2356985360,
"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": 0.9215701179284799,
"total_bits": 2406343680,
"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": 0.9546249022609309,
"total_bits": 2901196304,
"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": 0.9609983537934328,
"total_bits": 2950554624,
"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": 0.9757618493351498,
"total_bits": 3360210448,
"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": 0.9769380679844242,
"total_bits": 3477813504,
"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": 0.9819876943763933,
"total_bits": 3790289152,
"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": 0.9936715813206607,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7060202328782332,
"total_bits": 569401344,
"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": 0.7132947703725414,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.744671883551698,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8074673892636048,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8474180714080208,
"total_bits": 844115968,
"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": 0.8548874482512474,
"total_bits": 845691904,
"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": 0.8985690337262655,
"total_bits": 1083191296,
"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": 0.9093360273461593,
"total_bits": 1085291520,
"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": 0.9139499083945626,
"total_bits": 1095792640,
"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": 0.9179052334082753,
"total_bits": 1108393984,
"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": 0.9257848345135388,
"total_bits": 1114127360,
"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": 0.9299886352137515,
"total_bits": 1123055616,
"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": 0.9339433584950472,
"total_bits": 1179148288,
"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": 0.9386454042243331,
"total_bits": 1195950080,
"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": 0.9629155460156893,
"total_bits": 1440755712,
"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": 0.9693416191363021,
"total_bits": 1464385536,
"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": 0.9732992179495724,
"total_bits": 1620062208,
"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": 0.9849748714876018,
"total_bits": 1779482624,
"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": 0.9930461802225756,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6432125427220997,
"total_bits": 1238922592,
"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": 0.6510276574837535,
"total_bits": 1283618144,
"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": 0.7127047824232202,
"total_bits": 1433121792,
"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": 0.7307717117824053,
"total_bits": 1608233984,
"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": 0.8193714689267309,
"total_bits": 1812774416,
"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": 0.8335588009733903,
"total_bits": 1862132736,
"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": 0.8594872731911509,
"total_bits": 2002191888,
"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": 0.9071664214134216,
"total_bits": 2289099280,
"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": 0.9153562160698991,
"total_bits": 2322850816,
"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": 0.9083312616536492,
"total_bits": 2356985360,
"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": 0.9196271614024514,
"total_bits": 2406343680,
"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": 0.9534561387018153,
"total_bits": 2901196304,
"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": 0.9600295218590059,
"total_bits": 2950554624,
"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": 0.9751292898466712,
"total_bits": 3360210448,
"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": 0.9763408526778221,
"total_bits": 3477813504,
"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": 0.9815240120025057,
"total_bits": 3790289152,
"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": 0.9935281773489949,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7072872111671848,
"total_bits": 569401344,
"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": 0.7138793546902507,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7473777242397007,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8082528357443056,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8472136580630353,
"total_bits": 844115968,
"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": 0.8513950172223543,
"total_bits": 845691904,
"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": 0.8983162559176746,
"total_bits": 1083191296,
"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": 0.9036192874375143,
"total_bits": 1085291520,
"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": 0.9170046072257192,
"total_bits": 1095792640,
"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": 0.9202903709129283,
"total_bits": 1108393984,
"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": 0.9249783691607023,
"total_bits": 1114127360,
"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": 0.932255576511747,
"total_bits": 1123055616,
"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": 0.9330718599651989,
"total_bits": 1179148288,
"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": 0.9412018723393741,
"total_bits": 1195950080,
"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": 0.9627021194288605,
"total_bits": 1440755712,
"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": 0.970430516000641,
"total_bits": 1464385536,
"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": 0.9726726661779379,
"total_bits": 1620062208,
"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": 0.9852122907575808,
"total_bits": 1779482624,
"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": 0.9929606613801106,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6552215328342037,
"total_bits": 1238922592,
"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": 0.6627034639057361,
"total_bits": 1283618144,
"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": 0.7209535696004566,
"total_bits": 1433121792,
"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": 0.7379295700474789,
"total_bits": 1608233984,
"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": 0.8263275286084727,
"total_bits": 1812774416,
"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": 0.8406122891526473,
"total_bits": 1862132736,
"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": 0.8652797392324397,
"total_bits": 2002191888,
"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": 0.9105992297592916,
"total_bits": 2289099280,
"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": 0.9181609400793126,
"total_bits": 2322850816,
"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": 0.9119755076734644,
"total_bits": 2356985360,
"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": 0.9229371171248586,
"total_bits": 2406343680,
"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": 0.9552025018553985,
"total_bits": 2901196304,
"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": 0.9616615340897912,
"total_bits": 2950554624,
"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": 0.9758588204062298,
"total_bits": 3360210448,
"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": 0.9771833926635353,
"total_bits": 3477813504,
"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": 0.982159871313917,
"total_bits": 3790289152,
"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": 0.9935833553195392,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6806146745619022,
"total_bits": 569401344,
"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": 0.6879737424223047,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7270470104719463,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7910297905143938,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8331007549637243,
"total_bits": 844115968,
"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": 0.8376333399822837,
"total_bits": 845691904,
"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": 0.8862010518971243,
"total_bits": 1083191296,
"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": 0.891661849068968,
"total_bits": 1085291520,
"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": 0.9012984060927441,
"total_bits": 1095792640,
"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": 0.9043496588343068,
"total_bits": 1108393984,
"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": 0.9180443282974393,
"total_bits": 1114127360,
"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": 0.9231846936439213,
"total_bits": 1123055616,
"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": 0.9265394212776109,
"total_bits": 1179148288,
"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": 0.9323201018728708,
"total_bits": 1195950080,
"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": 0.9597017563095218,
"total_bits": 1440755712,
"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": 0.9660947713020601,
"total_bits": 1464385536,
"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": 0.9705451876905403,
"total_bits": 1620062208,
"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": 0.9838012118676775,
"total_bits": 1779482624,
"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": 0.9922212233374778,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6447967714384982,
"total_bits": 1238922592,
"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": 0.6528616124077847,
"total_bits": 1283618144,
"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": 0.7130956539982243,
"total_bits": 1433121792,
"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": 0.7304634326382687,
"total_bits": 1608233984,
"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": 0.8207565004888334,
"total_bits": 1812774416,
"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": 0.8347714135521337,
"total_bits": 1862132736,
"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": 0.8597444413523925,
"total_bits": 2002191888,
"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": 0.9080420822689408,
"total_bits": 2289099280,
"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": 0.9159125975872341,
"total_bits": 2322850816,
"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": 0.9090805610543803,
"total_bits": 2356985360,
"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": 0.9202261116159589,
"total_bits": 2406343680,
"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": 0.9538815884213698,
"total_bits": 2901196304,
"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": 0.9603352899614134,
"total_bits": 2950554624,
"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": 0.9753110627398679,
"total_bits": 3360210448,
"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": 0.9765373493887877,
"total_bits": 3477813504,
"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": 0.9814687274200352,
"total_bits": 3790289152,
"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": 0.9935327846507885,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.7003080601754941,
"total_bits": 569401344,
"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": 0.7072826396477849,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.742943431201734,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8057839635171389,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8443930996091742,
"total_bits": 844115968,
"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": 0.8500840373729405,
"total_bits": 845691904,
"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": 0.8981389971940141,
"total_bits": 1083191296,
"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": 0.9056164626228183,
"total_bits": 1085291520,
"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": 0.9150790744706204,
"total_bits": 1095792640,
"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": 0.9186334680569799,
"total_bits": 1108393984,
"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": 0.9238151137373949,
"total_bits": 1114127360,
"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": 0.9287569524818345,
"total_bits": 1123055616,
"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": 0.9328090726937118,
"total_bits": 1179148288,
"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": 0.9378622313471217,
"total_bits": 1195950080,
"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": 0.9625974938665566,
"total_bits": 1440755712,
"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": 0.9686672796348208,
"total_bits": 1464385536,
"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": 0.9732382333788433,
"total_bits": 1620062208,
"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": 0.9849048482352182,
"total_bits": 1779482624,
"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": 0.9930033576674759,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6420708173199704,
"total_bits": 1238922592,
"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": 0.6499877195609243,
"total_bits": 1283618144,
"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": 0.7105275659184707,
"total_bits": 1433121792,
"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": 0.728246343763251,
"total_bits": 1608233984,
"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": 0.8181356842580595,
"total_bits": 1812774416,
"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": 0.8326687514781952,
"total_bits": 1862132736,
"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": 0.8581630767960298,
"total_bits": 2002191888,
"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": 0.9063350570045019,
"total_bits": 2289099280,
"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": 0.9146260458387827,
"total_bits": 2322850816,
"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": 0.9074803214324147,
"total_bits": 2356985360,
"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": 0.9190164792694544,
"total_bits": 2406343680,
"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": 0.9529020286312229,
"total_bits": 2901196304,
"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": 0.9596539126022866,
"total_bits": 2950554624,
"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": 0.974712916796929,
"total_bits": 3360210448,
"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": 0.9760132530998242,
"total_bits": 3477813504,
"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": 0.9810281149846943,
"total_bits": 3790289152,
"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": 0.9933943502899063,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6811138171898692,
"total_bits": 569401344,
"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": 0.6883086350403333,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7257599697301262,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7876860648393631,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8335772674334676,
"total_bits": 844115968,
"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": 0.8390394864897979,
"total_bits": 845691904,
"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": 0.8892897681186074,
"total_bits": 1083191296,
"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": 0.8960545204187694,
"total_bits": 1085291520,
"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": 0.9047291761166171,
"total_bits": 1095792640,
"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": 0.9089425606162924,
"total_bits": 1108393984,
"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": 0.9188094084200106,
"total_bits": 1114127360,
"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": 0.9244818605090442,
"total_bits": 1123055616,
"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": 0.9275458847221575,
"total_bits": 1179148288,
"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": 0.9338835924863815,
"total_bits": 1195950080,
"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": 0.9598773210064361,
"total_bits": 1440755712,
"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": 0.9665802400559187,
"total_bits": 1464385536,
"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": 0.9709875243470857,
"total_bits": 1620062208,
"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": 0.9836511675660548,
"total_bits": 1779482624,
"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": 0.99247895279213,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.632301799560848,
"total_bits": 1238922592,
"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": 0.6405087144751298,
"total_bits": 1283618144,
"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": 0.7031259505372298,
"total_bits": 1433121792,
"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": 0.7213246665502849,
"total_bits": 1608233984,
"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": 0.8133322863202346,
"total_bits": 1812774416,
"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": 0.8281787539783277,
"total_bits": 1862132736,
"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": 0.8544925184626329,
"total_bits": 2002191888,
"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": 0.903876569318144,
"total_bits": 2289099280,
"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": 0.9123669733342371,
"total_bits": 2322850816,
"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": 0.9050076556833166,
"total_bits": 2356985360,
"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": 0.9168272324298558,
"total_bits": 2406343680,
"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": 0.9516522870644143,
"total_bits": 2901196304,
"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": 0.9585944484723241,
"total_bits": 2950554624,
"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": 0.9740770705240337,
"total_bits": 3360210448,
"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": 0.9753870389571315,
"total_bits": 3477813504,
"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": 0.980544550246314,
"total_bits": 3790289152,
"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": 0.9931884493286672,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6740728914737701,
"total_bits": 569401344,
"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": 0.6818644859288868,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7202856054431515,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7846320396975467,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8308817461917275,
"total_bits": 844115968,
"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": 0.8352970567188764,
"total_bits": 845691904,
"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": 0.8891382746790585,
"total_bits": 1083191296,
"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": 0.894704888525762,
"total_bits": 1085291520,
"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": 0.9045887546319711,
"total_bits": 1095792640,
"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": 0.9081605664993587,
"total_bits": 1108393984,
"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": 0.9160960685265691,
"total_bits": 1114127360,
"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": 0.9222573724232221,
"total_bits": 1123055616,
"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": 0.9255689884486952,
"total_bits": 1179148288,
"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": 0.9320815095776006,
"total_bits": 1195950080,
"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": 0.9590907090980756,
"total_bits": 1440755712,
"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": 0.9658540701983791,
"total_bits": 1464385536,
"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": 0.9706396481120273,
"total_bits": 1620062208,
"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": 0.9835002135093275,
"total_bits": 1779482624,
"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": 0.9923274813121871,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6352136511551707,
"total_bits": 1238922592,
"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": 0.6434144612989927,
"total_bits": 1283618144,
"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": 0.7055291455042989,
"total_bits": 1433121792,
"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": 0.7237034408669722,
"total_bits": 1608233984,
"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": 0.8146537056094721,
"total_bits": 1812774416,
"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": 0.8295542268376601,
"total_bits": 1862132736,
"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": 0.8556892220911226,
"total_bits": 2002191888,
"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": 0.9044638636865114,
"total_bits": 2289099280,
"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": 0.912989071325252,
"total_bits": 2322850816,
"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": 0.9057782359029117,
"total_bits": 2356985360,
"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": 0.9175144594750906,
"total_bits": 2406343680,
"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": 0.9520549862400481,
"total_bits": 2901196304,
"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": 0.958918897533103,
"total_bits": 2950554624,
"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": 0.9741988145795307,
"total_bits": 3360210448,
"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": 0.9755980267532562,
"total_bits": 3477813504,
"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": 0.9807273772005972,
"total_bits": 3790289152,
"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": 0.9932399698239016,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6857639033543437,
"total_bits": 569401344,
"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": 0.6937871387130337,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7302346684430775,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7945483060259568,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8363283080490012,
"total_bits": 844115968,
"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": 0.8413765257910678,
"total_bits": 845691904,
"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": 0.8947284292233617,
"total_bits": 1083191296,
"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": 0.9017202677695375,
"total_bits": 1085291520,
"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": 0.9095429374199164,
"total_bits": 1095792640,
"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": 0.9131462926927366,
"total_bits": 1108393984,
"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": 0.91952336226639,
"total_bits": 1114127360,
"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": 0.9249081392037242,
"total_bits": 1123055616,
"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": 0.9292840604719362,
"total_bits": 1179148288,
"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": 0.9351505003869534,
"total_bits": 1195950080,
"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": 0.961210424178525,
"total_bits": 1440755712,
"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": 0.9675535869441534,
"total_bits": 1464385536,
"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": 0.9726316470064615,
"total_bits": 1620062208,
"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": 0.9841352448259529,
"total_bits": 1779482624,
"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": 0.9928531630552913,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6330719480389042,
"total_bits": 1238922592,
"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": 0.6411938730039095,
"total_bits": 1283618144,
"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": 0.7039925110967535,
"total_bits": 1433121792,
"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": 0.722618590844305,
"total_bits": 1608233984,
"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": 0.8131686394151888,
"total_bits": 1812774416,
"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": 0.8281791970918053,
"total_bits": 1862132736,
"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": 0.8548546825584612,
"total_bits": 2002191888,
"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": 0.9034224470195017,
"total_bits": 2289099280,
"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": 0.9121887656418901,
"total_bits": 2322850816,
"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": 0.905037992878964,
"total_bits": 2356985360,
"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": 0.9168910737100401,
"total_bits": 2406343680,
"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": 0.9517030308121129,
"total_bits": 2901196304,
"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": 0.9586411085175841,
"total_bits": 2950554624,
"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": 0.9740187283605337,
"total_bits": 3360210448,
"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": 0.9754580556366005,
"total_bits": 3477813504,
"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": 0.9807724383120474,
"total_bits": 3790289152,
"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": 0.9931879994753552,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6605826035926217,
"total_bits": 569401344,
"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": 0.6684647469144118,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7087781492032503,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7780427054355019,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8249952126490443,
"total_bits": 844115968,
"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": 0.8292835519502038,
"total_bits": 845691904,
"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": 0.8878794692848858,
"total_bits": 1083191296,
"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": 0.8931704469417271,
"total_bits": 1085291520,
"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": 0.9015847656287646,
"total_bits": 1095792640,
"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": 0.9054694022787244,
"total_bits": 1108393984,
"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": 0.912583081737945,
"total_bits": 1114127360,
"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": 0.918895922601223,
"total_bits": 1123055616,
"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": 0.9226346859022191,
"total_bits": 1179148288,
"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": 0.929431530205827,
"total_bits": 1195950080,
"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": 0.9579528609388753,
"total_bits": 1440755712,
"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": 0.9646181413217595,
"total_bits": 1464385536,
"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": 0.9701696909768017,
"total_bits": 1620062208,
"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": 0.9828519392758608,
"total_bits": 1779482624,
"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": 0.992011710437701,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6356943516354812,
"total_bits": 1238922592,
"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": 0.6438819295481631,
"total_bits": 1283618144,
"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": 0.7044667209449567,
"total_bits": 1433121792,
"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": 0.7221561968326569,
"total_bits": 1608233984,
"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": 0.8149866822518801,
"total_bits": 1812774416,
"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": 0.8297491167720995,
"total_bits": 1862132736,
"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": 0.8551575172888606,
"total_bits": 2002191888,
"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": 0.9046587069567881,
"total_bits": 2289099280,
"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": 0.9129839739517162,
"total_bits": 2322850816,
"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": 0.9059153622702548,
"total_bits": 2356985360,
"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": 0.91757680083576,
"total_bits": 2406343680,
"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": 0.9521017896109506,
"total_bits": 2901196304,
"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": 0.9589341769092962,
"total_bits": 2950554624,
"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": 0.9742147052954686,
"total_bits": 3360210448,
"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": 0.975617475317497,
"total_bits": 3477813504,
"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": 0.9806046121214566,
"total_bits": 3790289152,
"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": 0.9931941750753475,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6687060578873283,
"total_bits": 569401344,
"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": 0.6766172569049032,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7178838527516315,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7831878975818032,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8288637305560865,
"total_bits": 844115968,
"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": 0.8330707761802172,
"total_bits": 845691904,
"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": 0.8883603634802919,
"total_bits": 1083191296,
"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": 0.8939662830610025,
"total_bits": 1085291520,
"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": 0.9036223213923604,
"total_bits": 1095792640,
"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": 0.9072916295967604,
"total_bits": 1108393984,
"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": 0.9152789429614419,
"total_bits": 1114127360,
"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": 0.9210865932859873,
"total_bits": 1123055616,
"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": 0.9248816825841603,
"total_bits": 1179148288,
"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": 0.931340366601944,
"total_bits": 1195950080,
"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": 0.9587014599850303,
"total_bits": 1440755712,
"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": 0.965499152282351,
"total_bits": 1464385536,
"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": 0.9702750129723235,
"total_bits": 1620062208,
"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": 0.9834323643559688,
"total_bits": 1779482624,
"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": 0.9923458706685587,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6253927836292668,
"total_bits": 1238922592,
"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": 0.6335805921178115,
"total_bits": 1283618144,
"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": 0.6952490257589441,
"total_bits": 1433121792,
"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": 0.7134909457282016,
"total_bits": 1608233984,
"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": 0.8093595363591847,
"total_bits": 1812774416,
"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": 0.8244935313337728,
"total_bits": 1862132736,
"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": 0.8505980356743461,
"total_bits": 2002191888,
"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": 0.9016914881373707,
"total_bits": 2289099280,
"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": 0.9103387700099694,
"total_bits": 2322850816,
"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": 0.903066858257118,
"total_bits": 2356985360,
"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": 0.9150399055920149,
"total_bits": 2406343680,
"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": 0.9506975857442931,
"total_bits": 2901196304,
"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": 0.9577082530840447,
"total_bits": 2950554624,
"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": 0.9735077438189795,
"total_bits": 3360210448,
"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": 0.9749186437969145,
"total_bits": 3477813504,
"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": 0.9800249696954301,
"total_bits": 3790289152,
"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": 0.9930696821722546,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6909299626162178,
"total_bits": 569401344,
"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": 0.6997758808888888,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7347843294080936,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7981345222184533,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.840781096565096,
"total_bits": 844115968,
"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": 0.8452074547347269,
"total_bits": 845691904,
"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": 0.9009124899381086,
"total_bits": 1083191296,
"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": 0.9070474477974992,
"total_bits": 1085291520,
"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": 0.9144198690590105,
"total_bits": 1095792640,
"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": 0.9180354617143932,
"total_bits": 1108393984,
"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": 0.9214114623242303,
"total_bits": 1114127360,
"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": 0.9267140375940424,
"total_bits": 1123055616,
"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": 0.9317224911953274,
"total_bits": 1179148288,
"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": 0.9375598087514702,
"total_bits": 1195950080,
"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": 0.9622433140481773,
"total_bits": 1440755712,
"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": 0.9685376372776533,
"total_bits": 1464385536,
"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": 0.9738368786087161,
"total_bits": 1620062208,
"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": 0.9843193868077115,
"total_bits": 1779482624,
"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": 0.9931909481896773,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6333271421884236,
"total_bits": 1238922592,
"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": 0.6412730389519742,
"total_bits": 1283618144,
"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": 0.7022424983350855,
"total_bits": 1433121792,
"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": 0.7207152004304684,
"total_bits": 1608233984,
"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": 0.8130538722402171,
"total_bits": 1812774416,
"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": 0.8279687503450796,
"total_bits": 1862132736,
"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": 0.8540579786426142,
"total_bits": 2002191888,
"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": 0.9032994826373301,
"total_bits": 2289099280,
"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": 0.9119360521435738,
"total_bits": 2322850816,
"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": 0.9049738963183603,
"total_bits": 2356985360,
"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": 0.9167398759409001,
"total_bits": 2406343680,
"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": 0.9516817035251542,
"total_bits": 2901196304,
"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": 0.9585407231199115,
"total_bits": 2950554624,
"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": 0.9740076675815018,
"total_bits": 3360210448,
"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": 0.9754507262259722,
"total_bits": 3477813504,
"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": 0.9806936494025745,
"total_bits": 3790289152,
"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": 0.9931612141234311,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6884443995199705,
"total_bits": 569401344,
"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": 0.6964752360394126,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7353918771994741,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7943536137279711,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8391749372607783,
"total_bits": 844115968,
"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": 0.8429979027101868,
"total_bits": 845691904,
"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": 0.8961482475462713,
"total_bits": 1083191296,
"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": 0.9011331208442387,
"total_bits": 1085291520,
"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": 0.9095505170131984,
"total_bits": 1095792640,
"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": 0.9133522075257803,
"total_bits": 1108393984,
"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": 0.9204240911885312,
"total_bits": 1114127360,
"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": 0.9259379312003914,
"total_bits": 1123055616,
"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": 0.9304760555295568,
"total_bits": 1179148288,
"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": 0.9362847648168865,
"total_bits": 1195950080,
"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": 0.9616782775050715,
"total_bits": 1440755712,
"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": 0.9676942638072529,
"total_bits": 1464385536,
"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": 0.9725180186919475,
"total_bits": 1620062208,
"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": 0.9840360555405679,
"total_bits": 1779482624,
"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": 0.9928764041366154,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.631011669573031,
"total_bits": 1238922592,
"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": 0.6389618547339189,
"total_bits": 1283618144,
"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": 0.6991312754781622,
"total_bits": 1433121792,
"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": 0.7173577060824947,
"total_bits": 1608233984,
"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": 0.8121385754723298,
"total_bits": 1812774416,
"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": 0.8267862820311597,
"total_bits": 1862132736,
"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": 0.8525993455397456,
"total_bits": 2002191888,
"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": 0.9030105048104337,
"total_bits": 2289099280,
"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": 0.9115159707634073,
"total_bits": 2322850816,
"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": 0.9046124972795185,
"total_bits": 2356985360,
"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": 0.916279384572255,
"total_bits": 2406343680,
"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": 0.951559920452143,
"total_bits": 2901196304,
"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": 0.9583408640403497,
"total_bits": 2950554624,
"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": 0.9739874866055814,
"total_bits": 3360210448,
"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": 0.9753877668592491,
"total_bits": 3477813504,
"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": 0.9804686074586291,
"total_bits": 3790289152,
"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": 0.9931347310101908,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6676833802147916,
"total_bits": 569401344,
"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": 0.676745175531036,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7148470808016627,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7809354553097173,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8285070551853431,
"total_bits": 844115968,
"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": 0.8330430937440771,
"total_bits": 845691904,
"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": 0.8903210892489082,
"total_bits": 1083191296,
"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": 0.896282826207186,
"total_bits": 1085291520,
"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": 0.9052165111428813,
"total_bits": 1095792640,
"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": 0.9091251006251887,
"total_bits": 1108393984,
"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": 0.915001416284787,
"total_bits": 1114127360,
"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": 0.9205849164802777,
"total_bits": 1123055616,
"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": 0.9250720035480826,
"total_bits": 1179148288,
"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": 0.9313997112606701,
"total_bits": 1195950080,
"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": 0.9586015628944886,
"total_bits": 1440755712,
"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": 0.9656259666540121,
"total_bits": 1464385536,
"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": 0.9705282088957334,
"total_bits": 1620062208,
"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": 0.9829829764111262,
"total_bits": 1779482624,
"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": 0.9923210731021276,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6239957448683287,
"total_bits": 1238922592,
"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": 0.6318945931760889,
"total_bits": 1283618144,
"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": 0.6926235337006419,
"total_bits": 1433121792,
"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": 0.711117684841156,
"total_bits": 1608233984,
"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": 0.8080241562504518,
"total_bits": 1812774416,
"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": 0.8231032145650763,
"total_bits": 1862132736,
"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": 0.8492321262234136,
"total_bits": 2002191888,
"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": 0.9008943356181446,
"total_bits": 2289099280,
"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": 0.9096420191620526,
"total_bits": 2322850816,
"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": 0.9025464673575602,
"total_bits": 2356985360,
"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": 0.9144492067004505,
"total_bits": 2406343680,
"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": 0.9505268913742743,
"total_bits": 2901196304,
"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": 0.9574269179842974,
"total_bits": 2950554624,
"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": 0.9734612871941767,
"total_bits": 3360210448,
"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": 0.9748696904433402,
"total_bits": 3477813504,
"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": 0.9800807203313238,
"total_bits": 3790289152,
"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": 0.9930063215035357,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6837779707030246,
"total_bits": 569401344,
"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": 0.6922926487107026,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7289278844469471,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7923072121645275,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.838725239822739,
"total_bits": 844115968,
"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": 0.8426956895150637,
"total_bits": 845691904,
"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": 0.8995934313065127,
"total_bits": 1083191296,
"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": 0.9049682581895276,
"total_bits": 1085291520,
"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": 0.9115315993365488,
"total_bits": 1095792640,
"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": 0.9157283682572215,
"total_bits": 1108393984,
"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": 0.9204361644622526,
"total_bits": 1114127360,
"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": 0.9251842477211827,
"total_bits": 1123055616,
"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": 0.9306290590841519,
"total_bits": 1179148288,
"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": 0.9358961372391174,
"total_bits": 1195950080,
"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": 0.9617053912462372,
"total_bits": 1440755712,
"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": 0.9679716915676468,
"total_bits": 1464385536,
"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": 0.9735290940458837,
"total_bits": 1620062208,
"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": 0.9839937681528298,
"total_bits": 1779482624,
"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": 0.9931117593261757,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6287749613586224,
"total_bits": 1238922592,
"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": 0.6366907138573497,
"total_bits": 1283618144,
"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": 0.6965988545041335,
"total_bits": 1433121792,
"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": 0.7152602601992457,
"total_bits": 1608233984,
"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": 0.8104101395920703,
"total_bits": 1812774416,
"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": 0.8253293272696043,
"total_bits": 1862132736,
"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": 0.8514814086650547,
"total_bits": 2002191888,
"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": 0.9018383026123047,
"total_bits": 2289099280,
"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": 0.9106119690757049,
"total_bits": 2322850816,
"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": 0.903747600160147,
"total_bits": 2356985360,
"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": 0.9155283429120716,
"total_bits": 2406343680,
"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": 0.9511166661977768,
"total_bits": 2901196304,
"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": 0.9579570342443491,
"total_bits": 2950554624,
"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": 0.9737114470059934,
"total_bits": 3360210448,
"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": 0.975175205716177,
"total_bits": 3477813504,
"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": 0.9804258202447703,
"total_bits": 3790289152,
"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": 0.9930326234990436,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6683254869360673,
"total_bits": 569401344,
"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": 0.6775426990107486,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7136455178260803,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.781587280725178,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8284177968376562,
"total_bits": 844115968,
"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": 0.8325494614086653,
"total_bits": 845691904,
"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": 0.8926087574739205,
"total_bits": 1083191296,
"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": 0.8978596004216295,
"total_bits": 1085291520,
"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": 0.9061939594777007,
"total_bits": 1095792640,
"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": 0.9103794846879809,
"total_bits": 1108393984,
"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": 0.915468929237441,
"total_bits": 1114127360,
"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": 0.9209753765087378,
"total_bits": 1123055616,
"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": 0.9259165876397961,
"total_bits": 1179148288,
"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": 0.9320812995888685,
"total_bits": 1195950080,
"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": 0.9592848535823195,
"total_bits": 1440755712,
"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": 0.9659319622932296,
"total_bits": 1464385536,
"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": 0.9719721873928058,
"total_bits": 1620062208,
"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": 0.9829154679747788,
"total_bits": 1779482624,
"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": 0.9926673491044264,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6319807767868042,
"total_bits": 1238922592,
"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": 0.6397669189854672,
"total_bits": 1283618144,
"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": 0.6990429382575185,
"total_bits": 1433121792,
"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": 0.717269241809845,
"total_bits": 1608233984,
"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": 0.8122761641678057,
"total_bits": 1812774416,
"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": 0.8269182286764446,
"total_bits": 1862132736,
"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": 0.8525834491378382,
"total_bits": 2002191888,
"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": 0.9029196226283124,
"total_bits": 2289099280,
"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": 0.9115662100283723,
"total_bits": 2322850816,
"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": 0.904673476752482,
"total_bits": 2356985360,
"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": 0.9163150971657351,
"total_bits": 2406343680,
"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": 0.9516066469644245,
"total_bits": 2901196304,
"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": 0.9583639122153583,
"total_bits": 2950554624,
"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": 0.9740282691230899,
"total_bits": 3360210448,
"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": 0.9754208308693609,
"total_bits": 3477813504,
"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": 0.9805457415549379,
"total_bits": 3790289152,
"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": 0.9931047715393728,
"total_bits": 4447746304,
"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.parallel_decoder": {
"attn": [
{
"accuracy": 0.6415771214585555,
"total_bits": 569401344,
"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": 0.6511312989812148,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.701411529591209,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7658656741443434,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8154507883285221,
"total_bits": 844115968,
"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": 0.8205945640802383,
"total_bits": 845691904,
"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": 0.8928601055553085,
"total_bits": 1083191296,
"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": 0.8985613792350418,
"total_bits": 1085291520,
"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": 0.9055977102957273,
"total_bits": 1095792640,
"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": 0.9098374635765427,
"total_bits": 1108393984,
"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": 0.9087679884151408,
"total_bits": 1114127360,
"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": 0.9146381314647825,
"total_bits": 1123055616,
"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": 0.9252244492894725,
"total_bits": 1179148288,
"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": 0.9308582381590417,
"total_bits": 1195950080,
"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": 0.9591090759556544,
"total_bits": 1440755712,
"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": 0.9654360782159002,
"total_bits": 1464385536,
"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": 0.9722806545660684,
"total_bits": 1620062208,
"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": 0.9830420052044486,
"total_bits": 1779482624,
"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": 0.992840751245814,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6276157890495502,
"total_bits": 1238922592,
"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": 0.6354428702279141,
"total_bits": 1283618144,
"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": 0.6941211160860563,
"total_bits": 1433121792,
"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": 0.7121970716275667,
"total_bits": 1608233984,
"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": 0.8093796584166979,
"total_bits": 1812774416,
"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": 0.8244250753992483,
"total_bits": 1862132736,
"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": 0.8499687941450822,
"total_bits": 2002191888,
"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": 0.901524184957931,
"total_bits": 2289099280,
"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": 0.9102779756251135,
"total_bits": 2322850816,
"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": 0.9030357328684706,
"total_bits": 2356985360,
"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": 0.9150669268871608,
"total_bits": 2406343680,
"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": 0.9505066342259708,
"total_bits": 2901196304,
"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": 0.9576571968040968,
"total_bits": 2950554624,
"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": 0.9733679164201021,
"total_bits": 3360210448,
"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": 0.9747011669372257,
"total_bits": 3477813504,
"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": 0.9796706891961788,
"total_bits": 3790289152,
"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": 0.9924808130599558,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.26.parallel_decoder": {
"attn": [
{
"accuracy": 0.6454185435646458,
"total_bits": 569401344,
"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": 0.6549653818732815,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7048700930256593,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7710686898545215,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8174705873978765,
"total_bits": 844115968,
"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": 0.8205993991149099,
"total_bits": 845691904,
"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": 0.8936765511569223,
"total_bits": 1083191296,
"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": 0.8982575284807306,
"total_bits": 1085291520,
"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": 0.9055535612921966,
"total_bits": 1095792640,
"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": 0.910133277115069,
"total_bits": 1108393984,
"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": 0.9092061605892683,
"total_bits": 1114127360,
"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": 0.9149004742503166,
"total_bits": 1123055616,
"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": 0.9255642971318019,
"total_bits": 1179148288,
"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": 0.9314739053186617,
"total_bits": 1195950080,
"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": 0.9592531868501714,
"total_bits": 1440755712,
"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": 0.9657937276520228,
"total_bits": 1464385536,
"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": 0.9721996289138731,
"total_bits": 1620062208,
"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": 0.9829628307764467,
"total_bits": 1779482624,
"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": 0.9928980585873911,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6146051240594763,
"total_bits": 1238922592,
"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": 0.6226195410678261,
"total_bits": 1283618144,
"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": 0.6825941722644002,
"total_bits": 1433121792,
"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": 0.7008350256242251,
"total_bits": 1608233984,
"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": 0.8025667996783006,
"total_bits": 1812774416,
"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": 0.8179802353444853,
"total_bits": 1862132736,
"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": 0.8440247311403877,
"total_bits": 2002191888,
"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": 0.8981652820580884,
"total_bits": 2289099280,
"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": 0.9072265056403059,
"total_bits": 2322850816,
"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": 0.8997537760358107,
"total_bits": 2356985360,
"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": 0.912005005698455,
"total_bits": 2406343680,
"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": 0.9491450180740733,
"total_bits": 2901196304,
"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": 0.9562372151566179,
"total_bits": 2950554624,
"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": 0.9727864117410622,
"total_bits": 3360210448,
"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": 0.9741698040773994,
"total_bits": 3477813504,
"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": 0.9793117275755656,
"total_bits": 3790289152,
"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": 0.9928375635002005,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.27.parallel_decoder": {
"attn": [
{
"accuracy": 0.6316209059012563,
"total_bits": 569401344,
"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": 0.6441263430996945,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.6942405959493236,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7608435914704674,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8133628493861148,
"total_bits": 844115968,
"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": 0.8180494912360844,
"total_bits": 845691904,
"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": 0.8895473876281789,
"total_bits": 1083191296,
"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": 0.8942034820977011,
"total_bits": 1085291520,
"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": 0.9012413087644076,
"total_bits": 1095792640,
"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": 0.9055549690597936,
"total_bits": 1108393984,
"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": 0.9075235657786068,
"total_bits": 1114127360,
"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": 0.9132430576964429,
"total_bits": 1123055616,
"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": 0.922334059680763,
"total_bits": 1179148288,
"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": 0.9284619951718732,
"total_bits": 1195950080,
"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": 0.9572478228885877,
"total_bits": 1440755712,
"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": 0.9640564260710227,
"total_bits": 1464385536,
"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": 0.9703561896948438,
"total_bits": 1620062208,
"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": 0.9819978855942425,
"total_bits": 1779482624,
"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": 0.9923565229155907,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6085809704504515,
"total_bits": 1238922592,
"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": 0.6168242959599746,
"total_bits": 1283618144,
"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": 0.6767605386282268,
"total_bits": 1433121792,
"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": 0.6948527364354384,
"total_bits": 1608233984,
"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": 0.7995657905151969,
"total_bits": 1812774416,
"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": 0.8152049528925043,
"total_bits": 1862132736,
"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": 0.8411439382716229,
"total_bits": 2002191888,
"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": 0.8968532822634044,
"total_bits": 2289099280,
"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": 0.9059231794978443,
"total_bits": 2322850816,
"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": 0.8982880652735108,
"total_bits": 2356985360,
"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": 0.9106766769760534,
"total_bits": 2406343680,
"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": 0.9484355381052745,
"total_bits": 2901196304,
"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": 0.9555850227020288,
"total_bits": 2950554624,
"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": 0.9724511102234062,
"total_bits": 3360210448,
"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": 0.9738085757156736,
"total_bits": 3477813504,
"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": 0.9788692153402065,
"total_bits": 3790289152,
"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": 0.992757303272619,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.28.parallel_decoder": {
"attn": [
{
"accuracy": 0.640038937330246,
"total_bits": 569401344,
"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": 0.6507547497749329,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.700255780627853,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7636244995029349,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8160751438454578,
"total_bits": 844115968,
"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": 0.8193507108249163,
"total_bits": 845691904,
"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": 0.8944602722400113,
"total_bits": 1083191296,
"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": 0.8983559471211935,
"total_bits": 1085291520,
"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": 0.9052740955039075,
"total_bits": 1095792640,
"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": 0.9095435311135492,
"total_bits": 1108393984,
"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": 0.9087496897891948,
"total_bits": 1114127360,
"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": 0.9143305879674459,
"total_bits": 1123055616,
"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": 0.9245690097542185,
"total_bits": 1179148288,
"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": 0.9303841561471161,
"total_bits": 1195950080,
"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": 0.9587813629523704,
"total_bits": 1440755712,
"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": 0.9650872800695269,
"total_bits": 1464385536,
"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": 0.9723610758389297,
"total_bits": 1620062208,
"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": 0.9826551354245135,
"total_bits": 1779482624,
"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": 0.9928219855125797,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6033016568736026,
"total_bits": 1238922592,
"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": 0.611611284707722,
"total_bits": 1283618144,
"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": 0.6711850683940084,
"total_bits": 1433121792,
"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": 0.6892087145855552,
"total_bits": 1608233984,
"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": 0.7966055689673675,
"total_bits": 1812774416,
"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": 0.8124285865771144,
"total_bits": 1862132736,
"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": 0.8383260447728007,
"total_bits": 2002191888,
"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": 0.8953291722818425,
"total_bits": 2289099280,
"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": 0.904512632834284,
"total_bits": 2322850816,
"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": 0.896818384920296,
"total_bits": 2356985360,
"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": 0.9093387444552622,
"total_bits": 2406343680,
"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": 0.9477017735572237,
"total_bits": 2901196304,
"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": 0.9549251288959855,
"total_bits": 2950554624,
"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": 0.9720859871687073,
"total_bits": 3360210448,
"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": 0.9734486511075183,
"total_bits": 3477813504,
"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": 0.9785163275113231,
"total_bits": 3790289152,
"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": 0.9926809896224815,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.29.parallel_decoder": {
"attn": [
{
"accuracy": 0.6385805246077085,
"total_bits": 569401344,
"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": 0.6483267545700073,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7047126716689059,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7651289332854121,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.813574952514548,
"total_bits": 844115968,
"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": 0.8180784802687795,
"total_bits": 845691904,
"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": 0.8937087533505339,
"total_bits": 1083191296,
"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": 0.8985263607219646,
"total_bits": 1085291520,
"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": 0.9056222654486957,
"total_bits": 1095792640,
"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": 0.909725229598974,
"total_bits": 1108393984,
"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": 0.9071109083138014,
"total_bits": 1114127360,
"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": 0.9136686905434257,
"total_bits": 1123055616,
"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": 0.9259366322504847,
"total_bits": 1179148288,
"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": 0.9315436915739587,
"total_bits": 1195950080,
"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": 0.9597015886714584,
"total_bits": 1440755712,
"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": 0.9657574795970791,
"total_bits": 1464385536,
"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": 0.9726059343665838,
"total_bits": 1620062208,
"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": 0.9833461325615644,
"total_bits": 1779482624,
"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": 0.9928528621155572,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6002251819560402,
"total_bits": 1238922592,
"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": 0.6087676694518642,
"total_bits": 1283618144,
"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": 0.6670165830536893,
"total_bits": 1433121792,
"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": 0.6847355757888995,
"total_bits": 1608233984,
"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": 0.7948666912944693,
"total_bits": 1812774416,
"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": 0.8108586080764469,
"total_bits": 1862132736,
"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": 0.8362211506617696,
"total_bits": 2002191888,
"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": 0.894373643947275,
"total_bits": 2289099280,
"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": 0.903621260664965,
"total_bits": 2322850816,
"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": 0.8959144981283891,
"total_bits": 2356985360,
"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": 0.9085589702192106,
"total_bits": 2406343680,
"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": 0.9472457246952936,
"total_bits": 2901196304,
"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": 0.9545374177396297,
"total_bits": 2950554624,
"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": 0.9718159919506625,
"total_bits": 3360210448,
"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": 0.9732080044126824,
"total_bits": 3477813504,
"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": 0.9781564489791268,
"total_bits": 3790289152,
"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": 0.9925729546107744,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.30.parallel_decoder": {
"attn": [
{
"accuracy": 0.6070975629906905,
"total_bits": 569401344,
"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": 0.6179946002207304,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.6795666672681507,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7447563488232463,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7965659154088873,
"total_bits": 844115968,
"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": 0.8020994271102705,
"total_bits": 845691904,
"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": 0.884879154594321,
"total_bits": 1083191296,
"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": 0.8899983457828823,
"total_bits": 1085291520,
"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": 0.8975679674430898,
"total_bits": 1095792640,
"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": 0.9016001377450793,
"total_bits": 1108393984,
"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": 0.898614024253268,
"total_bits": 1114127360,
"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": 0.9059575874554484,
"total_bits": 1123055616,
"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": 0.9198297462181041,
"total_bits": 1179148288,
"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": 0.9256258163797229,
"total_bits": 1195950080,
"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": 0.9562845349703964,
"total_bits": 1440755712,
"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": 0.9629553676043686,
"total_bits": 1464385536,
"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": 0.970199947196402,
"total_bits": 1620062208,
"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": 0.9819454317912459,
"total_bits": 1779482624,
"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": 0.9923424144324503,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.5968017515383268,
"total_bits": 1238922592,
"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": 0.6054277184762453,
"total_bits": 1283618144,
"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": 0.6627508982231742,
"total_bits": 1433121792,
"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": 0.6800436032445807,
"total_bits": 1608233984,
"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": 0.7928170710802078,
"total_bits": 1812774416,
"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": 0.808986244233031,
"total_bits": 1862132736,
"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": 0.8338963538408279,
"total_bits": 2002191888,
"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": 0.8933157269891939,
"total_bits": 2289099280,
"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": 0.9026911337124675,
"total_bits": 2322850816,
"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": 0.8948355662195306,
"total_bits": 2356985360,
"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": 0.9076380419887995,
"total_bits": 2406343680,
"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": 0.9466749437545475,
"total_bits": 2901196304,
"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": 0.9540680799829333,
"total_bits": 2950554624,
"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": 0.9714938120818452,
"total_bits": 3360210448,
"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": 0.9729370018957477,
"total_bits": 3477813504,
"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": 0.9777911090733189,
"total_bits": 3790289152,
"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": 0.9925345173350683,
"total_bits": 4447746304,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
]
},
"model.layers.31.parallel_decoder": {
"attn": [
{
"accuracy": 0.6110926254799491,
"total_bits": 569401344,
"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": 0.6260434436170679,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.6870943182393124,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7508228797661631,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8002250923922187,
"total_bits": 844115968,
"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": 0.8037349142526325,
"total_bits": 845691904,
"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": 0.8866623545947828,
"total_bits": 1083191296,
"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": 0.891693544230963,
"total_bits": 1085291520,
"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": 0.8985181274382692,
"total_bits": 1095792640,
"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": 0.9025690218335704,
"total_bits": 1108393984,
"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": 0.9005848004629737,
"total_bits": 1114127360,
"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": 0.9075740909105853,
"total_bits": 1123055616,
"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": 0.9213807167191255,
"total_bits": 1179148288,
"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": 0.9274644573268137,
"total_bits": 1195950080,
"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": 0.9573622902757243,
"total_bits": 1440755712,
"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": 0.9637917429208755,
"total_bits": 1464385536,
"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": 0.9706283426402431,
"total_bits": 1620062208,
"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": 0.9823014198949462,
"total_bits": 1779482624,
"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": 0.9923756315225833,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.5917604095057437,
"total_bits": 1238922592,
"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": 0.6006523951103813,
"total_bits": 1283618144,
"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": 0.6572132392933494,
"total_bits": 1433121792,
"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": 0.674357040932304,
"total_bits": 1608233984,
"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": 0.7901132130309155,
"total_bits": 1812774416,
"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": 0.8065273714692969,
"total_bits": 1862132736,
"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": 0.831144895992781,
"total_bits": 2002191888,
"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": 0.8919408517448526,
"total_bits": 2289099280,
"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": 0.9014052701623816,
"total_bits": 2322850816,
"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": 0.8935016405425573,
"total_bits": 2356985360,
"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": 0.9064366935115111,
"total_bits": 2406343680,
"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": 0.9460286296119815,
"total_bits": 2901196304,
"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": 0.953472695460445,
"total_bits": 2950554624,
"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": 0.9711597404981914,
"total_bits": 3360210448,
"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": 0.9725909580133463,
"total_bits": 3477813504,
"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": 0.9773633473209644,
"total_bits": 3790289152,
"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": 0.99241310755085,
"total_bits": 4447746304,
"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.32.parallel_decoder": {
"attn": [
{
"accuracy": 0.6299126501146115,
"total_bits": 569401344,
"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": 0.6425659766322689,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.6971385078994852,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7552546247055656,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8100783573953729,
"total_bits": 844115968,
"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": 0.8141838492531526,
"total_bits": 845691904,
"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": 0.8924402963174016,
"total_bits": 1083191296,
"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": 0.896641895567116,
"total_bits": 1085291520,
"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": 0.9039139496652704,
"total_bits": 1095792640,
"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": 0.9078168547467181,
"total_bits": 1108393984,
"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": 0.9063080172789724,
"total_bits": 1114127360,
"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": 0.9119661481756913,
"total_bits": 1123055616,
"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": 0.9245154877242289,
"total_bits": 1179148288,
"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": 0.9304967585362887,
"total_bits": 1195950080,
"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": 0.9591338754092392,
"total_bits": 1440755712,
"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": 0.9654451703750774,
"total_bits": 1464385536,
"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": 0.972312867641449,
"total_bits": 1620062208,
"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": 0.9827793266526178,
"total_bits": 1779482624,
"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": 0.9927799325240286,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.5858678472669501,
"total_bits": 1238922592,
"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": 0.5949741021582955,
"total_bits": 1283618144,
"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": 0.6511183334024329,
"total_bits": 1433121792,
"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": 0.6680631245437421,
"total_bits": 1608233984,
"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": 0.7867142275760048,
"total_bits": 1812774416,
"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": 0.8035224283996382,
"total_bits": 1862132736,
"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": 0.8279722936843571,
"total_bits": 2002191888,
"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": 0.8902567111347851,
"total_bits": 2289099280,
"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": 0.8997994869163162,
"total_bits": 2322850816,
"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": 0.8917160540034896,
"total_bits": 2356985360,
"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": 0.9049372261292056,
"total_bits": 2406343680,
"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": 0.9450861799873804,
"total_bits": 2901196304,
"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": 0.9527057338702052,
"total_bits": 2950554624,
"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": 0.9705948720833188,
"total_bits": 3360210448,
"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": 0.9721348184700075,
"total_bits": 3477813504,
"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": 0.9768359734627762,
"total_bits": 3790289152,
"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": 0.9922984892649478,
"total_bits": 4447746304,
"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.33.parallel_decoder": {
"attn": [
{
"accuracy": 0.640726078497736,
"total_bits": 569401344,
"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": 0.6522314266154641,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7003191395809776,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7615780446090197,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8129753846871226,
"total_bits": 844115968,
"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": 0.8151350523296156,
"total_bits": 845691904,
"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": 0.894299181668382,
"total_bits": 1083191296,
"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": 0.896803431605038,
"total_bits": 1085291520,
"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": 0.9043257577638877,
"total_bits": 1095792640,
"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": 0.908356663035719,
"total_bits": 1108393984,
"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": 0.9061856034554934,
"total_bits": 1114127360,
"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": 0.912338194486342,
"total_bits": 1123055616,
"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": 0.9243822215419066,
"total_bits": 1179148288,
"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": 0.9304221397952029,
"total_bits": 1195950080,
"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": 0.9594694564217016,
"total_bits": 1440755712,
"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": 0.9653669466313562,
"total_bits": 1464385536,
"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": 0.972766180375689,
"total_bits": 1620062208,
"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": 0.9829727345587391,
"total_bits": 1779482624,
"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": 0.9929425085796729,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.5879478187937486,
"total_bits": 1238922592,
"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": 0.5970923335928666,
"total_bits": 1283618144,
"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": 0.651789969519565,
"total_bits": 1433121792,
"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": 0.6684821577448594,
"total_bits": 1608233984,
"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": 0.7877233734256343,
"total_bits": 1812774416,
"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": 0.8043441058773744,
"total_bits": 1862132736,
"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": 0.8282079571171811,
"total_bits": 2002191888,
"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": 0.8906613968704876,
"total_bits": 2289099280,
"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": 0.9002021496233187,
"total_bits": 2322850816,
"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": 0.8922715492938694,
"total_bits": 2356985360,
"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": 0.9053345412800187,
"total_bits": 2406343680,
"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": 0.9453923366963863,
"total_bits": 2901196304,
"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": 0.9529277088218614,
"total_bits": 2950554624,
"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": 0.9708236252006731,
"total_bits": 3360210448,
"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": 0.9722617619523877,
"total_bits": 3477813504,
"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": 0.9768811738804767,
"total_bits": 3790289152,
"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": 0.9923407042301015,
"total_bits": 4447746304,
"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.34.parallel_decoder": {
"attn": [
{
"accuracy": 0.6690214260628349,
"total_bits": 569401344,
"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": 0.6779401302337646,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7166163662546559,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7768551024951433,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8317825115825,
"total_bits": 844115968,
"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": 0.8344941178434774,
"total_bits": 845691904,
"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": 0.9003256880923322,
"total_bits": 1083191296,
"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": 0.9032399062263339,
"total_bits": 1085291520,
"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": 0.9104317407074728,
"total_bits": 1095792640,
"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": 0.914304188599712,
"total_bits": 1108393984,
"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": 0.9162702727082529,
"total_bits": 1114127360,
"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": 0.9215055678627992,
"total_bits": 1123055616,
"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": 0.9280754900292346,
"total_bits": 1179148288,
"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": 0.933455765051277,
"total_bits": 1195950080,
"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": 0.9612037265967381,
"total_bits": 1440755712,
"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": 0.9669678923919013,
"total_bits": 1464385536,
"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": 0.9742367419561273,
"total_bits": 1620062208,
"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": 0.9836187957246837,
"total_bits": 1779482624,
"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": 0.993331400285426,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6026556021288821,
"total_bits": 1238922592,
"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": 0.6115019666521173,
"total_bits": 1283618144,
"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": 0.6630146942640606,
"total_bits": 1433121792,
"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": 0.6789547252027612,
"total_bits": 1608233984,
"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": 0.7952204479982978,
"total_bits": 1812774416,
"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": 0.8113293616395247,
"total_bits": 1862132736,
"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": 0.833868712971085,
"total_bits": 2002191888,
"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": 0.8941278908597796,
"total_bits": 2289099280,
"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": 0.9034012638424572,
"total_bits": 2322850816,
"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": 0.8959445549469245,
"total_bits": 2356985360,
"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": 0.9086241459375933,
"total_bits": 2406343680,
"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": 0.9471900235665471,
"total_bits": 2901196304,
"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": 0.9545012706596601,
"total_bits": 2950554624,
"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": 0.9715948271516123,
"total_bits": 3360210448,
"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": 0.9731835114132417,
"total_bits": 3477813504,
"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": 0.9775261610354248,
"total_bits": 3790289152,
"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": 0.9924516774419891,
"total_bits": 4447746304,
"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.35.parallel_decoder": {
"attn": [
{
"accuracy": 0.6761380316395509,
"total_bits": 569401344,
"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": 0.6876343127928282,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7246147284382267,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7830455954137602,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8352807721025065,
"total_bits": 844115968,
"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": 0.8373239867781338,
"total_bits": 845691904,
"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": 0.90183963077633,
"total_bits": 1083191296,
"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": 0.9037701283630571,
"total_bits": 1085291520,
"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": 0.9115435335981219,
"total_bits": 1095792640,
"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": 0.9152212644878187,
"total_bits": 1108393984,
"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": 0.9171895031866274,
"total_bits": 1114127360,
"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": 0.922803972308573,
"total_bits": 1123055616,
"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": 0.9287821858337051,
"total_bits": 1179148288,
"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": 0.93476745623507,
"total_bits": 1195950080,
"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": 0.9615729160998997,
"total_bits": 1440755712,
"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": 0.9673066797029031,
"total_bits": 1464385536,
"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": 0.9740692418264715,
"total_bits": 1620062208,
"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": 0.9837371523148919,
"total_bits": 1779482624,
"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": 0.9932309408476087,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6180715623654818,
"total_bits": 1238922592,
"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": 0.6266903940000033,
"total_bits": 1283618144,
"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": 0.6754499783641414,
"total_bits": 1433121792,
"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": 0.6910713424808101,
"total_bits": 1608233984,
"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": 0.803187335008069,
"total_bits": 1812774416,
"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": 0.8187072920171838,
"total_bits": 1862132736,
"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": 0.8401506582373067,
"total_bits": 2002191888,
"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": 0.8978716907532591,
"total_bits": 2289099280,
"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": 0.9068741033736029,
"total_bits": 2322850816,
"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": 0.900031606225591,
"total_bits": 2356985360,
"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": 0.9122053506342989,
"total_bits": 2406343680,
"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": 0.949292844651561,
"total_bits": 2901196304,
"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": 0.9563146323749894,
"total_bits": 2950554624,
"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": 0.9726968024318156,
"total_bits": 3360210448,
"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": 0.9742456170681276,
"total_bits": 3477813504,
"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": 0.9784094316786841,
"total_bits": 3790289152,
"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": 0.9927180743237075,
"total_bits": 4447746304,
"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.36.parallel_decoder": {
"attn": [
{
"accuracy": 0.6759921131949675,
"total_bits": 569401344,
"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": 0.6860632817996175,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7245600521564484,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7858527271371138,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8351044827385953,
"total_bits": 844115968,
"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": 0.8371424643616927,
"total_bits": 845691904,
"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": 0.9026609012170842,
"total_bits": 1083191296,
"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": 0.9053609265308631,
"total_bits": 1085291520,
"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": 0.9122907578160888,
"total_bits": 1095792640,
"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": 0.916508527962785,
"total_bits": 1108393984,
"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": 0.9168802444872103,
"total_bits": 1114127360,
"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": 0.9225738656363989,
"total_bits": 1123055616,
"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": 0.9289579375794059,
"total_bits": 1179148288,
"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": 0.9350364827796033,
"total_bits": 1195950080,
"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": 0.9616822274891954,
"total_bits": 1440755712,
"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": 0.967545373561351,
"total_bits": 1464385536,
"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": 0.9740374556889659,
"total_bits": 1620062208,
"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": 0.9839312904269287,
"total_bits": 1779482624,
"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": 0.9932215021383998,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.6543641262932828,
"total_bits": 1238922592,
"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": 0.6620304192367352,
"total_bits": 1283618144,
"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": 0.706206572683234,
"total_bits": 1433121792,
"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": 0.7209354105748629,
"total_bits": 1608233984,
"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": 0.8217865620788775,
"total_bits": 1812774416,
"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": 0.8357619815751126,
"total_bits": 1862132736,
"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": 0.8553736256925684,
"total_bits": 2002191888,
"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": 0.9069977618361774,
"total_bits": 2289099280,
"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": 0.9153610185572976,
"total_bits": 2322850816,
"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": 0.909483052790165,
"total_bits": 2356985360,
"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": 0.9204554836216726,
"total_bits": 2406343680,
"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": 0.9540831928974703,
"total_bits": 2901196304,
"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": 0.9604147974597781,
"total_bits": 2950554624,
"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": 0.975176637698161,
"total_bits": 3360210448,
"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": 0.9767107324380624,
"total_bits": 3477813504,
"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": 0.9806061595874397,
"total_bits": 3790289152,
"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": 0.9933355768995458,
"total_bits": 4447746304,
"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.37.parallel_decoder": {
"attn": [
{
"accuracy": 0.719902375026753,
"total_bits": 569401344,
"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": 0.7290053406828327,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7629507364411103,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8161222228878423,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8582289352228767,
"total_bits": 844115968,
"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": 0.8603930673316905,
"total_bits": 845691904,
"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": 0.9184123585490804,
"total_bits": 1083191296,
"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": 0.920791708129017,
"total_bits": 1085291520,
"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": 0.9266083448341018,
"total_bits": 1095792640,
"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": 0.9298872420270192,
"total_bits": 1108393984,
"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": 0.9294003932491729,
"total_bits": 1114127360,
"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": 0.9336396824372442,
"total_bits": 1123055616,
"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": 0.9401608080063996,
"total_bits": 1179148288,
"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": 0.9446849889661136,
"total_bits": 1195950080,
"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": 0.9677454588052473,
"total_bits": 1440755712,
"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": 0.9724625817156936,
"total_bits": 1464385536,
"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": 0.9789758740287078,
"total_bits": 1620062208,
"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": 0.9862559594606098,
"total_bits": 1779482624,
"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": 0.9945423636634491,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.7035647301297439,
"total_bits": 1238922592,
"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": 0.70995295518323,
"total_bits": 1283618144,
"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": 0.7481388645736795,
"total_bits": 1433121792,
"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": 0.7615541141284139,
"total_bits": 1608233984,
"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": 0.8466819791417373,
"total_bits": 1812774416,
"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": 0.8588746809645703,
"total_bits": 1862132736,
"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": 0.8761088914777103,
"total_bits": 2002191888,
"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": 0.91934044070934,
"total_bits": 2289099280,
"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": 0.9269114181791481,
"total_bits": 2322850816,
"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": 0.9220720909928021,
"total_bits": 2356985360,
"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": 0.9316247485970196,
"total_bits": 2406343680,
"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": 0.9604188197929608,
"total_bits": 2901196304,
"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": 0.9659406789823582,
"total_bits": 2950554624,
"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": 0.9784131967707684,
"total_bits": 3360210448,
"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": 0.9799719300905341,
"total_bits": 3477813504,
"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": 0.9835081883755169,
"total_bits": 3790289152,
"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": 0.9941546820398224,
"total_bits": 4447746304,
"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.38.parallel_decoder": {
"attn": [
{
"accuracy": 0.7355329261014336,
"total_bits": 569401344,
"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": 0.7463625442040593,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.7779648978459208,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8300937100460655,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8642328173706406,
"total_bits": 844115968,
"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": 0.8660682404511854,
"total_bits": 845691904,
"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": 0.9273402710494242,
"total_bits": 1083191296,
"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": 0.9287896073962513,
"total_bits": 1085291520,
"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": 0.9335490653389379,
"total_bits": 1095792640,
"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": 0.936850566024843,
"total_bits": 1108393984,
"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": 0.9321994338380664,
"total_bits": 1114127360,
"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": 0.9365658354210227,
"total_bits": 1123055616,
"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": 0.9450494866622121,
"total_bits": 1179148288,
"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": 0.9491907177787078,
"total_bits": 1195950080,
"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": 0.9703349454426452,
"total_bits": 1440755712,
"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": 0.974712689162085,
"total_bits": 1464385536,
"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": 0.9812876783605469,
"total_bits": 1620062208,
"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": 0.9873579738563613,
"total_bits": 1779482624,
"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": 0.9951338947530052,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.7378086722210834,
"total_bits": 1238922592,
"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": 0.7432973337800879,
"total_bits": 1283618144,
"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": 0.7791152486675664,
"total_bits": 1433121792,
"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": 0.791997671911591,
"total_bits": 1608233984,
"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": 0.8636591548198148,
"total_bits": 1812774416,
"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": 0.8749852627515793,
"total_bits": 1862132736,
"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": 0.8912379541679433,
"total_bits": 2002191888,
"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": 0.9277460390799924,
"total_bits": 2289099280,
"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": 0.9349431319064215,
"total_bits": 2322850816,
"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": 0.9303920510175981,
"total_bits": 2356985360,
"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": 0.939228749981052,
"total_bits": 2406343680,
"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": 0.9644969820388054,
"total_bits": 2901196304,
"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": 0.9696334995525447,
"total_bits": 2950554624,
"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": 0.9804502255038211,
"total_bits": 3360210448,
"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": 0.9819700995362118,
"total_bits": 3477813504,
"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": 0.9854927756088344,
"total_bits": 3790289152,
"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": 0.9944961422711218,
"total_bits": 4447746304,
"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.39.parallel_decoder": {
"attn": [
{
"accuracy": 0.7950398349448254,
"total_bits": 569401344,
"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": 0.8040020767795413,
"total_bits": 591945728,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8350759830914045,
"total_bits": 647517184,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.881459705923733,
"total_bits": 713050112,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.8983170245038836,
"total_bits": 844115968,
"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": 0.899737512986911,
"total_bits": 845691904,
"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": 0.9420377027831579,
"total_bits": 1083191296,
"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": 0.9433401705403077,
"total_bits": 1085291520,
"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": 0.947364759680472,
"total_bits": 1095792640,
"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": 0.9498296799628358,
"total_bits": 1108393984,
"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": 0.9493058857164884,
"total_bits": 1114127360,
"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": 0.9524960320835051,
"total_bits": 1123055616,
"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": 0.9586989812945065,
"total_bits": 1179148288,
"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": 0.9619346016919926,
"total_bits": 1195950080,
"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": 0.9776070954670247,
"total_bits": 1440755712,
"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": 0.98093107180964,
"total_bits": 1464385536,
"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": 0.9850358791453274,
"total_bits": 1620062208,
"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": 0.9900984565825447,
"total_bits": 1779482624,
"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": 0.9960984445275053,
"total_bits": 2156933120,
"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
}
}
],
"mlp": [
{
"accuracy": 0.8549274331644962,
"total_bits": 1238922592,
"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": 0.8581229816926152,
"total_bits": 1283618144,
"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": 0.880668190749068,
"total_bits": 1433121792,
"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": 0.8883051095824492,
"total_bits": 1608233984,
"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": 0.9254113151446769,
"total_bits": 1812774416,
"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": 0.9323858256010633,
"total_bits": 1862132736,
"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": 0.9416114972218087,
"total_bits": 2002191888,
"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": 0.9603368932087171,
"total_bits": 2289099280,
"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": 0.9642958925350716,
"total_bits": 2322850816,
"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": 0.9617331655401933,
"total_bits": 2356985360,
"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": 0.9668800246558691,
"total_bits": 2406343680,
"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": 0.9802676994647634,
"total_bits": 2901196304,
"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": 0.9833279732908857,
"total_bits": 2950554624,
"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": 0.9889906577912992,
"total_bits": 3360210448,
"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": 0.9898536756242576,
"total_bits": 3477813504,
"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": 0.9920510497728461,
"total_bits": 3790289152,
"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": 0.9966997058057275,
"total_bits": 4447746304,
"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": 42
}