Upload folder using huggingface_hub
Browse files- hidden_states.safetensors +1 -1
- job_new.json +921 -873
- measurement.json +13 -13
- output-00001-of-00002.safetensors +2 -2
- output-00002-of-00002.safetensors +2 -2
hidden_states.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 2097160800
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:51fa1df539e6d53a8e76a55c37fbd0bb987b123d7af8b307f99b56fe68593797
|
| 3 |
size 2097160800
|
job_new.json
CHANGED
|
@@ -16952,7 +16952,7 @@
|
|
| 16952 |
}
|
| 16953 |
},
|
| 16954 |
{
|
| 16955 |
-
"accuracy": 0.
|
| 16956 |
"total_bits": 899923248,
|
| 16957 |
"gate_proj": {
|
| 16958 |
"group_size": {
|
|
@@ -17056,7 +17056,7 @@
|
|
| 17056 |
}
|
| 17057 |
},
|
| 17058 |
{
|
| 17059 |
-
"accuracy": 0.
|
| 17060 |
"total_bits": 995125808,
|
| 17061 |
"gate_proj": {
|
| 17062 |
"group_size": {
|
|
@@ -17105,7 +17105,7 @@
|
|
| 17105 |
}
|
| 17106 |
},
|
| 17107 |
{
|
| 17108 |
-
"accuracy": 0.
|
| 17109 |
"total_bits": 1137906608,
|
| 17110 |
"gate_proj": {
|
| 17111 |
"group_size": {
|
|
@@ -17899,7 +17899,7 @@
|
|
| 17899 |
}
|
| 17900 |
},
|
| 17901 |
{
|
| 17902 |
-
"accuracy": 0.
|
| 17903 |
"total_bits": 206079360,
|
| 17904 |
"q_proj": {
|
| 17905 |
"group_size": {
|
|
@@ -21545,7 +21545,7 @@
|
|
| 21545 |
}
|
| 21546 |
},
|
| 21547 |
{
|
| 21548 |
-
"accuracy": 0.
|
| 21549 |
"total_bits": 143375360,
|
| 21550 |
"q_proj": {
|
| 21551 |
"group_size": {
|
|
@@ -24554,7 +24554,7 @@
|
|
| 24554 |
],
|
| 24555 |
"model.layers.12.mlp": [
|
| 24556 |
{
|
| 24557 |
-
"accuracy": 0.
|
| 24558 |
"total_bits": 614790432,
|
| 24559 |
"gate_proj": {
|
| 24560 |
"group_size": {
|
|
@@ -24606,7 +24606,7 @@
|
|
| 24606 |
}
|
| 24607 |
},
|
| 24608 |
{
|
| 24609 |
-
"accuracy": 0.
|
| 24610 |
"total_bits": 637728032,
|
| 24611 |
"gate_proj": {
|
| 24612 |
"group_size": {
|
|
@@ -46269,7 +46269,7 @@
|
|
| 46269 |
}
|
| 46270 |
},
|
| 46271 |
{
|
| 46272 |
-
"accuracy": 0.
|
| 46273 |
"total_bits": 925328640,
|
| 46274 |
"gate_proj": {
|
| 46275 |
"group_size": {
|
|
@@ -51868,7 +51868,7 @@
|
|
| 51868 |
],
|
| 51869 |
"model.layers.26.mlp": [
|
| 51870 |
{
|
| 51871 |
-
"accuracy": 0.
|
| 51872 |
"total_bits": 614790432,
|
| 51873 |
"gate_proj": {
|
| 51874 |
"group_size": {
|
|
@@ -55136,7 +55136,7 @@
|
|
| 55136 |
}
|
| 55137 |
},
|
| 55138 |
{
|
| 55139 |
-
"accuracy": 0.
|
| 55140 |
"total_bits": 267324160,
|
| 55141 |
"q_proj": {
|
| 55142 |
"group_size": {
|
|
@@ -55188,7 +55188,7 @@
|
|
| 55188 |
}
|
| 55189 |
},
|
| 55190 |
{
|
| 55191 |
-
"accuracy": 0.
|
| 55192 |
"total_bits": 271022080,
|
| 55193 |
"q_proj": {
|
| 55194 |
"group_size": {
|
|
@@ -62992,7 +62992,7 @@
|
|
| 62992 |
}
|
| 62993 |
},
|
| 62994 |
{
|
| 62995 |
-
"accuracy": 0.
|
| 62996 |
"total_bits": 271022080,
|
| 62997 |
"q_proj": {
|
| 62998 |
"group_size": {
|
|
@@ -64403,7 +64403,7 @@
|
|
| 64403 |
],
|
| 64404 |
"model.layers.33.self_attn": [
|
| 64405 |
{
|
| 64406 |
-
"accuracy": 0.
|
| 64407 |
"total_bits": 139115520,
|
| 64408 |
"q_proj": {
|
| 64409 |
"group_size": {
|
|
@@ -78063,14 +78063,14 @@
|
|
| 78063 |
},
|
| 78064 |
"strategy": {
|
| 78065 |
"model.layers.0.self_attn": {
|
| 78066 |
-
"accuracy": 0.
|
| 78067 |
-
"total_bits":
|
| 78068 |
"q_proj": {
|
| 78069 |
"group_size": {
|
| 78070 |
-
"
|
| 78071 |
},
|
| 78072 |
"bits": [
|
| 78073 |
-
|
| 78074 |
],
|
| 78075 |
"bits_prop": [
|
| 78076 |
1
|
|
@@ -78079,10 +78079,10 @@
|
|
| 78079 |
},
|
| 78080 |
"k_proj": {
|
| 78081 |
"group_size": {
|
| 78082 |
-
"
|
| 78083 |
},
|
| 78084 |
"bits": [
|
| 78085 |
-
|
| 78086 |
],
|
| 78087 |
"bits_prop": [
|
| 78088 |
1
|
|
@@ -78091,7 +78091,7 @@
|
|
| 78091 |
},
|
| 78092 |
"v_proj": {
|
| 78093 |
"group_size": {
|
| 78094 |
-
"8":
|
| 78095 |
},
|
| 78096 |
"bits": [
|
| 78097 |
8
|
|
@@ -78103,10 +78103,10 @@
|
|
| 78103 |
},
|
| 78104 |
"o_proj": {
|
| 78105 |
"group_size": {
|
| 78106 |
-
"
|
| 78107 |
},
|
| 78108 |
"bits": [
|
| 78109 |
-
|
| 78110 |
],
|
| 78111 |
"bits_prop": [
|
| 78112 |
1
|
|
@@ -78167,14 +78167,14 @@
|
|
| 78167 |
}
|
| 78168 |
},
|
| 78169 |
"model.layers.1.self_attn": {
|
| 78170 |
-
"accuracy": 0.
|
| 78171 |
-
"total_bits":
|
| 78172 |
"q_proj": {
|
| 78173 |
"group_size": {
|
| 78174 |
-
"
|
| 78175 |
},
|
| 78176 |
"bits": [
|
| 78177 |
-
|
| 78178 |
],
|
| 78179 |
"bits_prop": [
|
| 78180 |
1
|
|
@@ -78183,10 +78183,10 @@
|
|
| 78183 |
},
|
| 78184 |
"k_proj": {
|
| 78185 |
"group_size": {
|
| 78186 |
-
"
|
| 78187 |
},
|
| 78188 |
"bits": [
|
| 78189 |
-
|
| 78190 |
],
|
| 78191 |
"bits_prop": [
|
| 78192 |
1
|
|
@@ -78195,7 +78195,7 @@
|
|
| 78195 |
},
|
| 78196 |
"v_proj": {
|
| 78197 |
"group_size": {
|
| 78198 |
-
"8":
|
| 78199 |
},
|
| 78200 |
"bits": [
|
| 78201 |
8
|
|
@@ -78207,10 +78207,10 @@
|
|
| 78207 |
},
|
| 78208 |
"o_proj": {
|
| 78209 |
"group_size": {
|
| 78210 |
-
"
|
| 78211 |
},
|
| 78212 |
"bits": [
|
| 78213 |
-
|
| 78214 |
],
|
| 78215 |
"bits_prop": [
|
| 78216 |
1
|
|
@@ -78271,38 +78271,44 @@
|
|
| 78271 |
}
|
| 78272 |
},
|
| 78273 |
"model.layers.2.self_attn": {
|
| 78274 |
-
"accuracy": 0.
|
| 78275 |
-
"total_bits":
|
| 78276 |
"q_proj": {
|
| 78277 |
"group_size": {
|
| 78278 |
-
"6": 32
|
|
|
|
| 78279 |
},
|
| 78280 |
"bits": [
|
| 78281 |
-
6
|
|
|
|
| 78282 |
],
|
| 78283 |
"bits_prop": [
|
| 78284 |
-
1
|
|
|
|
| 78285 |
],
|
| 78286 |
"scale_bits": 4
|
| 78287 |
},
|
| 78288 |
"k_proj": {
|
| 78289 |
"group_size": {
|
| 78290 |
-
"6": 32
|
|
|
|
| 78291 |
},
|
| 78292 |
"bits": [
|
| 78293 |
-
6
|
|
|
|
| 78294 |
],
|
| 78295 |
"bits_prop": [
|
| 78296 |
-
1
|
|
|
|
| 78297 |
],
|
| 78298 |
"scale_bits": 4
|
| 78299 |
},
|
| 78300 |
"v_proj": {
|
| 78301 |
"group_size": {
|
| 78302 |
-
"
|
| 78303 |
},
|
| 78304 |
"bits": [
|
| 78305 |
-
|
| 78306 |
],
|
| 78307 |
"bits_prop": [
|
| 78308 |
1
|
|
@@ -78311,20 +78317,23 @@
|
|
| 78311 |
},
|
| 78312 |
"o_proj": {
|
| 78313 |
"group_size": {
|
| 78314 |
-
"6": 32
|
|
|
|
| 78315 |
},
|
| 78316 |
"bits": [
|
| 78317 |
-
6
|
|
|
|
| 78318 |
],
|
| 78319 |
"bits_prop": [
|
| 78320 |
-
1
|
|
|
|
| 78321 |
],
|
| 78322 |
"scale_bits": 4
|
| 78323 |
}
|
| 78324 |
},
|
| 78325 |
"model.layers.2.mlp": {
|
| 78326 |
-
"accuracy": 0.
|
| 78327 |
-
"total_bits":
|
| 78328 |
"gate_proj": {
|
| 78329 |
"group_size": {
|
| 78330 |
"8": 128,
|
|
@@ -78357,27 +78366,24 @@
|
|
| 78357 |
},
|
| 78358 |
"down_proj": {
|
| 78359 |
"group_size": {
|
| 78360 |
-
"8": 128
|
| 78361 |
-
"6": 128
|
| 78362 |
},
|
| 78363 |
"bits": [
|
| 78364 |
-
8
|
| 78365 |
-
6
|
| 78366 |
],
|
| 78367 |
"bits_prop": [
|
| 78368 |
-
|
| 78369 |
-
0.85
|
| 78370 |
],
|
| 78371 |
"scale_bits": 4
|
| 78372 |
}
|
| 78373 |
},
|
| 78374 |
"model.layers.3.self_attn": {
|
| 78375 |
-
"accuracy": 0.
|
| 78376 |
-
"total_bits":
|
| 78377 |
"q_proj": {
|
| 78378 |
"group_size": {
|
| 78379 |
-
"6":
|
| 78380 |
-
"5":
|
| 78381 |
},
|
| 78382 |
"bits": [
|
| 78383 |
6,
|
|
@@ -78391,8 +78397,8 @@
|
|
| 78391 |
},
|
| 78392 |
"k_proj": {
|
| 78393 |
"group_size": {
|
| 78394 |
-
"6":
|
| 78395 |
-
"5":
|
| 78396 |
},
|
| 78397 |
"bits": [
|
| 78398 |
6,
|
|
@@ -78406,7 +78412,7 @@
|
|
| 78406 |
},
|
| 78407 |
"v_proj": {
|
| 78408 |
"group_size": {
|
| 78409 |
-
"6":
|
| 78410 |
},
|
| 78411 |
"bits": [
|
| 78412 |
6
|
|
@@ -78418,8 +78424,8 @@
|
|
| 78418 |
},
|
| 78419 |
"o_proj": {
|
| 78420 |
"group_size": {
|
| 78421 |
-
"6":
|
| 78422 |
-
"5":
|
| 78423 |
},
|
| 78424 |
"bits": [
|
| 78425 |
6,
|
|
@@ -78656,16 +78662,16 @@
|
|
| 78656 |
}
|
| 78657 |
},
|
| 78658 |
"model.layers.5.mlp": {
|
| 78659 |
-
"accuracy": 0.
|
| 78660 |
-
"total_bits":
|
| 78661 |
"gate_proj": {
|
| 78662 |
"group_size": {
|
| 78663 |
-
"
|
| 78664 |
-
"
|
| 78665 |
},
|
| 78666 |
"bits": [
|
| 78667 |
-
|
| 78668 |
-
|
| 78669 |
],
|
| 78670 |
"bits_prop": [
|
| 78671 |
0.1,
|
|
@@ -78675,71 +78681,74 @@
|
|
| 78675 |
},
|
| 78676 |
"up_proj": {
|
| 78677 |
"group_size": {
|
| 78678 |
-
"
|
| 78679 |
-
"
|
| 78680 |
},
|
| 78681 |
"bits": [
|
| 78682 |
-
|
| 78683 |
-
|
| 78684 |
],
|
| 78685 |
"bits_prop": [
|
| 78686 |
-
0.
|
| 78687 |
-
0.
|
| 78688 |
],
|
| 78689 |
"scale_bits": 4
|
| 78690 |
},
|
| 78691 |
"down_proj": {
|
| 78692 |
"group_size": {
|
| 78693 |
-
"8":
|
| 78694 |
-
"6": 128
|
| 78695 |
-
"5": 128
|
| 78696 |
},
|
| 78697 |
"bits": [
|
| 78698 |
8,
|
| 78699 |
-
6
|
| 78700 |
-
5
|
| 78701 |
],
|
| 78702 |
"bits_prop": [
|
| 78703 |
-
0.
|
| 78704 |
-
0.1,
|
| 78705 |
0.85
|
| 78706 |
],
|
| 78707 |
"scale_bits": 4
|
| 78708 |
}
|
| 78709 |
},
|
| 78710 |
"model.layers.6.self_attn": {
|
| 78711 |
-
"accuracy": 0.
|
| 78712 |
-
"total_bits":
|
| 78713 |
"q_proj": {
|
| 78714 |
"group_size": {
|
| 78715 |
-
"6": 32
|
|
|
|
| 78716 |
},
|
| 78717 |
"bits": [
|
| 78718 |
-
6
|
|
|
|
| 78719 |
],
|
| 78720 |
"bits_prop": [
|
| 78721 |
-
1
|
|
|
|
| 78722 |
],
|
| 78723 |
"scale_bits": 4
|
| 78724 |
},
|
| 78725 |
"k_proj": {
|
| 78726 |
"group_size": {
|
| 78727 |
-
"6": 32
|
|
|
|
| 78728 |
},
|
| 78729 |
"bits": [
|
| 78730 |
-
6
|
|
|
|
| 78731 |
],
|
| 78732 |
"bits_prop": [
|
| 78733 |
-
1
|
|
|
|
| 78734 |
],
|
| 78735 |
"scale_bits": 4
|
| 78736 |
},
|
| 78737 |
"v_proj": {
|
| 78738 |
"group_size": {
|
| 78739 |
-
"
|
| 78740 |
},
|
| 78741 |
"bits": [
|
| 78742 |
-
|
| 78743 |
],
|
| 78744 |
"bits_prop": [
|
| 78745 |
1
|
|
@@ -78748,13 +78757,16 @@
|
|
| 78748 |
},
|
| 78749 |
"o_proj": {
|
| 78750 |
"group_size": {
|
| 78751 |
-
"6": 32
|
|
|
|
| 78752 |
},
|
| 78753 |
"bits": [
|
| 78754 |
-
6
|
|
|
|
| 78755 |
],
|
| 78756 |
"bits_prop": [
|
| 78757 |
-
1
|
|
|
|
| 78758 |
],
|
| 78759 |
"scale_bits": 4
|
| 78760 |
}
|
|
@@ -78812,11 +78824,11 @@
|
|
| 78812 |
}
|
| 78813 |
},
|
| 78814 |
"model.layers.7.self_attn": {
|
| 78815 |
-
"accuracy": 0.
|
| 78816 |
-
"total_bits":
|
| 78817 |
"q_proj": {
|
| 78818 |
"group_size": {
|
| 78819 |
-
"6":
|
| 78820 |
},
|
| 78821 |
"bits": [
|
| 78822 |
6
|
|
@@ -78828,7 +78840,7 @@
|
|
| 78828 |
},
|
| 78829 |
"k_proj": {
|
| 78830 |
"group_size": {
|
| 78831 |
-
"6":
|
| 78832 |
},
|
| 78833 |
"bits": [
|
| 78834 |
6
|
|
@@ -78840,10 +78852,10 @@
|
|
| 78840 |
},
|
| 78841 |
"v_proj": {
|
| 78842 |
"group_size": {
|
| 78843 |
-
"
|
| 78844 |
},
|
| 78845 |
"bits": [
|
| 78846 |
-
|
| 78847 |
],
|
| 78848 |
"bits_prop": [
|
| 78849 |
1
|
|
@@ -78852,7 +78864,7 @@
|
|
| 78852 |
},
|
| 78853 |
"o_proj": {
|
| 78854 |
"group_size": {
|
| 78855 |
-
"6":
|
| 78856 |
},
|
| 78857 |
"bits": [
|
| 78858 |
6
|
|
@@ -78864,7 +78876,120 @@
|
|
| 78864 |
}
|
| 78865 |
},
|
| 78866 |
"model.layers.7.mlp": {
|
| 78867 |
-
"accuracy": 0.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78868 |
"total_bits": 1466656000,
|
| 78869 |
"gate_proj": {
|
| 78870 |
"group_size": {
|
|
@@ -78915,40 +79040,10 @@
|
|
| 78915 |
"scale_bits": 4
|
| 78916 |
}
|
| 78917 |
},
|
| 78918 |
-
"model.layers.
|
| 78919 |
-
"accuracy": 0.
|
| 78920 |
-
"total_bits":
|
| 78921 |
"q_proj": {
|
| 78922 |
-
"group_size": {
|
| 78923 |
-
"6": 32,
|
| 78924 |
-
"5": 32
|
| 78925 |
-
},
|
| 78926 |
-
"bits": [
|
| 78927 |
-
6,
|
| 78928 |
-
5
|
| 78929 |
-
],
|
| 78930 |
-
"bits_prop": [
|
| 78931 |
-
0.1,
|
| 78932 |
-
0.9
|
| 78933 |
-
],
|
| 78934 |
-
"scale_bits": 4
|
| 78935 |
-
},
|
| 78936 |
-
"k_proj": {
|
| 78937 |
-
"group_size": {
|
| 78938 |
-
"6": 32,
|
| 78939 |
-
"5": 32
|
| 78940 |
-
},
|
| 78941 |
-
"bits": [
|
| 78942 |
-
6,
|
| 78943 |
-
5
|
| 78944 |
-
],
|
| 78945 |
-
"bits_prop": [
|
| 78946 |
-
0.1,
|
| 78947 |
-
0.9
|
| 78948 |
-
],
|
| 78949 |
-
"scale_bits": 4
|
| 78950 |
-
},
|
| 78951 |
-
"v_proj": {
|
| 78952 |
"group_size": {
|
| 78953 |
"6": 32
|
| 78954 |
},
|
|
@@ -78960,95 +79055,12 @@
|
|
| 78960 |
],
|
| 78961 |
"scale_bits": 4
|
| 78962 |
},
|
| 78963 |
-
"o_proj": {
|
| 78964 |
-
"group_size": {
|
| 78965 |
-
"6": 32,
|
| 78966 |
-
"5": 32
|
| 78967 |
-
},
|
| 78968 |
-
"bits": [
|
| 78969 |
-
6,
|
| 78970 |
-
5
|
| 78971 |
-
],
|
| 78972 |
-
"bits_prop": [
|
| 78973 |
-
0.1,
|
| 78974 |
-
0.9
|
| 78975 |
-
],
|
| 78976 |
-
"scale_bits": 4
|
| 78977 |
-
}
|
| 78978 |
-
},
|
| 78979 |
-
"model.layers.8.mlp": {
|
| 78980 |
-
"accuracy": 0.9908282831311226,
|
| 78981 |
-
"total_bits": 1441250608,
|
| 78982 |
-
"gate_proj": {
|
| 78983 |
-
"group_size": {
|
| 78984 |
-
"6": 128,
|
| 78985 |
-
"5": 128
|
| 78986 |
-
},
|
| 78987 |
-
"bits": [
|
| 78988 |
-
6,
|
| 78989 |
-
5
|
| 78990 |
-
],
|
| 78991 |
-
"bits_prop": [
|
| 78992 |
-
0.1,
|
| 78993 |
-
0.9
|
| 78994 |
-
],
|
| 78995 |
-
"scale_bits": 4
|
| 78996 |
-
},
|
| 78997 |
-
"up_proj": {
|
| 78998 |
-
"group_size": {
|
| 78999 |
-
"6": 128,
|
| 79000 |
-
"5": 128
|
| 79001 |
-
},
|
| 79002 |
-
"bits": [
|
| 79003 |
-
6,
|
| 79004 |
-
5
|
| 79005 |
-
],
|
| 79006 |
-
"bits_prop": [
|
| 79007 |
-
0.25,
|
| 79008 |
-
0.75
|
| 79009 |
-
],
|
| 79010 |
-
"scale_bits": 4
|
| 79011 |
-
},
|
| 79012 |
-
"down_proj": {
|
| 79013 |
-
"group_size": {
|
| 79014 |
-
"8": 32,
|
| 79015 |
-
"6": 128,
|
| 79016 |
-
"5": 128
|
| 79017 |
-
},
|
| 79018 |
-
"bits": [
|
| 79019 |
-
8,
|
| 79020 |
-
6,
|
| 79021 |
-
5
|
| 79022 |
-
],
|
| 79023 |
-
"bits_prop": [
|
| 79024 |
-
0.05,
|
| 79025 |
-
0.1,
|
| 79026 |
-
0.85
|
| 79027 |
-
],
|
| 79028 |
-
"scale_bits": 4
|
| 79029 |
-
}
|
| 79030 |
-
},
|
| 79031 |
-
"model.layers.9.self_attn": {
|
| 79032 |
-
"accuracy": 0.9989123591221869,
|
| 79033 |
-
"total_bits": 526999040,
|
| 79034 |
-
"q_proj": {
|
| 79035 |
-
"group_size": {
|
| 79036 |
-
"8": 128
|
| 79037 |
-
},
|
| 79038 |
-
"bits": [
|
| 79039 |
-
8
|
| 79040 |
-
],
|
| 79041 |
-
"bits_prop": [
|
| 79042 |
-
1
|
| 79043 |
-
],
|
| 79044 |
-
"scale_bits": 4
|
| 79045 |
-
},
|
| 79046 |
"k_proj": {
|
| 79047 |
"group_size": {
|
| 79048 |
-
"
|
| 79049 |
},
|
| 79050 |
"bits": [
|
| 79051 |
-
|
| 79052 |
],
|
| 79053 |
"bits_prop": [
|
| 79054 |
1
|
|
@@ -79057,7 +79069,7 @@
|
|
| 79057 |
},
|
| 79058 |
"v_proj": {
|
| 79059 |
"group_size": {
|
| 79060 |
-
"8":
|
| 79061 |
},
|
| 79062 |
"bits": [
|
| 79063 |
8
|
|
@@ -79069,10 +79081,10 @@
|
|
| 79069 |
},
|
| 79070 |
"o_proj": {
|
| 79071 |
"group_size": {
|
| 79072 |
-
"
|
| 79073 |
},
|
| 79074 |
"bits": [
|
| 79075 |
-
|
| 79076 |
],
|
| 79077 |
"bits_prop": [
|
| 79078 |
1
|
|
@@ -79081,16 +79093,16 @@
|
|
| 79081 |
}
|
| 79082 |
},
|
| 79083 |
"model.layers.9.mlp": {
|
| 79084 |
-
"accuracy": 0.
|
| 79085 |
-
"total_bits":
|
| 79086 |
"gate_proj": {
|
| 79087 |
"group_size": {
|
| 79088 |
-
"
|
| 79089 |
-
"
|
| 79090 |
},
|
| 79091 |
"bits": [
|
| 79092 |
-
|
| 79093 |
-
|
| 79094 |
],
|
| 79095 |
"bits_prop": [
|
| 79096 |
0.1,
|
|
@@ -79100,45 +79112,39 @@
|
|
| 79100 |
},
|
| 79101 |
"up_proj": {
|
| 79102 |
"group_size": {
|
| 79103 |
-
"
|
| 79104 |
-
"
|
| 79105 |
},
|
| 79106 |
"bits": [
|
| 79107 |
-
|
| 79108 |
-
|
| 79109 |
],
|
| 79110 |
"bits_prop": [
|
| 79111 |
-
0.
|
| 79112 |
-
0.
|
| 79113 |
],
|
| 79114 |
"scale_bits": 4
|
| 79115 |
},
|
| 79116 |
"down_proj": {
|
| 79117 |
"group_size": {
|
| 79118 |
-
"8":
|
| 79119 |
-
"6": 32,
|
| 79120 |
-
"5": 32
|
| 79121 |
},
|
| 79122 |
"bits": [
|
| 79123 |
-
8
|
| 79124 |
-
6,
|
| 79125 |
-
5
|
| 79126 |
],
|
| 79127 |
"bits_prop": [
|
| 79128 |
-
|
| 79129 |
-
0.1,
|
| 79130 |
-
0.85
|
| 79131 |
],
|
| 79132 |
"scale_bits": 4
|
| 79133 |
}
|
| 79134 |
},
|
| 79135 |
"model.layers.10.self_attn": {
|
| 79136 |
-
"accuracy": 0.
|
| 79137 |
-
"total_bits":
|
| 79138 |
"q_proj": {
|
| 79139 |
"group_size": {
|
| 79140 |
-
"6":
|
| 79141 |
-
"5":
|
| 79142 |
},
|
| 79143 |
"bits": [
|
| 79144 |
6,
|
|
@@ -79152,8 +79158,8 @@
|
|
| 79152 |
},
|
| 79153 |
"k_proj": {
|
| 79154 |
"group_size": {
|
| 79155 |
-
"6":
|
| 79156 |
-
"5":
|
| 79157 |
},
|
| 79158 |
"bits": [
|
| 79159 |
6,
|
|
@@ -79167,7 +79173,7 @@
|
|
| 79167 |
},
|
| 79168 |
"v_proj": {
|
| 79169 |
"group_size": {
|
| 79170 |
-
"6":
|
| 79171 |
},
|
| 79172 |
"bits": [
|
| 79173 |
6
|
|
@@ -79179,8 +79185,8 @@
|
|
| 79179 |
},
|
| 79180 |
"o_proj": {
|
| 79181 |
"group_size": {
|
| 79182 |
-
"6":
|
| 79183 |
-
"5":
|
| 79184 |
},
|
| 79185 |
"bits": [
|
| 79186 |
6,
|
|
@@ -79194,90 +79200,87 @@
|
|
| 79194 |
}
|
| 79195 |
},
|
| 79196 |
"model.layers.10.mlp": {
|
| 79197 |
-
"accuracy": 0.
|
| 79198 |
-
"total_bits":
|
| 79199 |
"gate_proj": {
|
| 79200 |
"group_size": {
|
| 79201 |
-
"6":
|
| 79202 |
-
"5": 32
|
| 79203 |
},
|
| 79204 |
"bits": [
|
| 79205 |
-
6
|
| 79206 |
-
5
|
| 79207 |
],
|
| 79208 |
"bits_prop": [
|
| 79209 |
-
|
| 79210 |
-
0.9
|
| 79211 |
],
|
| 79212 |
"scale_bits": 4
|
| 79213 |
},
|
| 79214 |
"up_proj": {
|
| 79215 |
"group_size": {
|
| 79216 |
-
"6":
|
| 79217 |
-
"5": 32
|
| 79218 |
},
|
| 79219 |
"bits": [
|
| 79220 |
-
6
|
| 79221 |
-
5
|
| 79222 |
],
|
| 79223 |
"bits_prop": [
|
| 79224 |
-
|
| 79225 |
-
0.75
|
| 79226 |
],
|
| 79227 |
"scale_bits": 4
|
| 79228 |
},
|
| 79229 |
"down_proj": {
|
| 79230 |
"group_size": {
|
| 79231 |
"8": 32,
|
| 79232 |
-
"6":
|
| 79233 |
-
"5": 32
|
| 79234 |
},
|
| 79235 |
"bits": [
|
| 79236 |
8,
|
| 79237 |
-
6
|
| 79238 |
-
5
|
| 79239 |
],
|
| 79240 |
"bits_prop": [
|
| 79241 |
0.05,
|
| 79242 |
-
0.
|
| 79243 |
-
0.85
|
| 79244 |
],
|
| 79245 |
"scale_bits": 4
|
| 79246 |
}
|
| 79247 |
},
|
| 79248 |
"model.layers.11.self_attn": {
|
| 79249 |
-
"accuracy": 0.
|
| 79250 |
-
"total_bits":
|
| 79251 |
"q_proj": {
|
| 79252 |
"group_size": {
|
| 79253 |
-
"6": 32
|
|
|
|
| 79254 |
},
|
| 79255 |
"bits": [
|
| 79256 |
-
6
|
|
|
|
| 79257 |
],
|
| 79258 |
"bits_prop": [
|
| 79259 |
-
1
|
|
|
|
| 79260 |
],
|
| 79261 |
"scale_bits": 4
|
| 79262 |
},
|
| 79263 |
"k_proj": {
|
| 79264 |
"group_size": {
|
| 79265 |
-
"6": 32
|
|
|
|
| 79266 |
},
|
| 79267 |
"bits": [
|
| 79268 |
-
6
|
|
|
|
| 79269 |
],
|
| 79270 |
"bits_prop": [
|
| 79271 |
-
1
|
|
|
|
| 79272 |
],
|
| 79273 |
"scale_bits": 4
|
| 79274 |
},
|
| 79275 |
"v_proj": {
|
| 79276 |
"group_size": {
|
| 79277 |
-
"
|
| 79278 |
},
|
| 79279 |
"bits": [
|
| 79280 |
-
|
| 79281 |
],
|
| 79282 |
"bits_prop": [
|
| 79283 |
1
|
|
@@ -79286,20 +79289,23 @@
|
|
| 79286 |
},
|
| 79287 |
"o_proj": {
|
| 79288 |
"group_size": {
|
| 79289 |
-
"6": 32
|
|
|
|
| 79290 |
},
|
| 79291 |
"bits": [
|
| 79292 |
-
6
|
|
|
|
| 79293 |
],
|
| 79294 |
"bits_prop": [
|
| 79295 |
-
1
|
|
|
|
| 79296 |
],
|
| 79297 |
"scale_bits": 4
|
| 79298 |
}
|
| 79299 |
},
|
| 79300 |
"model.layers.11.mlp": {
|
| 79301 |
-
"accuracy": 0.
|
| 79302 |
-
"total_bits":
|
| 79303 |
"gate_proj": {
|
| 79304 |
"group_size": {
|
| 79305 |
"8": 128,
|
|
@@ -79332,29 +79338,26 @@
|
|
| 79332 |
},
|
| 79333 |
"down_proj": {
|
| 79334 |
"group_size": {
|
| 79335 |
-
"8": 128
|
| 79336 |
-
"6": 128
|
| 79337 |
},
|
| 79338 |
"bits": [
|
| 79339 |
-
8
|
| 79340 |
-
6
|
| 79341 |
],
|
| 79342 |
"bits_prop": [
|
| 79343 |
-
|
| 79344 |
-
0.85
|
| 79345 |
],
|
| 79346 |
"scale_bits": 4
|
| 79347 |
}
|
| 79348 |
},
|
| 79349 |
"model.layers.12.self_attn": {
|
| 79350 |
-
"accuracy": 0.
|
| 79351 |
-
"total_bits":
|
| 79352 |
"q_proj": {
|
| 79353 |
"group_size": {
|
| 79354 |
-
"
|
| 79355 |
},
|
| 79356 |
"bits": [
|
| 79357 |
-
|
| 79358 |
],
|
| 79359 |
"bits_prop": [
|
| 79360 |
1
|
|
@@ -79363,10 +79366,10 @@
|
|
| 79363 |
},
|
| 79364 |
"k_proj": {
|
| 79365 |
"group_size": {
|
| 79366 |
-
"
|
| 79367 |
},
|
| 79368 |
"bits": [
|
| 79369 |
-
|
| 79370 |
],
|
| 79371 |
"bits_prop": [
|
| 79372 |
1
|
|
@@ -79375,7 +79378,7 @@
|
|
| 79375 |
},
|
| 79376 |
"v_proj": {
|
| 79377 |
"group_size": {
|
| 79378 |
-
"8":
|
| 79379 |
},
|
| 79380 |
"bits": [
|
| 79381 |
8
|
|
@@ -79387,10 +79390,10 @@
|
|
| 79387 |
},
|
| 79388 |
"o_proj": {
|
| 79389 |
"group_size": {
|
| 79390 |
-
"
|
| 79391 |
},
|
| 79392 |
"bits": [
|
| 79393 |
-
|
| 79394 |
],
|
| 79395 |
"bits_prop": [
|
| 79396 |
1
|
|
@@ -79399,8 +79402,8 @@
|
|
| 79399 |
}
|
| 79400 |
},
|
| 79401 |
"model.layers.12.mlp": {
|
| 79402 |
-
"accuracy": 0.
|
| 79403 |
-
"total_bits":
|
| 79404 |
"gate_proj": {
|
| 79405 |
"group_size": {
|
| 79406 |
"8": 128,
|
|
@@ -79432,61 +79435,6 @@
|
|
| 79432 |
"scale_bits": 4
|
| 79433 |
},
|
| 79434 |
"down_proj": {
|
| 79435 |
-
"group_size": {
|
| 79436 |
-
"8": 128,
|
| 79437 |
-
"6": 128
|
| 79438 |
-
},
|
| 79439 |
-
"bits": [
|
| 79440 |
-
8,
|
| 79441 |
-
6
|
| 79442 |
-
],
|
| 79443 |
-
"bits_prop": [
|
| 79444 |
-
0.15,
|
| 79445 |
-
0.85
|
| 79446 |
-
],
|
| 79447 |
-
"scale_bits": 4
|
| 79448 |
-
}
|
| 79449 |
-
},
|
| 79450 |
-
"model.layers.13.self_attn": {
|
| 79451 |
-
"accuracy": 0.998663023696281,
|
| 79452 |
-
"total_bits": 526999040,
|
| 79453 |
-
"q_proj": {
|
| 79454 |
-
"group_size": {
|
| 79455 |
-
"8": 128
|
| 79456 |
-
},
|
| 79457 |
-
"bits": [
|
| 79458 |
-
8
|
| 79459 |
-
],
|
| 79460 |
-
"bits_prop": [
|
| 79461 |
-
1
|
| 79462 |
-
],
|
| 79463 |
-
"scale_bits": 4
|
| 79464 |
-
},
|
| 79465 |
-
"k_proj": {
|
| 79466 |
-
"group_size": {
|
| 79467 |
-
"8": 128
|
| 79468 |
-
},
|
| 79469 |
-
"bits": [
|
| 79470 |
-
8
|
| 79471 |
-
],
|
| 79472 |
-
"bits_prop": [
|
| 79473 |
-
1
|
| 79474 |
-
],
|
| 79475 |
-
"scale_bits": 4
|
| 79476 |
-
},
|
| 79477 |
-
"v_proj": {
|
| 79478 |
-
"group_size": {
|
| 79479 |
-
"8": 128
|
| 79480 |
-
},
|
| 79481 |
-
"bits": [
|
| 79482 |
-
8
|
| 79483 |
-
],
|
| 79484 |
-
"bits_prop": [
|
| 79485 |
-
1
|
| 79486 |
-
],
|
| 79487 |
-
"scale_bits": 4
|
| 79488 |
-
},
|
| 79489 |
-
"o_proj": {
|
| 79490 |
"group_size": {
|
| 79491 |
"8": 128
|
| 79492 |
},
|
|
@@ -79499,10 +79447,10 @@
|
|
| 79499 |
"scale_bits": 4
|
| 79500 |
}
|
| 79501 |
},
|
| 79502 |
-
"model.layers.13.
|
| 79503 |
-
"accuracy": 0.
|
| 79504 |
-
"total_bits":
|
| 79505 |
-
"
|
| 79506 |
"group_size": {
|
| 79507 |
"6": 128
|
| 79508 |
},
|
|
@@ -79514,7 +79462,7 @@
|
|
| 79514 |
],
|
| 79515 |
"scale_bits": 4
|
| 79516 |
},
|
| 79517 |
-
"
|
| 79518 |
"group_size": {
|
| 79519 |
"6": 128
|
| 79520 |
},
|
|
@@ -79526,29 +79474,10 @@
|
|
| 79526 |
],
|
| 79527 |
"scale_bits": 4
|
| 79528 |
},
|
| 79529 |
-
"
|
| 79530 |
"group_size": {
|
| 79531 |
-
"8": 32,
|
| 79532 |
"6": 128
|
| 79533 |
},
|
| 79534 |
-
"bits": [
|
| 79535 |
-
8,
|
| 79536 |
-
6
|
| 79537 |
-
],
|
| 79538 |
-
"bits_prop": [
|
| 79539 |
-
0.05,
|
| 79540 |
-
0.95
|
| 79541 |
-
],
|
| 79542 |
-
"scale_bits": 4
|
| 79543 |
-
}
|
| 79544 |
-
},
|
| 79545 |
-
"model.layers.14.self_attn": {
|
| 79546 |
-
"accuracy": 0.9961008524987847,
|
| 79547 |
-
"total_bits": 415201280,
|
| 79548 |
-
"q_proj": {
|
| 79549 |
-
"group_size": {
|
| 79550 |
-
"6": 32
|
| 79551 |
-
},
|
| 79552 |
"bits": [
|
| 79553 |
6
|
| 79554 |
],
|
|
@@ -79557,33 +79486,9 @@
|
|
| 79557 |
],
|
| 79558 |
"scale_bits": 4
|
| 79559 |
},
|
| 79560 |
-
"k_proj": {
|
| 79561 |
-
"group_size": {
|
| 79562 |
-
"6": 32
|
| 79563 |
-
},
|
| 79564 |
-
"bits": [
|
| 79565 |
-
6
|
| 79566 |
-
],
|
| 79567 |
-
"bits_prop": [
|
| 79568 |
-
1
|
| 79569 |
-
],
|
| 79570 |
-
"scale_bits": 4
|
| 79571 |
-
},
|
| 79572 |
-
"v_proj": {
|
| 79573 |
-
"group_size": {
|
| 79574 |
-
"8": 32
|
| 79575 |
-
},
|
| 79576 |
-
"bits": [
|
| 79577 |
-
8
|
| 79578 |
-
],
|
| 79579 |
-
"bits_prop": [
|
| 79580 |
-
1
|
| 79581 |
-
],
|
| 79582 |
-
"scale_bits": 4
|
| 79583 |
-
},
|
| 79584 |
"o_proj": {
|
| 79585 |
"group_size": {
|
| 79586 |
-
"6":
|
| 79587 |
},
|
| 79588 |
"bits": [
|
| 79589 |
6
|
|
@@ -79594,9 +79499,9 @@
|
|
| 79594 |
"scale_bits": 4
|
| 79595 |
}
|
| 79596 |
},
|
| 79597 |
-
"model.layers.
|
| 79598 |
-
"accuracy": 0.
|
| 79599 |
-
"total_bits":
|
| 79600 |
"gate_proj": {
|
| 79601 |
"group_size": {
|
| 79602 |
"8": 128,
|
|
@@ -79629,22 +79534,19 @@
|
|
| 79629 |
},
|
| 79630 |
"down_proj": {
|
| 79631 |
"group_size": {
|
| 79632 |
-
"8": 128
|
| 79633 |
-
"6": 128
|
| 79634 |
},
|
| 79635 |
"bits": [
|
| 79636 |
-
8
|
| 79637 |
-
6
|
| 79638 |
],
|
| 79639 |
"bits_prop": [
|
| 79640 |
-
|
| 79641 |
-
0.85
|
| 79642 |
],
|
| 79643 |
"scale_bits": 4
|
| 79644 |
}
|
| 79645 |
},
|
| 79646 |
-
"model.layers.
|
| 79647 |
-
"accuracy": 0.
|
| 79648 |
"total_bits": 415201280,
|
| 79649 |
"q_proj": {
|
| 79650 |
"group_size": {
|
|
@@ -79695,51 +79597,60 @@
|
|
| 79695 |
"scale_bits": 4
|
| 79696 |
}
|
| 79697 |
},
|
| 79698 |
-
"model.layers.
|
| 79699 |
-
"accuracy": 0.
|
| 79700 |
-
"total_bits":
|
| 79701 |
"gate_proj": {
|
| 79702 |
"group_size": {
|
| 79703 |
-
"6":
|
|
|
|
| 79704 |
},
|
| 79705 |
"bits": [
|
| 79706 |
-
6
|
|
|
|
| 79707 |
],
|
| 79708 |
"bits_prop": [
|
| 79709 |
-
1
|
|
|
|
| 79710 |
],
|
| 79711 |
"scale_bits": 4
|
| 79712 |
},
|
| 79713 |
"up_proj": {
|
| 79714 |
"group_size": {
|
| 79715 |
-
"6":
|
|
|
|
| 79716 |
},
|
| 79717 |
"bits": [
|
| 79718 |
-
6
|
|
|
|
| 79719 |
],
|
| 79720 |
"bits_prop": [
|
| 79721 |
-
|
|
|
|
| 79722 |
],
|
| 79723 |
"scale_bits": 4
|
| 79724 |
},
|
| 79725 |
"down_proj": {
|
| 79726 |
"group_size": {
|
| 79727 |
"8": 32,
|
| 79728 |
-
"6":
|
|
|
|
| 79729 |
},
|
| 79730 |
"bits": [
|
| 79731 |
8,
|
| 79732 |
-
6
|
|
|
|
| 79733 |
],
|
| 79734 |
"bits_prop": [
|
| 79735 |
0.05,
|
| 79736 |
-
0.
|
|
|
|
| 79737 |
],
|
| 79738 |
"scale_bits": 4
|
| 79739 |
}
|
| 79740 |
},
|
| 79741 |
-
"model.layers.
|
| 79742 |
-
"accuracy": 0.
|
| 79743 |
"total_bits": 526999040,
|
| 79744 |
"q_proj": {
|
| 79745 |
"group_size": {
|
|
@@ -79790,42 +79701,36 @@
|
|
| 79790 |
"scale_bits": 4
|
| 79791 |
}
|
| 79792 |
},
|
| 79793 |
-
"model.layers.
|
| 79794 |
-
"accuracy": 0.
|
| 79795 |
-
"total_bits":
|
| 79796 |
"gate_proj": {
|
| 79797 |
"group_size": {
|
| 79798 |
-
"8": 128,
|
| 79799 |
"6": 128
|
| 79800 |
},
|
| 79801 |
"bits": [
|
| 79802 |
-
8,
|
| 79803 |
6
|
| 79804 |
],
|
| 79805 |
"bits_prop": [
|
| 79806 |
-
|
| 79807 |
-
0.9
|
| 79808 |
],
|
| 79809 |
"scale_bits": 4
|
| 79810 |
},
|
| 79811 |
"up_proj": {
|
| 79812 |
"group_size": {
|
| 79813 |
-
"8": 128,
|
| 79814 |
"6": 128
|
| 79815 |
},
|
| 79816 |
"bits": [
|
| 79817 |
-
8,
|
| 79818 |
6
|
| 79819 |
],
|
| 79820 |
"bits_prop": [
|
| 79821 |
-
|
| 79822 |
-
0.9
|
| 79823 |
],
|
| 79824 |
"scale_bits": 4
|
| 79825 |
},
|
| 79826 |
"down_proj": {
|
| 79827 |
"group_size": {
|
| 79828 |
-
"8":
|
| 79829 |
"6": 128
|
| 79830 |
},
|
| 79831 |
"bits": [
|
|
@@ -79833,14 +79738,14 @@
|
|
| 79833 |
6
|
| 79834 |
],
|
| 79835 |
"bits_prop": [
|
| 79836 |
-
0.
|
| 79837 |
-
0.
|
| 79838 |
],
|
| 79839 |
"scale_bits": 4
|
| 79840 |
}
|
| 79841 |
},
|
| 79842 |
-
"model.layers.
|
| 79843 |
-
"accuracy": 0.
|
| 79844 |
"total_bits": 415201280,
|
| 79845 |
"q_proj": {
|
| 79846 |
"group_size": {
|
|
@@ -79891,8 +79796,8 @@
|
|
| 79891 |
"scale_bits": 4
|
| 79892 |
}
|
| 79893 |
},
|
| 79894 |
-
"model.layers.
|
| 79895 |
-
"accuracy": 0.
|
| 79896 |
"total_bits": 1881221440,
|
| 79897 |
"gate_proj": {
|
| 79898 |
"group_size": {
|
|
@@ -79937,15 +79842,15 @@
|
|
| 79937 |
"scale_bits": 4
|
| 79938 |
}
|
| 79939 |
},
|
| 79940 |
-
"model.layers.
|
| 79941 |
-
"accuracy": 0.
|
| 79942 |
-
"total_bits":
|
| 79943 |
"q_proj": {
|
| 79944 |
"group_size": {
|
| 79945 |
-
"
|
| 79946 |
},
|
| 79947 |
"bits": [
|
| 79948 |
-
|
| 79949 |
],
|
| 79950 |
"bits_prop": [
|
| 79951 |
1
|
|
@@ -79954,10 +79859,10 @@
|
|
| 79954 |
},
|
| 79955 |
"k_proj": {
|
| 79956 |
"group_size": {
|
| 79957 |
-
"
|
| 79958 |
},
|
| 79959 |
"bits": [
|
| 79960 |
-
|
| 79961 |
],
|
| 79962 |
"bits_prop": [
|
| 79963 |
1
|
|
@@ -79966,7 +79871,7 @@
|
|
| 79966 |
},
|
| 79967 |
"v_proj": {
|
| 79968 |
"group_size": {
|
| 79969 |
-
"8":
|
| 79970 |
},
|
| 79971 |
"bits": [
|
| 79972 |
8
|
|
@@ -79978,10 +79883,10 @@
|
|
| 79978 |
},
|
| 79979 |
"o_proj": {
|
| 79980 |
"group_size": {
|
| 79981 |
-
"
|
| 79982 |
},
|
| 79983 |
"bits": [
|
| 79984 |
-
|
| 79985 |
],
|
| 79986 |
"bits_prop": [
|
| 79987 |
1
|
|
@@ -79989,17 +79894,17 @@
|
|
| 79989 |
"scale_bits": 4
|
| 79990 |
}
|
| 79991 |
},
|
| 79992 |
-
"model.layers.
|
| 79993 |
-
"accuracy": 0.
|
| 79994 |
-
"total_bits":
|
| 79995 |
"gate_proj": {
|
| 79996 |
"group_size": {
|
| 79997 |
-
"
|
| 79998 |
-
"
|
| 79999 |
},
|
| 80000 |
"bits": [
|
| 80001 |
-
|
| 80002 |
-
|
| 80003 |
],
|
| 80004 |
"bits_prop": [
|
| 80005 |
0.1,
|
|
@@ -80009,41 +79914,44 @@
|
|
| 80009 |
},
|
| 80010 |
"up_proj": {
|
| 80011 |
"group_size": {
|
| 80012 |
-
"
|
| 80013 |
-
"
|
| 80014 |
},
|
| 80015 |
"bits": [
|
| 80016 |
-
|
| 80017 |
-
|
| 80018 |
],
|
| 80019 |
"bits_prop": [
|
| 80020 |
-
0.
|
| 80021 |
-
0.
|
| 80022 |
],
|
| 80023 |
"scale_bits": 4
|
| 80024 |
},
|
| 80025 |
"down_proj": {
|
| 80026 |
"group_size": {
|
| 80027 |
-
"8":
|
| 80028 |
-
"6":
|
|
|
|
| 80029 |
},
|
| 80030 |
"bits": [
|
| 80031 |
8,
|
| 80032 |
-
6
|
|
|
|
| 80033 |
],
|
| 80034 |
"bits_prop": [
|
| 80035 |
-
0.
|
|
|
|
| 80036 |
0.85
|
| 80037 |
],
|
| 80038 |
"scale_bits": 4
|
| 80039 |
}
|
| 80040 |
},
|
| 80041 |
-
"model.layers.
|
| 80042 |
-
"accuracy": 0.
|
| 80043 |
-
"total_bits":
|
| 80044 |
"q_proj": {
|
| 80045 |
"group_size": {
|
| 80046 |
-
"6":
|
| 80047 |
},
|
| 80048 |
"bits": [
|
| 80049 |
6
|
|
@@ -80055,7 +79963,7 @@
|
|
| 80055 |
},
|
| 80056 |
"k_proj": {
|
| 80057 |
"group_size": {
|
| 80058 |
-
"6":
|
| 80059 |
},
|
| 80060 |
"bits": [
|
| 80061 |
6
|
|
@@ -80067,10 +79975,10 @@
|
|
| 80067 |
},
|
| 80068 |
"v_proj": {
|
| 80069 |
"group_size": {
|
| 80070 |
-
"
|
| 80071 |
},
|
| 80072 |
"bits": [
|
| 80073 |
-
|
| 80074 |
],
|
| 80075 |
"bits_prop": [
|
| 80076 |
1
|
|
@@ -80079,7 +79987,7 @@
|
|
| 80079 |
},
|
| 80080 |
"o_proj": {
|
| 80081 |
"group_size": {
|
| 80082 |
-
"6":
|
| 80083 |
},
|
| 80084 |
"bits": [
|
| 80085 |
6
|
|
@@ -80090,8 +79998,8 @@
|
|
| 80090 |
"scale_bits": 4
|
| 80091 |
}
|
| 80092 |
},
|
| 80093 |
-
"model.layers.
|
| 80094 |
-
"accuracy": 0.
|
| 80095 |
"total_bits": 1670626608,
|
| 80096 |
"gate_proj": {
|
| 80097 |
"group_size": {
|
|
@@ -80133,8 +80041,8 @@
|
|
| 80133 |
"scale_bits": 4
|
| 80134 |
}
|
| 80135 |
},
|
| 80136 |
-
"model.layers.
|
| 80137 |
-
"accuracy": 0.
|
| 80138 |
"total_bits": 526999040,
|
| 80139 |
"q_proj": {
|
| 80140 |
"group_size": {
|
|
@@ -80185,8 +80093,8 @@
|
|
| 80185 |
"scale_bits": 4
|
| 80186 |
}
|
| 80187 |
},
|
| 80188 |
-
"model.layers.
|
| 80189 |
-
"accuracy": 0.
|
| 80190 |
"total_bits": 1725245760,
|
| 80191 |
"gate_proj": {
|
| 80192 |
"group_size": {
|
|
@@ -80234,39 +80142,45 @@
|
|
| 80234 |
"scale_bits": 4
|
| 80235 |
}
|
| 80236 |
},
|
| 80237 |
-
"model.layers.
|
| 80238 |
-
"accuracy": 0.
|
| 80239 |
-
"total_bits":
|
| 80240 |
"q_proj": {
|
| 80241 |
"group_size": {
|
| 80242 |
-
"6": 32
|
|
|
|
| 80243 |
},
|
| 80244 |
"bits": [
|
| 80245 |
-
6
|
|
|
|
| 80246 |
],
|
| 80247 |
"bits_prop": [
|
| 80248 |
-
1
|
|
|
|
| 80249 |
],
|
| 80250 |
"scale_bits": 4
|
| 80251 |
},
|
| 80252 |
"k_proj": {
|
| 80253 |
"group_size": {
|
| 80254 |
-
"6": 32
|
|
|
|
| 80255 |
},
|
| 80256 |
"bits": [
|
| 80257 |
-
6
|
|
|
|
| 80258 |
],
|
| 80259 |
"bits_prop": [
|
| 80260 |
-
1
|
|
|
|
| 80261 |
],
|
| 80262 |
"scale_bits": 4
|
| 80263 |
},
|
| 80264 |
"v_proj": {
|
| 80265 |
"group_size": {
|
| 80266 |
-
"
|
| 80267 |
},
|
| 80268 |
"bits": [
|
| 80269 |
-
|
| 80270 |
],
|
| 80271 |
"bits_prop": [
|
| 80272 |
1
|
|
@@ -80275,19 +80189,22 @@
|
|
| 80275 |
},
|
| 80276 |
"o_proj": {
|
| 80277 |
"group_size": {
|
| 80278 |
-
"6": 32
|
|
|
|
| 80279 |
},
|
| 80280 |
"bits": [
|
| 80281 |
-
6
|
|
|
|
| 80282 |
],
|
| 80283 |
"bits_prop": [
|
| 80284 |
-
1
|
|
|
|
| 80285 |
],
|
| 80286 |
"scale_bits": 4
|
| 80287 |
}
|
| 80288 |
},
|
| 80289 |
-
"model.layers.
|
| 80290 |
-
"accuracy": 0.
|
| 80291 |
"total_bits": 1670626608,
|
| 80292 |
"gate_proj": {
|
| 80293 |
"group_size": {
|
|
@@ -80329,8 +80246,8 @@
|
|
| 80329 |
"scale_bits": 4
|
| 80330 |
}
|
| 80331 |
},
|
| 80332 |
-
"model.layers.
|
| 80333 |
-
"accuracy": 0.
|
| 80334 |
"total_bits": 526999040,
|
| 80335 |
"q_proj": {
|
| 80336 |
"group_size": {
|
|
@@ -80381,6 +80298,104 @@
|
|
| 80381 |
"scale_bits": 4
|
| 80382 |
}
|
| 80383 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80384 |
"model.layers.22.mlp": {
|
| 80385 |
"accuracy": 0.9926852826029062,
|
| 80386 |
"total_bits": 1725245760,
|
|
@@ -80526,14 +80541,14 @@
|
|
| 80526 |
}
|
| 80527 |
},
|
| 80528 |
"model.layers.24.self_attn": {
|
| 80529 |
-
"accuracy": 0.
|
| 80530 |
-
"total_bits":
|
| 80531 |
"q_proj": {
|
| 80532 |
"group_size": {
|
| 80533 |
-
"
|
| 80534 |
},
|
| 80535 |
"bits": [
|
| 80536 |
-
|
| 80537 |
],
|
| 80538 |
"bits_prop": [
|
| 80539 |
1
|
|
@@ -80542,10 +80557,10 @@
|
|
| 80542 |
},
|
| 80543 |
"k_proj": {
|
| 80544 |
"group_size": {
|
| 80545 |
-
"
|
| 80546 |
},
|
| 80547 |
"bits": [
|
| 80548 |
-
|
| 80549 |
],
|
| 80550 |
"bits_prop": [
|
| 80551 |
1
|
|
@@ -80554,7 +80569,7 @@
|
|
| 80554 |
},
|
| 80555 |
"v_proj": {
|
| 80556 |
"group_size": {
|
| 80557 |
-
"8":
|
| 80558 |
},
|
| 80559 |
"bits": [
|
| 80560 |
8
|
|
@@ -80566,10 +80581,10 @@
|
|
| 80566 |
},
|
| 80567 |
"o_proj": {
|
| 80568 |
"group_size": {
|
| 80569 |
-
"
|
| 80570 |
},
|
| 80571 |
"bits": [
|
| 80572 |
-
|
| 80573 |
],
|
| 80574 |
"bits_prop": [
|
| 80575 |
1
|
|
@@ -80621,14 +80636,14 @@
|
|
| 80621 |
}
|
| 80622 |
},
|
| 80623 |
"model.layers.25.self_attn": {
|
| 80624 |
-
"accuracy": 0.
|
| 80625 |
-
"total_bits":
|
| 80626 |
"q_proj": {
|
| 80627 |
"group_size": {
|
| 80628 |
-
"
|
| 80629 |
},
|
| 80630 |
"bits": [
|
| 80631 |
-
|
| 80632 |
],
|
| 80633 |
"bits_prop": [
|
| 80634 |
1
|
|
@@ -80637,10 +80652,10 @@
|
|
| 80637 |
},
|
| 80638 |
"k_proj": {
|
| 80639 |
"group_size": {
|
| 80640 |
-
"
|
| 80641 |
},
|
| 80642 |
"bits": [
|
| 80643 |
-
|
| 80644 |
],
|
| 80645 |
"bits_prop": [
|
| 80646 |
1
|
|
@@ -80649,7 +80664,7 @@
|
|
| 80649 |
},
|
| 80650 |
"v_proj": {
|
| 80651 |
"group_size": {
|
| 80652 |
-
"8":
|
| 80653 |
},
|
| 80654 |
"bits": [
|
| 80655 |
8
|
|
@@ -80660,6 +80675,52 @@
|
|
| 80660 |
"scale_bits": 4
|
| 80661 |
},
|
| 80662 |
"o_proj": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80663 |
"group_size": {
|
| 80664 |
"8": 128
|
| 80665 |
},
|
|
@@ -80672,8 +80733,69 @@
|
|
| 80672 |
"scale_bits": 4
|
| 80673 |
}
|
| 80674 |
},
|
| 80675 |
-
"model.layers.
|
| 80676 |
-
"accuracy": 0.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80677 |
"total_bits": 1670626608,
|
| 80678 |
"gate_proj": {
|
| 80679 |
"group_size": {
|
|
@@ -80715,39 +80837,45 @@
|
|
| 80715 |
"scale_bits": 4
|
| 80716 |
}
|
| 80717 |
},
|
| 80718 |
-
"model.layers.
|
| 80719 |
-
"accuracy": 0.
|
| 80720 |
-
"total_bits":
|
| 80721 |
"q_proj": {
|
| 80722 |
"group_size": {
|
| 80723 |
-
"
|
|
|
|
| 80724 |
},
|
| 80725 |
"bits": [
|
| 80726 |
-
|
|
|
|
| 80727 |
],
|
| 80728 |
"bits_prop": [
|
| 80729 |
-
1
|
|
|
|
| 80730 |
],
|
| 80731 |
"scale_bits": 4
|
| 80732 |
},
|
| 80733 |
"k_proj": {
|
| 80734 |
"group_size": {
|
| 80735 |
-
"
|
|
|
|
| 80736 |
},
|
| 80737 |
"bits": [
|
| 80738 |
-
|
|
|
|
| 80739 |
],
|
| 80740 |
"bits_prop": [
|
| 80741 |
-
1
|
|
|
|
| 80742 |
],
|
| 80743 |
"scale_bits": 4
|
| 80744 |
},
|
| 80745 |
"v_proj": {
|
| 80746 |
"group_size": {
|
| 80747 |
-
"
|
| 80748 |
},
|
| 80749 |
"bits": [
|
| 80750 |
-
|
| 80751 |
],
|
| 80752 |
"bits_prop": [
|
| 80753 |
1
|
|
@@ -80756,19 +80884,22 @@
|
|
| 80756 |
},
|
| 80757 |
"o_proj": {
|
| 80758 |
"group_size": {
|
| 80759 |
-
"
|
|
|
|
| 80760 |
},
|
| 80761 |
"bits": [
|
| 80762 |
-
|
|
|
|
| 80763 |
],
|
| 80764 |
"bits_prop": [
|
| 80765 |
-
1
|
|
|
|
| 80766 |
],
|
| 80767 |
"scale_bits": 4
|
| 80768 |
}
|
| 80769 |
},
|
| 80770 |
-
"model.layers.
|
| 80771 |
-
"accuracy": 0.
|
| 80772 |
"total_bits": 1670626608,
|
| 80773 |
"gate_proj": {
|
| 80774 |
"group_size": {
|
|
@@ -80810,12 +80941,12 @@
|
|
| 80810 |
"scale_bits": 4
|
| 80811 |
}
|
| 80812 |
},
|
| 80813 |
-
"model.layers.
|
| 80814 |
-
"accuracy": 0.
|
| 80815 |
-
"total_bits":
|
| 80816 |
"q_proj": {
|
| 80817 |
"group_size": {
|
| 80818 |
-
"6":
|
| 80819 |
},
|
| 80820 |
"bits": [
|
| 80821 |
6
|
|
@@ -80827,7 +80958,7 @@
|
|
| 80827 |
},
|
| 80828 |
"k_proj": {
|
| 80829 |
"group_size": {
|
| 80830 |
-
"6":
|
| 80831 |
},
|
| 80832 |
"bits": [
|
| 80833 |
6
|
|
@@ -80839,10 +80970,10 @@
|
|
| 80839 |
},
|
| 80840 |
"v_proj": {
|
| 80841 |
"group_size": {
|
| 80842 |
-
"
|
| 80843 |
},
|
| 80844 |
"bits": [
|
| 80845 |
-
|
| 80846 |
],
|
| 80847 |
"bits_prop": [
|
| 80848 |
1
|
|
@@ -80851,7 +80982,7 @@
|
|
| 80851 |
},
|
| 80852 |
"o_proj": {
|
| 80853 |
"group_size": {
|
| 80854 |
-
"6":
|
| 80855 |
},
|
| 80856 |
"bits": [
|
| 80857 |
6
|
|
@@ -80862,54 +80993,51 @@
|
|
| 80862 |
"scale_bits": 4
|
| 80863 |
}
|
| 80864 |
},
|
| 80865 |
-
"model.layers.
|
| 80866 |
-
"accuracy": 0.
|
| 80867 |
-
"total_bits":
|
| 80868 |
"gate_proj": {
|
| 80869 |
"group_size": {
|
| 80870 |
-
"8": 128,
|
| 80871 |
"6": 128
|
| 80872 |
},
|
| 80873 |
"bits": [
|
| 80874 |
-
8,
|
| 80875 |
6
|
| 80876 |
],
|
| 80877 |
"bits_prop": [
|
| 80878 |
-
|
| 80879 |
-
0.9
|
| 80880 |
],
|
| 80881 |
"scale_bits": 4
|
| 80882 |
},
|
| 80883 |
"up_proj": {
|
| 80884 |
"group_size": {
|
| 80885 |
-
"8": 128,
|
| 80886 |
"6": 128
|
| 80887 |
},
|
| 80888 |
"bits": [
|
| 80889 |
-
8,
|
| 80890 |
6
|
| 80891 |
],
|
| 80892 |
"bits_prop": [
|
| 80893 |
-
|
| 80894 |
-
0.9
|
| 80895 |
],
|
| 80896 |
"scale_bits": 4
|
| 80897 |
},
|
| 80898 |
"down_proj": {
|
| 80899 |
"group_size": {
|
| 80900 |
-
"8":
|
|
|
|
| 80901 |
},
|
| 80902 |
"bits": [
|
| 80903 |
-
8
|
|
|
|
| 80904 |
],
|
| 80905 |
"bits_prop": [
|
| 80906 |
-
|
|
|
|
| 80907 |
],
|
| 80908 |
"scale_bits": 4
|
| 80909 |
}
|
| 80910 |
},
|
| 80911 |
-
"model.layers.
|
| 80912 |
-
"accuracy": 0.
|
| 80913 |
"total_bits": 415201280,
|
| 80914 |
"q_proj": {
|
| 80915 |
"group_size": {
|
|
@@ -80960,51 +81088,60 @@
|
|
| 80960 |
"scale_bits": 4
|
| 80961 |
}
|
| 80962 |
},
|
| 80963 |
-
"model.layers.
|
| 80964 |
-
"accuracy": 0.
|
| 80965 |
-
"total_bits":
|
| 80966 |
"gate_proj": {
|
| 80967 |
"group_size": {
|
| 80968 |
-
"6": 128
|
|
|
|
| 80969 |
},
|
| 80970 |
"bits": [
|
| 80971 |
-
6
|
|
|
|
| 80972 |
],
|
| 80973 |
"bits_prop": [
|
| 80974 |
-
1
|
|
|
|
| 80975 |
],
|
| 80976 |
"scale_bits": 4
|
| 80977 |
},
|
| 80978 |
"up_proj": {
|
| 80979 |
"group_size": {
|
| 80980 |
-
"6": 128
|
|
|
|
| 80981 |
},
|
| 80982 |
"bits": [
|
| 80983 |
-
6
|
|
|
|
| 80984 |
],
|
| 80985 |
"bits_prop": [
|
| 80986 |
-
|
|
|
|
| 80987 |
],
|
| 80988 |
"scale_bits": 4
|
| 80989 |
},
|
| 80990 |
"down_proj": {
|
| 80991 |
"group_size": {
|
| 80992 |
"8": 32,
|
| 80993 |
-
"6": 128
|
|
|
|
| 80994 |
},
|
| 80995 |
"bits": [
|
| 80996 |
8,
|
| 80997 |
-
6
|
|
|
|
| 80998 |
],
|
| 80999 |
"bits_prop": [
|
| 81000 |
0.05,
|
| 81001 |
-
0.
|
|
|
|
| 81002 |
],
|
| 81003 |
"scale_bits": 4
|
| 81004 |
}
|
| 81005 |
},
|
| 81006 |
-
"model.layers.
|
| 81007 |
-
"accuracy": 0.
|
| 81008 |
"total_bits": 415201280,
|
| 81009 |
"q_proj": {
|
| 81010 |
"group_size": {
|
|
@@ -81055,64 +81192,55 @@
|
|
| 81055 |
"scale_bits": 4
|
| 81056 |
}
|
| 81057 |
},
|
| 81058 |
-
"model.layers.
|
| 81059 |
-
"accuracy": 0.
|
| 81060 |
-
"total_bits":
|
| 81061 |
"gate_proj": {
|
| 81062 |
"group_size": {
|
| 81063 |
-
"6":
|
| 81064 |
-
"5": 32
|
| 81065 |
},
|
| 81066 |
"bits": [
|
| 81067 |
-
6
|
| 81068 |
-
5
|
| 81069 |
],
|
| 81070 |
"bits_prop": [
|
| 81071 |
-
|
| 81072 |
-
0.9
|
| 81073 |
],
|
| 81074 |
"scale_bits": 4
|
| 81075 |
},
|
| 81076 |
"up_proj": {
|
| 81077 |
"group_size": {
|
| 81078 |
-
"6":
|
| 81079 |
-
"5": 32
|
| 81080 |
},
|
| 81081 |
"bits": [
|
| 81082 |
-
6
|
| 81083 |
-
5
|
| 81084 |
],
|
| 81085 |
"bits_prop": [
|
| 81086 |
-
|
| 81087 |
-
0.75
|
| 81088 |
],
|
| 81089 |
"scale_bits": 4
|
| 81090 |
},
|
| 81091 |
"down_proj": {
|
| 81092 |
"group_size": {
|
| 81093 |
"8": 32,
|
| 81094 |
-
"6":
|
| 81095 |
-
"5": 32
|
| 81096 |
},
|
| 81097 |
"bits": [
|
| 81098 |
8,
|
| 81099 |
-
6
|
| 81100 |
-
5
|
| 81101 |
],
|
| 81102 |
"bits_prop": [
|
| 81103 |
0.05,
|
| 81104 |
-
0.
|
| 81105 |
-
0.85
|
| 81106 |
],
|
| 81107 |
"scale_bits": 4
|
| 81108 |
}
|
| 81109 |
},
|
| 81110 |
-
"model.layers.
|
| 81111 |
-
"accuracy": 0.
|
| 81112 |
-
"total_bits":
|
| 81113 |
"q_proj": {
|
| 81114 |
"group_size": {
|
| 81115 |
-
"6":
|
| 81116 |
},
|
| 81117 |
"bits": [
|
| 81118 |
6
|
|
@@ -81124,7 +81252,7 @@
|
|
| 81124 |
},
|
| 81125 |
"k_proj": {
|
| 81126 |
"group_size": {
|
| 81127 |
-
"6":
|
| 81128 |
},
|
| 81129 |
"bits": [
|
| 81130 |
6
|
|
@@ -81136,10 +81264,10 @@
|
|
| 81136 |
},
|
| 81137 |
"v_proj": {
|
| 81138 |
"group_size": {
|
| 81139 |
-
"
|
| 81140 |
},
|
| 81141 |
"bits": [
|
| 81142 |
-
|
| 81143 |
],
|
| 81144 |
"bits_prop": [
|
| 81145 |
1
|
|
@@ -81148,7 +81276,7 @@
|
|
| 81148 |
},
|
| 81149 |
"o_proj": {
|
| 81150 |
"group_size": {
|
| 81151 |
-
"6":
|
| 81152 |
},
|
| 81153 |
"bits": [
|
| 81154 |
6
|
|
@@ -81159,91 +81287,88 @@
|
|
| 81159 |
"scale_bits": 4
|
| 81160 |
}
|
| 81161 |
},
|
| 81162 |
-
"model.layers.
|
| 81163 |
-
"accuracy": 0.
|
| 81164 |
-
"total_bits":
|
| 81165 |
"gate_proj": {
|
| 81166 |
"group_size": {
|
| 81167 |
-
"6":
|
| 81168 |
-
"5": 32
|
| 81169 |
},
|
| 81170 |
"bits": [
|
| 81171 |
-
6
|
| 81172 |
-
5
|
| 81173 |
],
|
| 81174 |
"bits_prop": [
|
| 81175 |
-
|
| 81176 |
-
0.9
|
| 81177 |
],
|
| 81178 |
"scale_bits": 4
|
| 81179 |
},
|
| 81180 |
"up_proj": {
|
| 81181 |
"group_size": {
|
| 81182 |
-
"6":
|
| 81183 |
-
"5": 32
|
| 81184 |
},
|
| 81185 |
"bits": [
|
| 81186 |
-
6
|
| 81187 |
-
5
|
| 81188 |
],
|
| 81189 |
"bits_prop": [
|
| 81190 |
-
|
| 81191 |
-
0.75
|
| 81192 |
],
|
| 81193 |
"scale_bits": 4
|
| 81194 |
},
|
| 81195 |
"down_proj": {
|
| 81196 |
"group_size": {
|
| 81197 |
"8": 32,
|
| 81198 |
-
"6":
|
| 81199 |
-
"5": 32
|
| 81200 |
},
|
| 81201 |
"bits": [
|
| 81202 |
8,
|
| 81203 |
-
6
|
| 81204 |
-
5
|
| 81205 |
],
|
| 81206 |
"bits_prop": [
|
| 81207 |
0.05,
|
| 81208 |
-
0.
|
| 81209 |
-
0.85
|
| 81210 |
],
|
| 81211 |
"scale_bits": 4
|
| 81212 |
}
|
| 81213 |
},
|
| 81214 |
-
"model.layers.
|
| 81215 |
-
"accuracy": 0.
|
| 81216 |
-
"total_bits":
|
| 81217 |
"q_proj": {
|
| 81218 |
"group_size": {
|
| 81219 |
-
"6":
|
|
|
|
| 81220 |
},
|
| 81221 |
"bits": [
|
| 81222 |
-
6
|
|
|
|
| 81223 |
],
|
| 81224 |
"bits_prop": [
|
| 81225 |
-
1
|
|
|
|
| 81226 |
],
|
| 81227 |
"scale_bits": 4
|
| 81228 |
},
|
| 81229 |
"k_proj": {
|
| 81230 |
"group_size": {
|
| 81231 |
-
"6":
|
|
|
|
| 81232 |
},
|
| 81233 |
"bits": [
|
| 81234 |
-
6
|
|
|
|
| 81235 |
],
|
| 81236 |
"bits_prop": [
|
| 81237 |
-
1
|
|
|
|
| 81238 |
],
|
| 81239 |
"scale_bits": 4
|
| 81240 |
},
|
| 81241 |
"v_proj": {
|
| 81242 |
"group_size": {
|
| 81243 |
-
"
|
| 81244 |
},
|
| 81245 |
"bits": [
|
| 81246 |
-
|
| 81247 |
],
|
| 81248 |
"bits_prop": [
|
| 81249 |
1
|
|
@@ -81252,62 +81377,74 @@
|
|
| 81252 |
},
|
| 81253 |
"o_proj": {
|
| 81254 |
"group_size": {
|
| 81255 |
-
"6":
|
|
|
|
| 81256 |
},
|
| 81257 |
"bits": [
|
| 81258 |
-
6
|
|
|
|
| 81259 |
],
|
| 81260 |
"bits_prop": [
|
| 81261 |
-
1
|
|
|
|
| 81262 |
],
|
| 81263 |
"scale_bits": 4
|
| 81264 |
}
|
| 81265 |
},
|
| 81266 |
-
"model.layers.
|
| 81267 |
-
"accuracy": 0.
|
| 81268 |
-
"total_bits":
|
| 81269 |
"gate_proj": {
|
| 81270 |
"group_size": {
|
| 81271 |
-
"6": 128
|
|
|
|
| 81272 |
},
|
| 81273 |
"bits": [
|
| 81274 |
-
6
|
|
|
|
| 81275 |
],
|
| 81276 |
"bits_prop": [
|
| 81277 |
-
1
|
|
|
|
| 81278 |
],
|
| 81279 |
"scale_bits": 4
|
| 81280 |
},
|
| 81281 |
"up_proj": {
|
| 81282 |
"group_size": {
|
| 81283 |
-
"6": 128
|
|
|
|
| 81284 |
},
|
| 81285 |
"bits": [
|
| 81286 |
-
6
|
|
|
|
| 81287 |
],
|
| 81288 |
"bits_prop": [
|
| 81289 |
-
|
|
|
|
| 81290 |
],
|
| 81291 |
"scale_bits": 4
|
| 81292 |
},
|
| 81293 |
"down_proj": {
|
| 81294 |
"group_size": {
|
| 81295 |
"8": 32,
|
| 81296 |
-
"6": 128
|
|
|
|
| 81297 |
},
|
| 81298 |
"bits": [
|
| 81299 |
8,
|
| 81300 |
-
6
|
|
|
|
| 81301 |
],
|
| 81302 |
"bits_prop": [
|
| 81303 |
0.05,
|
| 81304 |
-
0.
|
|
|
|
| 81305 |
],
|
| 81306 |
"scale_bits": 4
|
| 81307 |
}
|
| 81308 |
},
|
| 81309 |
-
"model.layers.
|
| 81310 |
-
"accuracy": 0.
|
| 81311 |
"total_bits": 526999040,
|
| 81312 |
"q_proj": {
|
| 81313 |
"group_size": {
|
|
@@ -81358,36 +81495,42 @@
|
|
| 81358 |
"scale_bits": 4
|
| 81359 |
}
|
| 81360 |
},
|
| 81361 |
-
"model.layers.
|
| 81362 |
-
"accuracy": 0.
|
| 81363 |
-
"total_bits":
|
| 81364 |
"gate_proj": {
|
| 81365 |
"group_size": {
|
|
|
|
| 81366 |
"6": 128
|
| 81367 |
},
|
| 81368 |
"bits": [
|
|
|
|
| 81369 |
6
|
| 81370 |
],
|
| 81371 |
"bits_prop": [
|
| 81372 |
-
1
|
|
|
|
| 81373 |
],
|
| 81374 |
"scale_bits": 4
|
| 81375 |
},
|
| 81376 |
"up_proj": {
|
| 81377 |
"group_size": {
|
|
|
|
| 81378 |
"6": 128
|
| 81379 |
},
|
| 81380 |
"bits": [
|
|
|
|
| 81381 |
6
|
| 81382 |
],
|
| 81383 |
"bits_prop": [
|
| 81384 |
-
1
|
|
|
|
| 81385 |
],
|
| 81386 |
"scale_bits": 4
|
| 81387 |
},
|
| 81388 |
"down_proj": {
|
| 81389 |
"group_size": {
|
| 81390 |
-
"8":
|
| 81391 |
"6": 128
|
| 81392 |
},
|
| 81393 |
"bits": [
|
|
@@ -81395,45 +81538,51 @@
|
|
| 81395 |
6
|
| 81396 |
],
|
| 81397 |
"bits_prop": [
|
| 81398 |
-
0.
|
| 81399 |
-
0.
|
| 81400 |
],
|
| 81401 |
"scale_bits": 4
|
| 81402 |
}
|
| 81403 |
},
|
| 81404 |
-
"model.layers.
|
| 81405 |
-
"accuracy": 0.
|
| 81406 |
-
"total_bits":
|
| 81407 |
"q_proj": {
|
| 81408 |
"group_size": {
|
| 81409 |
-
"6": 32
|
|
|
|
| 81410 |
},
|
| 81411 |
"bits": [
|
| 81412 |
-
6
|
|
|
|
| 81413 |
],
|
| 81414 |
"bits_prop": [
|
| 81415 |
-
1
|
|
|
|
| 81416 |
],
|
| 81417 |
"scale_bits": 4
|
| 81418 |
},
|
| 81419 |
"k_proj": {
|
| 81420 |
"group_size": {
|
| 81421 |
-
"6": 32
|
|
|
|
| 81422 |
},
|
| 81423 |
"bits": [
|
| 81424 |
-
6
|
|
|
|
| 81425 |
],
|
| 81426 |
"bits_prop": [
|
| 81427 |
-
1
|
|
|
|
| 81428 |
],
|
| 81429 |
"scale_bits": 4
|
| 81430 |
},
|
| 81431 |
"v_proj": {
|
| 81432 |
"group_size": {
|
| 81433 |
-
"
|
| 81434 |
},
|
| 81435 |
"bits": [
|
| 81436 |
-
|
| 81437 |
],
|
| 81438 |
"bits_prop": [
|
| 81439 |
1
|
|
@@ -81442,108 +81591,96 @@
|
|
| 81442 |
},
|
| 81443 |
"o_proj": {
|
| 81444 |
"group_size": {
|
| 81445 |
-
"6": 32
|
|
|
|
| 81446 |
},
|
| 81447 |
"bits": [
|
| 81448 |
-
6
|
|
|
|
| 81449 |
],
|
| 81450 |
"bits_prop": [
|
| 81451 |
-
1
|
|
|
|
| 81452 |
],
|
| 81453 |
"scale_bits": 4
|
| 81454 |
}
|
| 81455 |
},
|
| 81456 |
-
"model.layers.
|
| 81457 |
-
"accuracy": 0.
|
| 81458 |
-
"total_bits":
|
| 81459 |
"gate_proj": {
|
| 81460 |
"group_size": {
|
| 81461 |
-
"6":
|
| 81462 |
-
"5": 32
|
| 81463 |
},
|
| 81464 |
"bits": [
|
| 81465 |
-
6
|
| 81466 |
-
5
|
| 81467 |
],
|
| 81468 |
"bits_prop": [
|
| 81469 |
-
|
| 81470 |
-
0.9
|
| 81471 |
],
|
| 81472 |
"scale_bits": 4
|
| 81473 |
},
|
| 81474 |
"up_proj": {
|
| 81475 |
"group_size": {
|
| 81476 |
-
"6":
|
| 81477 |
-
"5": 32
|
| 81478 |
},
|
| 81479 |
"bits": [
|
| 81480 |
-
6
|
| 81481 |
-
5
|
| 81482 |
],
|
| 81483 |
"bits_prop": [
|
| 81484 |
-
|
| 81485 |
-
0.75
|
| 81486 |
],
|
| 81487 |
"scale_bits": 4
|
| 81488 |
},
|
| 81489 |
"down_proj": {
|
| 81490 |
"group_size": {
|
| 81491 |
"8": 32,
|
| 81492 |
-
"6":
|
| 81493 |
-
"5": 32
|
| 81494 |
},
|
| 81495 |
"bits": [
|
| 81496 |
8,
|
| 81497 |
-
6
|
| 81498 |
-
5
|
| 81499 |
],
|
| 81500 |
"bits_prop": [
|
| 81501 |
0.05,
|
| 81502 |
-
0.
|
| 81503 |
-
0.85
|
| 81504 |
],
|
| 81505 |
"scale_bits": 4
|
| 81506 |
}
|
| 81507 |
},
|
| 81508 |
-
"model.layers.
|
| 81509 |
-
"accuracy": 0.
|
| 81510 |
-
"total_bits":
|
| 81511 |
"q_proj": {
|
| 81512 |
"group_size": {
|
| 81513 |
-
"
|
| 81514 |
-
"5": 32
|
| 81515 |
},
|
| 81516 |
"bits": [
|
| 81517 |
-
|
| 81518 |
-
5
|
| 81519 |
],
|
| 81520 |
"bits_prop": [
|
| 81521 |
-
|
| 81522 |
-
0.9
|
| 81523 |
],
|
| 81524 |
"scale_bits": 4
|
| 81525 |
},
|
| 81526 |
"k_proj": {
|
| 81527 |
"group_size": {
|
| 81528 |
-
"
|
| 81529 |
-
"5": 32
|
| 81530 |
},
|
| 81531 |
"bits": [
|
| 81532 |
-
|
| 81533 |
-
5
|
| 81534 |
],
|
| 81535 |
"bits_prop": [
|
| 81536 |
-
|
| 81537 |
-
0.9
|
| 81538 |
],
|
| 81539 |
"scale_bits": 4
|
| 81540 |
},
|
| 81541 |
"v_proj": {
|
| 81542 |
"group_size": {
|
| 81543 |
-
"
|
| 81544 |
},
|
| 81545 |
"bits": [
|
| 81546 |
-
|
| 81547 |
],
|
| 81548 |
"bits_prop": [
|
| 81549 |
1
|
|
@@ -81552,27 +81689,24 @@
|
|
| 81552 |
},
|
| 81553 |
"o_proj": {
|
| 81554 |
"group_size": {
|
| 81555 |
-
"
|
| 81556 |
-
"5": 32
|
| 81557 |
},
|
| 81558 |
"bits": [
|
| 81559 |
-
|
| 81560 |
-
5
|
| 81561 |
],
|
| 81562 |
"bits_prop": [
|
| 81563 |
-
|
| 81564 |
-
0.9
|
| 81565 |
],
|
| 81566 |
"scale_bits": 4
|
| 81567 |
}
|
| 81568 |
},
|
| 81569 |
-
"model.layers.
|
| 81570 |
-
"accuracy": 0.
|
| 81571 |
-
"total_bits":
|
| 81572 |
"gate_proj": {
|
| 81573 |
"group_size": {
|
| 81574 |
-
"6":
|
| 81575 |
-
"5":
|
| 81576 |
},
|
| 81577 |
"bits": [
|
| 81578 |
6,
|
|
@@ -81586,8 +81720,8 @@
|
|
| 81586 |
},
|
| 81587 |
"up_proj": {
|
| 81588 |
"group_size": {
|
| 81589 |
-
"6":
|
| 81590 |
-
"5":
|
| 81591 |
},
|
| 81592 |
"bits": [
|
| 81593 |
6,
|
|
@@ -81602,8 +81736,8 @@
|
|
| 81602 |
"down_proj": {
|
| 81603 |
"group_size": {
|
| 81604 |
"8": 32,
|
| 81605 |
-
"6":
|
| 81606 |
-
"5":
|
| 81607 |
},
|
| 81608 |
"bits": [
|
| 81609 |
8,
|
|
@@ -81618,8 +81752,8 @@
|
|
| 81618 |
"scale_bits": 4
|
| 81619 |
}
|
| 81620 |
},
|
| 81621 |
-
"model.layers.
|
| 81622 |
-
"accuracy": 0.
|
| 81623 |
"total_bits": 349009920,
|
| 81624 |
"q_proj": {
|
| 81625 |
"group_size": {
|
|
@@ -81679,108 +81813,13 @@
|
|
| 81679 |
"scale_bits": 4
|
| 81680 |
}
|
| 81681 |
},
|
| 81682 |
-
"model.layers.35.mlp": {
|
| 81683 |
-
"accuracy": 0.9944175193086267,
|
| 81684 |
-
"total_bits": 1670626608,
|
| 81685 |
-
"gate_proj": {
|
| 81686 |
-
"group_size": {
|
| 81687 |
-
"6": 128
|
| 81688 |
-
},
|
| 81689 |
-
"bits": [
|
| 81690 |
-
6
|
| 81691 |
-
],
|
| 81692 |
-
"bits_prop": [
|
| 81693 |
-
1
|
| 81694 |
-
],
|
| 81695 |
-
"scale_bits": 4
|
| 81696 |
-
},
|
| 81697 |
-
"up_proj": {
|
| 81698 |
-
"group_size": {
|
| 81699 |
-
"6": 128
|
| 81700 |
-
},
|
| 81701 |
-
"bits": [
|
| 81702 |
-
6
|
| 81703 |
-
],
|
| 81704 |
-
"bits_prop": [
|
| 81705 |
-
1
|
| 81706 |
-
],
|
| 81707 |
-
"scale_bits": 4
|
| 81708 |
-
},
|
| 81709 |
-
"down_proj": {
|
| 81710 |
-
"group_size": {
|
| 81711 |
-
"8": 32,
|
| 81712 |
-
"6": 128
|
| 81713 |
-
},
|
| 81714 |
-
"bits": [
|
| 81715 |
-
8,
|
| 81716 |
-
6
|
| 81717 |
-
],
|
| 81718 |
-
"bits_prop": [
|
| 81719 |
-
0.05,
|
| 81720 |
-
0.95
|
| 81721 |
-
],
|
| 81722 |
-
"scale_bits": 4
|
| 81723 |
-
}
|
| 81724 |
-
},
|
| 81725 |
-
"model.layers.36.self_attn": {
|
| 81726 |
-
"accuracy": 0.9982832411769778,
|
| 81727 |
-
"total_bits": 415201280,
|
| 81728 |
-
"q_proj": {
|
| 81729 |
-
"group_size": {
|
| 81730 |
-
"6": 32
|
| 81731 |
-
},
|
| 81732 |
-
"bits": [
|
| 81733 |
-
6
|
| 81734 |
-
],
|
| 81735 |
-
"bits_prop": [
|
| 81736 |
-
1
|
| 81737 |
-
],
|
| 81738 |
-
"scale_bits": 4
|
| 81739 |
-
},
|
| 81740 |
-
"k_proj": {
|
| 81741 |
-
"group_size": {
|
| 81742 |
-
"6": 32
|
| 81743 |
-
},
|
| 81744 |
-
"bits": [
|
| 81745 |
-
6
|
| 81746 |
-
],
|
| 81747 |
-
"bits_prop": [
|
| 81748 |
-
1
|
| 81749 |
-
],
|
| 81750 |
-
"scale_bits": 4
|
| 81751 |
-
},
|
| 81752 |
-
"v_proj": {
|
| 81753 |
-
"group_size": {
|
| 81754 |
-
"8": 32
|
| 81755 |
-
},
|
| 81756 |
-
"bits": [
|
| 81757 |
-
8
|
| 81758 |
-
],
|
| 81759 |
-
"bits_prop": [
|
| 81760 |
-
1
|
| 81761 |
-
],
|
| 81762 |
-
"scale_bits": 4
|
| 81763 |
-
},
|
| 81764 |
-
"o_proj": {
|
| 81765 |
-
"group_size": {
|
| 81766 |
-
"6": 32
|
| 81767 |
-
},
|
| 81768 |
-
"bits": [
|
| 81769 |
-
6
|
| 81770 |
-
],
|
| 81771 |
-
"bits_prop": [
|
| 81772 |
-
1
|
| 81773 |
-
],
|
| 81774 |
-
"scale_bits": 4
|
| 81775 |
-
}
|
| 81776 |
-
},
|
| 81777 |
"model.layers.36.mlp": {
|
| 81778 |
-
"accuracy": 0.
|
| 81779 |
-
"total_bits":
|
| 81780 |
"gate_proj": {
|
| 81781 |
"group_size": {
|
| 81782 |
-
"6":
|
| 81783 |
-
"5":
|
| 81784 |
},
|
| 81785 |
"bits": [
|
| 81786 |
6,
|
|
@@ -81794,8 +81833,8 @@
|
|
| 81794 |
},
|
| 81795 |
"up_proj": {
|
| 81796 |
"group_size": {
|
| 81797 |
-
"6":
|
| 81798 |
-
"5":
|
| 81799 |
},
|
| 81800 |
"bits": [
|
| 81801 |
6,
|
|
@@ -81810,8 +81849,8 @@
|
|
| 81810 |
"down_proj": {
|
| 81811 |
"group_size": {
|
| 81812 |
"8": 32,
|
| 81813 |
-
"6":
|
| 81814 |
-
"5":
|
| 81815 |
},
|
| 81816 |
"bits": [
|
| 81817 |
8,
|
|
@@ -81888,8 +81927,8 @@
|
|
| 81888 |
}
|
| 81889 |
},
|
| 81890 |
"model.layers.37.mlp": {
|
| 81891 |
-
"accuracy": 0.
|
| 81892 |
-
"total_bits":
|
| 81893 |
"gate_proj": {
|
| 81894 |
"group_size": {
|
| 81895 |
"8": 128,
|
|
@@ -81922,29 +81961,26 @@
|
|
| 81922 |
},
|
| 81923 |
"down_proj": {
|
| 81924 |
"group_size": {
|
| 81925 |
-
"8": 128
|
| 81926 |
-
"6": 128
|
| 81927 |
},
|
| 81928 |
"bits": [
|
| 81929 |
-
8
|
| 81930 |
-
6
|
| 81931 |
],
|
| 81932 |
"bits_prop": [
|
| 81933 |
-
|
| 81934 |
-
0.85
|
| 81935 |
],
|
| 81936 |
"scale_bits": 4
|
| 81937 |
}
|
| 81938 |
},
|
| 81939 |
"model.layers.38.self_attn": {
|
| 81940 |
-
"accuracy": 0.
|
| 81941 |
-
"total_bits":
|
| 81942 |
"q_proj": {
|
| 81943 |
"group_size": {
|
| 81944 |
-
"
|
| 81945 |
},
|
| 81946 |
"bits": [
|
| 81947 |
-
|
| 81948 |
],
|
| 81949 |
"bits_prop": [
|
| 81950 |
1
|
|
@@ -81953,10 +81989,10 @@
|
|
| 81953 |
},
|
| 81954 |
"k_proj": {
|
| 81955 |
"group_size": {
|
| 81956 |
-
"
|
| 81957 |
},
|
| 81958 |
"bits": [
|
| 81959 |
-
|
| 81960 |
],
|
| 81961 |
"bits_prop": [
|
| 81962 |
1
|
|
@@ -81965,7 +82001,7 @@
|
|
| 81965 |
},
|
| 81966 |
"v_proj": {
|
| 81967 |
"group_size": {
|
| 81968 |
-
"8":
|
| 81969 |
},
|
| 81970 |
"bits": [
|
| 81971 |
8
|
|
@@ -81977,10 +82013,10 @@
|
|
| 81977 |
},
|
| 81978 |
"o_proj": {
|
| 81979 |
"group_size": {
|
| 81980 |
-
"
|
| 81981 |
},
|
| 81982 |
"bits": [
|
| 81983 |
-
|
| 81984 |
],
|
| 81985 |
"bits_prop": [
|
| 81986 |
1
|
|
@@ -81989,12 +82025,12 @@
|
|
| 81989 |
}
|
| 81990 |
},
|
| 81991 |
"model.layers.38.mlp": {
|
| 81992 |
-
"accuracy": 0.
|
| 81993 |
-
"total_bits":
|
| 81994 |
"gate_proj": {
|
| 81995 |
"group_size": {
|
| 81996 |
-
"6":
|
| 81997 |
-
"5":
|
| 81998 |
},
|
| 81999 |
"bits": [
|
| 82000 |
6,
|
|
@@ -82008,8 +82044,8 @@
|
|
| 82008 |
},
|
| 82009 |
"up_proj": {
|
| 82010 |
"group_size": {
|
| 82011 |
-
"6":
|
| 82012 |
-
"5":
|
| 82013 |
},
|
| 82014 |
"bits": [
|
| 82015 |
6,
|
|
@@ -82024,8 +82060,8 @@
|
|
| 82024 |
"down_proj": {
|
| 82025 |
"group_size": {
|
| 82026 |
"8": 32,
|
| 82027 |
-
"6":
|
| 82028 |
-
"5":
|
| 82029 |
},
|
| 82030 |
"bits": [
|
| 82031 |
8,
|
|
@@ -82041,53 +82077,65 @@
|
|
| 82041 |
}
|
| 82042 |
},
|
| 82043 |
"model.layers.39.self_attn": {
|
| 82044 |
-
"accuracy": 0.
|
| 82045 |
-
"total_bits":
|
| 82046 |
"q_proj": {
|
| 82047 |
"group_size": {
|
|
|
|
| 82048 |
"4": 128
|
| 82049 |
},
|
| 82050 |
"bits": [
|
|
|
|
| 82051 |
4
|
| 82052 |
],
|
| 82053 |
"bits_prop": [
|
| 82054 |
-
1
|
|
|
|
| 82055 |
],
|
| 82056 |
"scale_bits": 4
|
| 82057 |
},
|
| 82058 |
"k_proj": {
|
| 82059 |
"group_size": {
|
|
|
|
| 82060 |
"4": 128
|
| 82061 |
},
|
| 82062 |
"bits": [
|
|
|
|
| 82063 |
4
|
| 82064 |
],
|
| 82065 |
"bits_prop": [
|
| 82066 |
-
1
|
|
|
|
| 82067 |
],
|
| 82068 |
"scale_bits": 4
|
| 82069 |
},
|
| 82070 |
"v_proj": {
|
| 82071 |
"group_size": {
|
| 82072 |
-
"
|
|
|
|
| 82073 |
},
|
| 82074 |
"bits": [
|
|
|
|
| 82075 |
4
|
| 82076 |
],
|
| 82077 |
"bits_prop": [
|
| 82078 |
-
1
|
|
|
|
| 82079 |
],
|
| 82080 |
"scale_bits": 4
|
| 82081 |
},
|
| 82082 |
"o_proj": {
|
| 82083 |
"group_size": {
|
|
|
|
| 82084 |
"4": 128
|
| 82085 |
},
|
| 82086 |
"bits": [
|
|
|
|
| 82087 |
4
|
| 82088 |
],
|
| 82089 |
"bits_prop": [
|
| 82090 |
-
1
|
|
|
|
| 82091 |
],
|
| 82092 |
"scale_bits": 4
|
| 82093 |
}
|
|
|
|
| 16952 |
}
|
| 16953 |
},
|
| 16954 |
{
|
| 16955 |
+
"accuracy": 0.9644066467881203,
|
| 16956 |
"total_bits": 899923248,
|
| 16957 |
"gate_proj": {
|
| 16958 |
"group_size": {
|
|
|
|
| 17056 |
}
|
| 17057 |
},
|
| 17058 |
{
|
| 17059 |
+
"accuracy": 0.9716205019503832,
|
| 17060 |
"total_bits": 995125808,
|
| 17061 |
"gate_proj": {
|
| 17062 |
"group_size": {
|
|
|
|
| 17105 |
}
|
| 17106 |
},
|
| 17107 |
{
|
| 17108 |
+
"accuracy": 0.9818738773465157,
|
| 17109 |
"total_bits": 1137906608,
|
| 17110 |
"gate_proj": {
|
| 17111 |
"group_size": {
|
|
|
|
| 17899 |
}
|
| 17900 |
},
|
| 17901 |
{
|
| 17902 |
+
"accuracy": 0.9735957626253366,
|
| 17903 |
"total_bits": 206079360,
|
| 17904 |
"q_proj": {
|
| 17905 |
"group_size": {
|
|
|
|
| 21545 |
}
|
| 21546 |
},
|
| 21547 |
{
|
| 21548 |
+
"accuracy": 0.9386431761085987,
|
| 21549 |
"total_bits": 143375360,
|
| 21550 |
"q_proj": {
|
| 21551 |
"group_size": {
|
|
|
|
| 24554 |
],
|
| 24555 |
"model.layers.12.mlp": [
|
| 24556 |
{
|
| 24557 |
+
"accuracy": 0.9032091200351715,
|
| 24558 |
"total_bits": 614790432,
|
| 24559 |
"gate_proj": {
|
| 24560 |
"group_size": {
|
|
|
|
| 24606 |
}
|
| 24607 |
},
|
| 24608 |
{
|
| 24609 |
+
"accuracy": 0.9056077301502228,
|
| 24610 |
"total_bits": 637728032,
|
| 24611 |
"gate_proj": {
|
| 24612 |
"group_size": {
|
|
|
|
| 46269 |
}
|
| 46270 |
},
|
| 46271 |
{
|
| 46272 |
+
"accuracy": 0.949563205242157,
|
| 46273 |
"total_bits": 925328640,
|
| 46274 |
"gate_proj": {
|
| 46275 |
"group_size": {
|
|
|
|
| 51868 |
],
|
| 51869 |
"model.layers.26.mlp": [
|
| 51870 |
{
|
| 51871 |
+
"accuracy": 0.8996819406747818,
|
| 51872 |
"total_bits": 614790432,
|
| 51873 |
"gate_proj": {
|
| 51874 |
"group_size": {
|
|
|
|
| 55136 |
}
|
| 55137 |
},
|
| 55138 |
{
|
| 55139 |
+
"accuracy": 0.9861949309706688,
|
| 55140 |
"total_bits": 267324160,
|
| 55141 |
"q_proj": {
|
| 55142 |
"group_size": {
|
|
|
|
| 55188 |
}
|
| 55189 |
},
|
| 55190 |
{
|
| 55191 |
+
"accuracy": 0.9868358941748738,
|
| 55192 |
"total_bits": 271022080,
|
| 55193 |
"q_proj": {
|
| 55194 |
"group_size": {
|
|
|
|
| 62992 |
}
|
| 62993 |
},
|
| 62994 |
{
|
| 62995 |
+
"accuracy": 0.9891389207914472,
|
| 62996 |
"total_bits": 271022080,
|
| 62997 |
"q_proj": {
|
| 62998 |
"group_size": {
|
|
|
|
| 64403 |
],
|
| 64404 |
"model.layers.33.self_attn": [
|
| 64405 |
{
|
| 64406 |
+
"accuracy": 0.9657053500413895,
|
| 64407 |
"total_bits": 139115520,
|
| 64408 |
"q_proj": {
|
| 64409 |
"group_size": {
|
|
|
|
| 78063 |
},
|
| 78064 |
"strategy": {
|
| 78065 |
"model.layers.0.self_attn": {
|
| 78066 |
+
"accuracy": 0.9954015873372555,
|
| 78067 |
+
"total_bits": 415201280,
|
| 78068 |
"q_proj": {
|
| 78069 |
"group_size": {
|
| 78070 |
+
"6": 32
|
| 78071 |
},
|
| 78072 |
"bits": [
|
| 78073 |
+
6
|
| 78074 |
],
|
| 78075 |
"bits_prop": [
|
| 78076 |
1
|
|
|
|
| 78079 |
},
|
| 78080 |
"k_proj": {
|
| 78081 |
"group_size": {
|
| 78082 |
+
"6": 32
|
| 78083 |
},
|
| 78084 |
"bits": [
|
| 78085 |
+
6
|
| 78086 |
],
|
| 78087 |
"bits_prop": [
|
| 78088 |
1
|
|
|
|
| 78091 |
},
|
| 78092 |
"v_proj": {
|
| 78093 |
"group_size": {
|
| 78094 |
+
"8": 32
|
| 78095 |
},
|
| 78096 |
"bits": [
|
| 78097 |
8
|
|
|
|
| 78103 |
},
|
| 78104 |
"o_proj": {
|
| 78105 |
"group_size": {
|
| 78106 |
+
"6": 32
|
| 78107 |
},
|
| 78108 |
"bits": [
|
| 78109 |
+
6
|
| 78110 |
],
|
| 78111 |
"bits_prop": [
|
| 78112 |
1
|
|
|
|
| 78167 |
}
|
| 78168 |
},
|
| 78169 |
"model.layers.1.self_attn": {
|
| 78170 |
+
"accuracy": 0.9952727472409606,
|
| 78171 |
+
"total_bits": 415201280,
|
| 78172 |
"q_proj": {
|
| 78173 |
"group_size": {
|
| 78174 |
+
"6": 32
|
| 78175 |
},
|
| 78176 |
"bits": [
|
| 78177 |
+
6
|
| 78178 |
],
|
| 78179 |
"bits_prop": [
|
| 78180 |
1
|
|
|
|
| 78183 |
},
|
| 78184 |
"k_proj": {
|
| 78185 |
"group_size": {
|
| 78186 |
+
"6": 32
|
| 78187 |
},
|
| 78188 |
"bits": [
|
| 78189 |
+
6
|
| 78190 |
],
|
| 78191 |
"bits_prop": [
|
| 78192 |
1
|
|
|
|
| 78195 |
},
|
| 78196 |
"v_proj": {
|
| 78197 |
"group_size": {
|
| 78198 |
+
"8": 32
|
| 78199 |
},
|
| 78200 |
"bits": [
|
| 78201 |
8
|
|
|
|
| 78207 |
},
|
| 78208 |
"o_proj": {
|
| 78209 |
"group_size": {
|
| 78210 |
+
"6": 32
|
| 78211 |
},
|
| 78212 |
"bits": [
|
| 78213 |
+
6
|
| 78214 |
],
|
| 78215 |
"bits_prop": [
|
| 78216 |
1
|
|
|
|
| 78271 |
}
|
| 78272 |
},
|
| 78273 |
"model.layers.2.self_attn": {
|
| 78274 |
+
"accuracy": 0.9947458985261619,
|
| 78275 |
+
"total_bits": 349009920,
|
| 78276 |
"q_proj": {
|
| 78277 |
"group_size": {
|
| 78278 |
+
"6": 32,
|
| 78279 |
+
"5": 32
|
| 78280 |
},
|
| 78281 |
"bits": [
|
| 78282 |
+
6,
|
| 78283 |
+
5
|
| 78284 |
],
|
| 78285 |
"bits_prop": [
|
| 78286 |
+
0.1,
|
| 78287 |
+
0.9
|
| 78288 |
],
|
| 78289 |
"scale_bits": 4
|
| 78290 |
},
|
| 78291 |
"k_proj": {
|
| 78292 |
"group_size": {
|
| 78293 |
+
"6": 32,
|
| 78294 |
+
"5": 32
|
| 78295 |
},
|
| 78296 |
"bits": [
|
| 78297 |
+
6,
|
| 78298 |
+
5
|
| 78299 |
],
|
| 78300 |
"bits_prop": [
|
| 78301 |
+
0.1,
|
| 78302 |
+
0.9
|
| 78303 |
],
|
| 78304 |
"scale_bits": 4
|
| 78305 |
},
|
| 78306 |
"v_proj": {
|
| 78307 |
"group_size": {
|
| 78308 |
+
"6": 32
|
| 78309 |
},
|
| 78310 |
"bits": [
|
| 78311 |
+
6
|
| 78312 |
],
|
| 78313 |
"bits_prop": [
|
| 78314 |
1
|
|
|
|
| 78317 |
},
|
| 78318 |
"o_proj": {
|
| 78319 |
"group_size": {
|
| 78320 |
+
"6": 32,
|
| 78321 |
+
"5": 32
|
| 78322 |
},
|
| 78323 |
"bits": [
|
| 78324 |
+
6,
|
| 78325 |
+
5
|
| 78326 |
],
|
| 78327 |
"bits_prop": [
|
| 78328 |
+
0.1,
|
| 78329 |
+
0.9
|
| 78330 |
],
|
| 78331 |
"scale_bits": 4
|
| 78332 |
}
|
| 78333 |
},
|
| 78334 |
"model.layers.2.mlp": {
|
| 78335 |
+
"accuracy": 0.9968946953304112,
|
| 78336 |
+
"total_bits": 1881221440,
|
| 78337 |
"gate_proj": {
|
| 78338 |
"group_size": {
|
| 78339 |
"8": 128,
|
|
|
|
| 78366 |
},
|
| 78367 |
"down_proj": {
|
| 78368 |
"group_size": {
|
| 78369 |
+
"8": 128
|
|
|
|
| 78370 |
},
|
| 78371 |
"bits": [
|
| 78372 |
+
8
|
|
|
|
| 78373 |
],
|
| 78374 |
"bits_prop": [
|
| 78375 |
+
1
|
|
|
|
| 78376 |
],
|
| 78377 |
"scale_bits": 4
|
| 78378 |
}
|
| 78379 |
},
|
| 78380 |
"model.layers.3.self_attn": {
|
| 78381 |
+
"accuracy": 0.9970399681478739,
|
| 78382 |
+
"total_bits": 349009920,
|
| 78383 |
"q_proj": {
|
| 78384 |
"group_size": {
|
| 78385 |
+
"6": 32,
|
| 78386 |
+
"5": 32
|
| 78387 |
},
|
| 78388 |
"bits": [
|
| 78389 |
6,
|
|
|
|
| 78397 |
},
|
| 78398 |
"k_proj": {
|
| 78399 |
"group_size": {
|
| 78400 |
+
"6": 32,
|
| 78401 |
+
"5": 32
|
| 78402 |
},
|
| 78403 |
"bits": [
|
| 78404 |
6,
|
|
|
|
| 78412 |
},
|
| 78413 |
"v_proj": {
|
| 78414 |
"group_size": {
|
| 78415 |
+
"6": 32
|
| 78416 |
},
|
| 78417 |
"bits": [
|
| 78418 |
6
|
|
|
|
| 78424 |
},
|
| 78425 |
"o_proj": {
|
| 78426 |
"group_size": {
|
| 78427 |
+
"6": 32,
|
| 78428 |
+
"5": 32
|
| 78429 |
},
|
| 78430 |
"bits": [
|
| 78431 |
6,
|
|
|
|
| 78662 |
}
|
| 78663 |
},
|
| 78664 |
"model.layers.5.mlp": {
|
| 78665 |
+
"accuracy": 0.9969794370699674,
|
| 78666 |
+
"total_bits": 1725245760,
|
| 78667 |
"gate_proj": {
|
| 78668 |
"group_size": {
|
| 78669 |
+
"8": 128,
|
| 78670 |
+
"6": 128
|
| 78671 |
},
|
| 78672 |
"bits": [
|
| 78673 |
+
8,
|
| 78674 |
+
6
|
| 78675 |
],
|
| 78676 |
"bits_prop": [
|
| 78677 |
0.1,
|
|
|
|
| 78681 |
},
|
| 78682 |
"up_proj": {
|
| 78683 |
"group_size": {
|
| 78684 |
+
"8": 128,
|
| 78685 |
+
"6": 128
|
| 78686 |
},
|
| 78687 |
"bits": [
|
| 78688 |
+
8,
|
| 78689 |
+
6
|
| 78690 |
],
|
| 78691 |
"bits_prop": [
|
| 78692 |
+
0.1,
|
| 78693 |
+
0.9
|
| 78694 |
],
|
| 78695 |
"scale_bits": 4
|
| 78696 |
},
|
| 78697 |
"down_proj": {
|
| 78698 |
"group_size": {
|
| 78699 |
+
"8": 128,
|
| 78700 |
+
"6": 128
|
|
|
|
| 78701 |
},
|
| 78702 |
"bits": [
|
| 78703 |
8,
|
| 78704 |
+
6
|
|
|
|
| 78705 |
],
|
| 78706 |
"bits_prop": [
|
| 78707 |
+
0.15,
|
|
|
|
| 78708 |
0.85
|
| 78709 |
],
|
| 78710 |
"scale_bits": 4
|
| 78711 |
}
|
| 78712 |
},
|
| 78713 |
"model.layers.6.self_attn": {
|
| 78714 |
+
"accuracy": 0.9961103489622474,
|
| 78715 |
+
"total_bits": 349009920,
|
| 78716 |
"q_proj": {
|
| 78717 |
"group_size": {
|
| 78718 |
+
"6": 32,
|
| 78719 |
+
"5": 32
|
| 78720 |
},
|
| 78721 |
"bits": [
|
| 78722 |
+
6,
|
| 78723 |
+
5
|
| 78724 |
],
|
| 78725 |
"bits_prop": [
|
| 78726 |
+
0.1,
|
| 78727 |
+
0.9
|
| 78728 |
],
|
| 78729 |
"scale_bits": 4
|
| 78730 |
},
|
| 78731 |
"k_proj": {
|
| 78732 |
"group_size": {
|
| 78733 |
+
"6": 32,
|
| 78734 |
+
"5": 32
|
| 78735 |
},
|
| 78736 |
"bits": [
|
| 78737 |
+
6,
|
| 78738 |
+
5
|
| 78739 |
],
|
| 78740 |
"bits_prop": [
|
| 78741 |
+
0.1,
|
| 78742 |
+
0.9
|
| 78743 |
],
|
| 78744 |
"scale_bits": 4
|
| 78745 |
},
|
| 78746 |
"v_proj": {
|
| 78747 |
"group_size": {
|
| 78748 |
+
"6": 32
|
| 78749 |
},
|
| 78750 |
"bits": [
|
| 78751 |
+
6
|
| 78752 |
],
|
| 78753 |
"bits_prop": [
|
| 78754 |
1
|
|
|
|
| 78757 |
},
|
| 78758 |
"o_proj": {
|
| 78759 |
"group_size": {
|
| 78760 |
+
"6": 32,
|
| 78761 |
+
"5": 32
|
| 78762 |
},
|
| 78763 |
"bits": [
|
| 78764 |
+
6,
|
| 78765 |
+
5
|
| 78766 |
],
|
| 78767 |
"bits_prop": [
|
| 78768 |
+
0.1,
|
| 78769 |
+
0.9
|
| 78770 |
],
|
| 78771 |
"scale_bits": 4
|
| 78772 |
}
|
|
|
|
| 78824 |
}
|
| 78825 |
},
|
| 78826 |
"model.layers.7.self_attn": {
|
| 78827 |
+
"accuracy": 0.9964394683483988,
|
| 78828 |
+
"total_bits": 395927040,
|
| 78829 |
"q_proj": {
|
| 78830 |
"group_size": {
|
| 78831 |
+
"6": 128
|
| 78832 |
},
|
| 78833 |
"bits": [
|
| 78834 |
6
|
|
|
|
| 78840 |
},
|
| 78841 |
"k_proj": {
|
| 78842 |
"group_size": {
|
| 78843 |
+
"6": 128
|
| 78844 |
},
|
| 78845 |
"bits": [
|
| 78846 |
6
|
|
|
|
| 78852 |
},
|
| 78853 |
"v_proj": {
|
| 78854 |
"group_size": {
|
| 78855 |
+
"6": 128
|
| 78856 |
},
|
| 78857 |
"bits": [
|
| 78858 |
+
6
|
| 78859 |
],
|
| 78860 |
"bits_prop": [
|
| 78861 |
1
|
|
|
|
| 78864 |
},
|
| 78865 |
"o_proj": {
|
| 78866 |
"group_size": {
|
| 78867 |
+
"6": 128
|
| 78868 |
},
|
| 78869 |
"bits": [
|
| 78870 |
6
|
|
|
|
| 78876 |
}
|
| 78877 |
},
|
| 78878 |
"model.layers.7.mlp": {
|
| 78879 |
+
"accuracy": 0.9841652419418097,
|
| 78880 |
+
"total_bits": 1170586928,
|
| 78881 |
+
"gate_proj": {
|
| 78882 |
+
"group_size": {
|
| 78883 |
+
"5": 128,
|
| 78884 |
+
"4": 128
|
| 78885 |
+
},
|
| 78886 |
+
"bits": [
|
| 78887 |
+
5,
|
| 78888 |
+
4
|
| 78889 |
+
],
|
| 78890 |
+
"bits_prop": [
|
| 78891 |
+
0.1,
|
| 78892 |
+
0.9
|
| 78893 |
+
],
|
| 78894 |
+
"scale_bits": 4
|
| 78895 |
+
},
|
| 78896 |
+
"up_proj": {
|
| 78897 |
+
"group_size": {
|
| 78898 |
+
"5": 128,
|
| 78899 |
+
"4": 128
|
| 78900 |
+
},
|
| 78901 |
+
"bits": [
|
| 78902 |
+
5,
|
| 78903 |
+
4
|
| 78904 |
+
],
|
| 78905 |
+
"bits_prop": [
|
| 78906 |
+
0.25,
|
| 78907 |
+
0.75
|
| 78908 |
+
],
|
| 78909 |
+
"scale_bits": 4
|
| 78910 |
+
},
|
| 78911 |
+
"down_proj": {
|
| 78912 |
+
"group_size": {
|
| 78913 |
+
"8": 32,
|
| 78914 |
+
"5": 128,
|
| 78915 |
+
"4": 128
|
| 78916 |
+
},
|
| 78917 |
+
"bits": [
|
| 78918 |
+
8,
|
| 78919 |
+
5,
|
| 78920 |
+
4
|
| 78921 |
+
],
|
| 78922 |
+
"bits_prop": [
|
| 78923 |
+
0.05,
|
| 78924 |
+
0.1,
|
| 78925 |
+
0.85
|
| 78926 |
+
],
|
| 78927 |
+
"scale_bits": 4
|
| 78928 |
+
}
|
| 78929 |
+
},
|
| 78930 |
+
"model.layers.8.self_attn": {
|
| 78931 |
+
"accuracy": 0.9952646759338677,
|
| 78932 |
+
"total_bits": 349009920,
|
| 78933 |
+
"q_proj": {
|
| 78934 |
+
"group_size": {
|
| 78935 |
+
"6": 32,
|
| 78936 |
+
"5": 32
|
| 78937 |
+
},
|
| 78938 |
+
"bits": [
|
| 78939 |
+
6,
|
| 78940 |
+
5
|
| 78941 |
+
],
|
| 78942 |
+
"bits_prop": [
|
| 78943 |
+
0.1,
|
| 78944 |
+
0.9
|
| 78945 |
+
],
|
| 78946 |
+
"scale_bits": 4
|
| 78947 |
+
},
|
| 78948 |
+
"k_proj": {
|
| 78949 |
+
"group_size": {
|
| 78950 |
+
"6": 32,
|
| 78951 |
+
"5": 32
|
| 78952 |
+
},
|
| 78953 |
+
"bits": [
|
| 78954 |
+
6,
|
| 78955 |
+
5
|
| 78956 |
+
],
|
| 78957 |
+
"bits_prop": [
|
| 78958 |
+
0.1,
|
| 78959 |
+
0.9
|
| 78960 |
+
],
|
| 78961 |
+
"scale_bits": 4
|
| 78962 |
+
},
|
| 78963 |
+
"v_proj": {
|
| 78964 |
+
"group_size": {
|
| 78965 |
+
"6": 32
|
| 78966 |
+
},
|
| 78967 |
+
"bits": [
|
| 78968 |
+
6
|
| 78969 |
+
],
|
| 78970 |
+
"bits_prop": [
|
| 78971 |
+
1
|
| 78972 |
+
],
|
| 78973 |
+
"scale_bits": 4
|
| 78974 |
+
},
|
| 78975 |
+
"o_proj": {
|
| 78976 |
+
"group_size": {
|
| 78977 |
+
"6": 32,
|
| 78978 |
+
"5": 32
|
| 78979 |
+
},
|
| 78980 |
+
"bits": [
|
| 78981 |
+
6,
|
| 78982 |
+
5
|
| 78983 |
+
],
|
| 78984 |
+
"bits_prop": [
|
| 78985 |
+
0.1,
|
| 78986 |
+
0.9
|
| 78987 |
+
],
|
| 78988 |
+
"scale_bits": 4
|
| 78989 |
+
}
|
| 78990 |
+
},
|
| 78991 |
+
"model.layers.8.mlp": {
|
| 78992 |
+
"accuracy": 0.9921151725575328,
|
| 78993 |
"total_bits": 1466656000,
|
| 78994 |
"gate_proj": {
|
| 78995 |
"group_size": {
|
|
|
|
| 79040 |
"scale_bits": 4
|
| 79041 |
}
|
| 79042 |
},
|
| 79043 |
+
"model.layers.9.self_attn": {
|
| 79044 |
+
"accuracy": 0.9976912483107299,
|
| 79045 |
+
"total_bits": 415201280,
|
| 79046 |
"q_proj": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79047 |
"group_size": {
|
| 79048 |
"6": 32
|
| 79049 |
},
|
|
|
|
| 79055 |
],
|
| 79056 |
"scale_bits": 4
|
| 79057 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79058 |
"k_proj": {
|
| 79059 |
"group_size": {
|
| 79060 |
+
"6": 32
|
| 79061 |
},
|
| 79062 |
"bits": [
|
| 79063 |
+
6
|
| 79064 |
],
|
| 79065 |
"bits_prop": [
|
| 79066 |
1
|
|
|
|
| 79069 |
},
|
| 79070 |
"v_proj": {
|
| 79071 |
"group_size": {
|
| 79072 |
+
"8": 32
|
| 79073 |
},
|
| 79074 |
"bits": [
|
| 79075 |
8
|
|
|
|
| 79081 |
},
|
| 79082 |
"o_proj": {
|
| 79083 |
"group_size": {
|
| 79084 |
+
"6": 32
|
| 79085 |
},
|
| 79086 |
"bits": [
|
| 79087 |
+
6
|
| 79088 |
],
|
| 79089 |
"bits_prop": [
|
| 79090 |
1
|
|
|
|
| 79093 |
}
|
| 79094 |
},
|
| 79095 |
"model.layers.9.mlp": {
|
| 79096 |
+
"accuracy": 0.995745147112757,
|
| 79097 |
+
"total_bits": 1881221440,
|
| 79098 |
"gate_proj": {
|
| 79099 |
"group_size": {
|
| 79100 |
+
"8": 128,
|
| 79101 |
+
"6": 128
|
| 79102 |
},
|
| 79103 |
"bits": [
|
| 79104 |
+
8,
|
| 79105 |
+
6
|
| 79106 |
],
|
| 79107 |
"bits_prop": [
|
| 79108 |
0.1,
|
|
|
|
| 79112 |
},
|
| 79113 |
"up_proj": {
|
| 79114 |
"group_size": {
|
| 79115 |
+
"8": 128,
|
| 79116 |
+
"6": 128
|
| 79117 |
},
|
| 79118 |
"bits": [
|
| 79119 |
+
8,
|
| 79120 |
+
6
|
| 79121 |
],
|
| 79122 |
"bits_prop": [
|
| 79123 |
+
0.1,
|
| 79124 |
+
0.9
|
| 79125 |
],
|
| 79126 |
"scale_bits": 4
|
| 79127 |
},
|
| 79128 |
"down_proj": {
|
| 79129 |
"group_size": {
|
| 79130 |
+
"8": 128
|
|
|
|
|
|
|
| 79131 |
},
|
| 79132 |
"bits": [
|
| 79133 |
+
8
|
|
|
|
|
|
|
| 79134 |
],
|
| 79135 |
"bits_prop": [
|
| 79136 |
+
1
|
|
|
|
|
|
|
| 79137 |
],
|
| 79138 |
"scale_bits": 4
|
| 79139 |
}
|
| 79140 |
},
|
| 79141 |
"model.layers.10.self_attn": {
|
| 79142 |
+
"accuracy": 0.9932247875258327,
|
| 79143 |
+
"total_bits": 342842880,
|
| 79144 |
"q_proj": {
|
| 79145 |
"group_size": {
|
| 79146 |
+
"6": 128,
|
| 79147 |
+
"5": 128
|
| 79148 |
},
|
| 79149 |
"bits": [
|
| 79150 |
6,
|
|
|
|
| 79158 |
},
|
| 79159 |
"k_proj": {
|
| 79160 |
"group_size": {
|
| 79161 |
+
"6": 128,
|
| 79162 |
+
"5": 128
|
| 79163 |
},
|
| 79164 |
"bits": [
|
| 79165 |
6,
|
|
|
|
| 79173 |
},
|
| 79174 |
"v_proj": {
|
| 79175 |
"group_size": {
|
| 79176 |
+
"6": 128
|
| 79177 |
},
|
| 79178 |
"bits": [
|
| 79179 |
6
|
|
|
|
| 79185 |
},
|
| 79186 |
"o_proj": {
|
| 79187 |
"group_size": {
|
| 79188 |
+
"6": 128,
|
| 79189 |
+
"5": 128
|
| 79190 |
},
|
| 79191 |
"bits": [
|
| 79192 |
6,
|
|
|
|
| 79200 |
}
|
| 79201 |
},
|
| 79202 |
"model.layers.10.mlp": {
|
| 79203 |
+
"accuracy": 0.9939747792668641,
|
| 79204 |
+
"total_bits": 1670626608,
|
| 79205 |
"gate_proj": {
|
| 79206 |
"group_size": {
|
| 79207 |
+
"6": 128
|
|
|
|
| 79208 |
},
|
| 79209 |
"bits": [
|
| 79210 |
+
6
|
|
|
|
| 79211 |
],
|
| 79212 |
"bits_prop": [
|
| 79213 |
+
1
|
|
|
|
| 79214 |
],
|
| 79215 |
"scale_bits": 4
|
| 79216 |
},
|
| 79217 |
"up_proj": {
|
| 79218 |
"group_size": {
|
| 79219 |
+
"6": 128
|
|
|
|
| 79220 |
},
|
| 79221 |
"bits": [
|
| 79222 |
+
6
|
|
|
|
| 79223 |
],
|
| 79224 |
"bits_prop": [
|
| 79225 |
+
1
|
|
|
|
| 79226 |
],
|
| 79227 |
"scale_bits": 4
|
| 79228 |
},
|
| 79229 |
"down_proj": {
|
| 79230 |
"group_size": {
|
| 79231 |
"8": 32,
|
| 79232 |
+
"6": 128
|
|
|
|
| 79233 |
},
|
| 79234 |
"bits": [
|
| 79235 |
8,
|
| 79236 |
+
6
|
|
|
|
| 79237 |
],
|
| 79238 |
"bits_prop": [
|
| 79239 |
0.05,
|
| 79240 |
+
0.95
|
|
|
|
| 79241 |
],
|
| 79242 |
"scale_bits": 4
|
| 79243 |
}
|
| 79244 |
},
|
| 79245 |
"model.layers.11.self_attn": {
|
| 79246 |
+
"accuracy": 0.99367392109707,
|
| 79247 |
+
"total_bits": 349009920,
|
| 79248 |
"q_proj": {
|
| 79249 |
"group_size": {
|
| 79250 |
+
"6": 32,
|
| 79251 |
+
"5": 32
|
| 79252 |
},
|
| 79253 |
"bits": [
|
| 79254 |
+
6,
|
| 79255 |
+
5
|
| 79256 |
],
|
| 79257 |
"bits_prop": [
|
| 79258 |
+
0.1,
|
| 79259 |
+
0.9
|
| 79260 |
],
|
| 79261 |
"scale_bits": 4
|
| 79262 |
},
|
| 79263 |
"k_proj": {
|
| 79264 |
"group_size": {
|
| 79265 |
+
"6": 32,
|
| 79266 |
+
"5": 32
|
| 79267 |
},
|
| 79268 |
"bits": [
|
| 79269 |
+
6,
|
| 79270 |
+
5
|
| 79271 |
],
|
| 79272 |
"bits_prop": [
|
| 79273 |
+
0.1,
|
| 79274 |
+
0.9
|
| 79275 |
],
|
| 79276 |
"scale_bits": 4
|
| 79277 |
},
|
| 79278 |
"v_proj": {
|
| 79279 |
"group_size": {
|
| 79280 |
+
"6": 32
|
| 79281 |
},
|
| 79282 |
"bits": [
|
| 79283 |
+
6
|
| 79284 |
],
|
| 79285 |
"bits_prop": [
|
| 79286 |
1
|
|
|
|
| 79289 |
},
|
| 79290 |
"o_proj": {
|
| 79291 |
"group_size": {
|
| 79292 |
+
"6": 32,
|
| 79293 |
+
"5": 32
|
| 79294 |
},
|
| 79295 |
"bits": [
|
| 79296 |
+
6,
|
| 79297 |
+
5
|
| 79298 |
],
|
| 79299 |
"bits_prop": [
|
| 79300 |
+
0.1,
|
| 79301 |
+
0.9
|
| 79302 |
],
|
| 79303 |
"scale_bits": 4
|
| 79304 |
}
|
| 79305 |
},
|
| 79306 |
"model.layers.11.mlp": {
|
| 79307 |
+
"accuracy": 0.9950231495313346,
|
| 79308 |
+
"total_bits": 1881221440,
|
| 79309 |
"gate_proj": {
|
| 79310 |
"group_size": {
|
| 79311 |
"8": 128,
|
|
|
|
| 79338 |
},
|
| 79339 |
"down_proj": {
|
| 79340 |
"group_size": {
|
| 79341 |
+
"8": 128
|
|
|
|
| 79342 |
},
|
| 79343 |
"bits": [
|
| 79344 |
+
8
|
|
|
|
| 79345 |
],
|
| 79346 |
"bits_prop": [
|
| 79347 |
+
1
|
|
|
|
| 79348 |
],
|
| 79349 |
"scale_bits": 4
|
| 79350 |
}
|
| 79351 |
},
|
| 79352 |
"model.layers.12.self_attn": {
|
| 79353 |
+
"accuracy": 0.9964596622157842,
|
| 79354 |
+
"total_bits": 415201280,
|
| 79355 |
"q_proj": {
|
| 79356 |
"group_size": {
|
| 79357 |
+
"6": 32
|
| 79358 |
},
|
| 79359 |
"bits": [
|
| 79360 |
+
6
|
| 79361 |
],
|
| 79362 |
"bits_prop": [
|
| 79363 |
1
|
|
|
|
| 79366 |
},
|
| 79367 |
"k_proj": {
|
| 79368 |
"group_size": {
|
| 79369 |
+
"6": 32
|
| 79370 |
},
|
| 79371 |
"bits": [
|
| 79372 |
+
6
|
| 79373 |
],
|
| 79374 |
"bits_prop": [
|
| 79375 |
1
|
|
|
|
| 79378 |
},
|
| 79379 |
"v_proj": {
|
| 79380 |
"group_size": {
|
| 79381 |
+
"8": 32
|
| 79382 |
},
|
| 79383 |
"bits": [
|
| 79384 |
8
|
|
|
|
| 79390 |
},
|
| 79391 |
"o_proj": {
|
| 79392 |
"group_size": {
|
| 79393 |
+
"6": 32
|
| 79394 |
},
|
| 79395 |
"bits": [
|
| 79396 |
+
6
|
| 79397 |
],
|
| 79398 |
"bits_prop": [
|
| 79399 |
1
|
|
|
|
| 79402 |
}
|
| 79403 |
},
|
| 79404 |
"model.layers.12.mlp": {
|
| 79405 |
+
"accuracy": 0.9948012055829167,
|
| 79406 |
+
"total_bits": 1881221440,
|
| 79407 |
"gate_proj": {
|
| 79408 |
"group_size": {
|
| 79409 |
"8": 128,
|
|
|
|
| 79435 |
"scale_bits": 4
|
| 79436 |
},
|
| 79437 |
"down_proj": {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79438 |
"group_size": {
|
| 79439 |
"8": 128
|
| 79440 |
},
|
|
|
|
| 79447 |
"scale_bits": 4
|
| 79448 |
}
|
| 79449 |
},
|
| 79450 |
+
"model.layers.13.self_attn": {
|
| 79451 |
+
"accuracy": 0.9948677592910826,
|
| 79452 |
+
"total_bits": 395927040,
|
| 79453 |
+
"q_proj": {
|
| 79454 |
"group_size": {
|
| 79455 |
"6": 128
|
| 79456 |
},
|
|
|
|
| 79462 |
],
|
| 79463 |
"scale_bits": 4
|
| 79464 |
},
|
| 79465 |
+
"k_proj": {
|
| 79466 |
"group_size": {
|
| 79467 |
"6": 128
|
| 79468 |
},
|
|
|
|
| 79474 |
],
|
| 79475 |
"scale_bits": 4
|
| 79476 |
},
|
| 79477 |
+
"v_proj": {
|
| 79478 |
"group_size": {
|
|
|
|
| 79479 |
"6": 128
|
| 79480 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79481 |
"bits": [
|
| 79482 |
6
|
| 79483 |
],
|
|
|
|
| 79486 |
],
|
| 79487 |
"scale_bits": 4
|
| 79488 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79489 |
"o_proj": {
|
| 79490 |
"group_size": {
|
| 79491 |
+
"6": 128
|
| 79492 |
},
|
| 79493 |
"bits": [
|
| 79494 |
6
|
|
|
|
| 79499 |
"scale_bits": 4
|
| 79500 |
}
|
| 79501 |
},
|
| 79502 |
+
"model.layers.13.mlp": {
|
| 79503 |
+
"accuracy": 0.994574514683336,
|
| 79504 |
+
"total_bits": 1881221440,
|
| 79505 |
"gate_proj": {
|
| 79506 |
"group_size": {
|
| 79507 |
"8": 128,
|
|
|
|
| 79534 |
},
|
| 79535 |
"down_proj": {
|
| 79536 |
"group_size": {
|
| 79537 |
+
"8": 128
|
|
|
|
| 79538 |
},
|
| 79539 |
"bits": [
|
| 79540 |
+
8
|
|
|
|
| 79541 |
],
|
| 79542 |
"bits_prop": [
|
| 79543 |
+
1
|
|
|
|
| 79544 |
],
|
| 79545 |
"scale_bits": 4
|
| 79546 |
}
|
| 79547 |
},
|
| 79548 |
+
"model.layers.14.self_attn": {
|
| 79549 |
+
"accuracy": 0.9961008524987847,
|
| 79550 |
"total_bits": 415201280,
|
| 79551 |
"q_proj": {
|
| 79552 |
"group_size": {
|
|
|
|
| 79597 |
"scale_bits": 4
|
| 79598 |
}
|
| 79599 |
},
|
| 79600 |
+
"model.layers.14.mlp": {
|
| 79601 |
+
"accuracy": 0.9884496573358774,
|
| 79602 |
+
"total_bits": 1466656000,
|
| 79603 |
"gate_proj": {
|
| 79604 |
"group_size": {
|
| 79605 |
+
"6": 32,
|
| 79606 |
+
"5": 32
|
| 79607 |
},
|
| 79608 |
"bits": [
|
| 79609 |
+
6,
|
| 79610 |
+
5
|
| 79611 |
],
|
| 79612 |
"bits_prop": [
|
| 79613 |
+
0.1,
|
| 79614 |
+
0.9
|
| 79615 |
],
|
| 79616 |
"scale_bits": 4
|
| 79617 |
},
|
| 79618 |
"up_proj": {
|
| 79619 |
"group_size": {
|
| 79620 |
+
"6": 32,
|
| 79621 |
+
"5": 32
|
| 79622 |
},
|
| 79623 |
"bits": [
|
| 79624 |
+
6,
|
| 79625 |
+
5
|
| 79626 |
],
|
| 79627 |
"bits_prop": [
|
| 79628 |
+
0.25,
|
| 79629 |
+
0.75
|
| 79630 |
],
|
| 79631 |
"scale_bits": 4
|
| 79632 |
},
|
| 79633 |
"down_proj": {
|
| 79634 |
"group_size": {
|
| 79635 |
"8": 32,
|
| 79636 |
+
"6": 32,
|
| 79637 |
+
"5": 32
|
| 79638 |
},
|
| 79639 |
"bits": [
|
| 79640 |
8,
|
| 79641 |
+
6,
|
| 79642 |
+
5
|
| 79643 |
],
|
| 79644 |
"bits_prop": [
|
| 79645 |
0.05,
|
| 79646 |
+
0.1,
|
| 79647 |
+
0.85
|
| 79648 |
],
|
| 79649 |
"scale_bits": 4
|
| 79650 |
}
|
| 79651 |
},
|
| 79652 |
+
"model.layers.15.self_attn": {
|
| 79653 |
+
"accuracy": 0.9985494788270444,
|
| 79654 |
"total_bits": 526999040,
|
| 79655 |
"q_proj": {
|
| 79656 |
"group_size": {
|
|
|
|
| 79701 |
"scale_bits": 4
|
| 79702 |
}
|
| 79703 |
},
|
| 79704 |
+
"model.layers.15.mlp": {
|
| 79705 |
+
"accuracy": 0.9926823936402798,
|
| 79706 |
+
"total_bits": 1670626608,
|
| 79707 |
"gate_proj": {
|
| 79708 |
"group_size": {
|
|
|
|
| 79709 |
"6": 128
|
| 79710 |
},
|
| 79711 |
"bits": [
|
|
|
|
| 79712 |
6
|
| 79713 |
],
|
| 79714 |
"bits_prop": [
|
| 79715 |
+
1
|
|
|
|
| 79716 |
],
|
| 79717 |
"scale_bits": 4
|
| 79718 |
},
|
| 79719 |
"up_proj": {
|
| 79720 |
"group_size": {
|
|
|
|
| 79721 |
"6": 128
|
| 79722 |
},
|
| 79723 |
"bits": [
|
|
|
|
| 79724 |
6
|
| 79725 |
],
|
| 79726 |
"bits_prop": [
|
| 79727 |
+
1
|
|
|
|
| 79728 |
],
|
| 79729 |
"scale_bits": 4
|
| 79730 |
},
|
| 79731 |
"down_proj": {
|
| 79732 |
"group_size": {
|
| 79733 |
+
"8": 32,
|
| 79734 |
"6": 128
|
| 79735 |
},
|
| 79736 |
"bits": [
|
|
|
|
| 79738 |
6
|
| 79739 |
],
|
| 79740 |
"bits_prop": [
|
| 79741 |
+
0.05,
|
| 79742 |
+
0.95
|
| 79743 |
],
|
| 79744 |
"scale_bits": 4
|
| 79745 |
}
|
| 79746 |
},
|
| 79747 |
+
"model.layers.16.self_attn": {
|
| 79748 |
+
"accuracy": 0.9957664785906672,
|
| 79749 |
"total_bits": 415201280,
|
| 79750 |
"q_proj": {
|
| 79751 |
"group_size": {
|
|
|
|
| 79796 |
"scale_bits": 4
|
| 79797 |
}
|
| 79798 |
},
|
| 79799 |
+
"model.layers.16.mlp": {
|
| 79800 |
+
"accuracy": 0.9944000607356429,
|
| 79801 |
"total_bits": 1881221440,
|
| 79802 |
"gate_proj": {
|
| 79803 |
"group_size": {
|
|
|
|
| 79842 |
"scale_bits": 4
|
| 79843 |
}
|
| 79844 |
},
|
| 79845 |
+
"model.layers.17.self_attn": {
|
| 79846 |
+
"accuracy": 0.9957772488705814,
|
| 79847 |
+
"total_bits": 415201280,
|
| 79848 |
"q_proj": {
|
| 79849 |
"group_size": {
|
| 79850 |
+
"6": 32
|
| 79851 |
},
|
| 79852 |
"bits": [
|
| 79853 |
+
6
|
| 79854 |
],
|
| 79855 |
"bits_prop": [
|
| 79856 |
1
|
|
|
|
| 79859 |
},
|
| 79860 |
"k_proj": {
|
| 79861 |
"group_size": {
|
| 79862 |
+
"6": 32
|
| 79863 |
},
|
| 79864 |
"bits": [
|
| 79865 |
+
6
|
| 79866 |
],
|
| 79867 |
"bits_prop": [
|
| 79868 |
1
|
|
|
|
| 79871 |
},
|
| 79872 |
"v_proj": {
|
| 79873 |
"group_size": {
|
| 79874 |
+
"8": 32
|
| 79875 |
},
|
| 79876 |
"bits": [
|
| 79877 |
8
|
|
|
|
| 79883 |
},
|
| 79884 |
"o_proj": {
|
| 79885 |
"group_size": {
|
| 79886 |
+
"6": 32
|
| 79887 |
},
|
| 79888 |
"bits": [
|
| 79889 |
+
6
|
| 79890 |
],
|
| 79891 |
"bits_prop": [
|
| 79892 |
1
|
|
|
|
| 79894 |
"scale_bits": 4
|
| 79895 |
}
|
| 79896 |
},
|
| 79897 |
+
"model.layers.17.mlp": {
|
| 79898 |
+
"accuracy": 0.987975318916142,
|
| 79899 |
+
"total_bits": 1466656000,
|
| 79900 |
"gate_proj": {
|
| 79901 |
"group_size": {
|
| 79902 |
+
"6": 32,
|
| 79903 |
+
"5": 32
|
| 79904 |
},
|
| 79905 |
"bits": [
|
| 79906 |
+
6,
|
| 79907 |
+
5
|
| 79908 |
],
|
| 79909 |
"bits_prop": [
|
| 79910 |
0.1,
|
|
|
|
| 79914 |
},
|
| 79915 |
"up_proj": {
|
| 79916 |
"group_size": {
|
| 79917 |
+
"6": 32,
|
| 79918 |
+
"5": 32
|
| 79919 |
},
|
| 79920 |
"bits": [
|
| 79921 |
+
6,
|
| 79922 |
+
5
|
| 79923 |
],
|
| 79924 |
"bits_prop": [
|
| 79925 |
+
0.25,
|
| 79926 |
+
0.75
|
| 79927 |
],
|
| 79928 |
"scale_bits": 4
|
| 79929 |
},
|
| 79930 |
"down_proj": {
|
| 79931 |
"group_size": {
|
| 79932 |
+
"8": 32,
|
| 79933 |
+
"6": 32,
|
| 79934 |
+
"5": 32
|
| 79935 |
},
|
| 79936 |
"bits": [
|
| 79937 |
8,
|
| 79938 |
+
6,
|
| 79939 |
+
5
|
| 79940 |
],
|
| 79941 |
"bits_prop": [
|
| 79942 |
+
0.05,
|
| 79943 |
+
0.1,
|
| 79944 |
0.85
|
| 79945 |
],
|
| 79946 |
"scale_bits": 4
|
| 79947 |
}
|
| 79948 |
},
|
| 79949 |
+
"model.layers.18.self_attn": {
|
| 79950 |
+
"accuracy": 0.9938682704232633,
|
| 79951 |
+
"total_bits": 395927040,
|
| 79952 |
"q_proj": {
|
| 79953 |
"group_size": {
|
| 79954 |
+
"6": 128
|
| 79955 |
},
|
| 79956 |
"bits": [
|
| 79957 |
6
|
|
|
|
| 79963 |
},
|
| 79964 |
"k_proj": {
|
| 79965 |
"group_size": {
|
| 79966 |
+
"6": 128
|
| 79967 |
},
|
| 79968 |
"bits": [
|
| 79969 |
6
|
|
|
|
| 79975 |
},
|
| 79976 |
"v_proj": {
|
| 79977 |
"group_size": {
|
| 79978 |
+
"6": 128
|
| 79979 |
},
|
| 79980 |
"bits": [
|
| 79981 |
+
6
|
| 79982 |
],
|
| 79983 |
"bits_prop": [
|
| 79984 |
1
|
|
|
|
| 79987 |
},
|
| 79988 |
"o_proj": {
|
| 79989 |
"group_size": {
|
| 79990 |
+
"6": 128
|
| 79991 |
},
|
| 79992 |
"bits": [
|
| 79993 |
6
|
|
|
|
| 79998 |
"scale_bits": 4
|
| 79999 |
}
|
| 80000 |
},
|
| 80001 |
+
"model.layers.18.mlp": {
|
| 80002 |
+
"accuracy": 0.9924099026247859,
|
| 80003 |
"total_bits": 1670626608,
|
| 80004 |
"gate_proj": {
|
| 80005 |
"group_size": {
|
|
|
|
| 80041 |
"scale_bits": 4
|
| 80042 |
}
|
| 80043 |
},
|
| 80044 |
+
"model.layers.19.self_attn": {
|
| 80045 |
+
"accuracy": 0.9984620407922193,
|
| 80046 |
"total_bits": 526999040,
|
| 80047 |
"q_proj": {
|
| 80048 |
"group_size": {
|
|
|
|
| 80093 |
"scale_bits": 4
|
| 80094 |
}
|
| 80095 |
},
|
| 80096 |
+
"model.layers.19.mlp": {
|
| 80097 |
+
"accuracy": 0.9926131763495505,
|
| 80098 |
"total_bits": 1725245760,
|
| 80099 |
"gate_proj": {
|
| 80100 |
"group_size": {
|
|
|
|
| 80142 |
"scale_bits": 4
|
| 80143 |
}
|
| 80144 |
},
|
| 80145 |
+
"model.layers.20.self_attn": {
|
| 80146 |
+
"accuracy": 0.9922737274318933,
|
| 80147 |
+
"total_bits": 349009920,
|
| 80148 |
"q_proj": {
|
| 80149 |
"group_size": {
|
| 80150 |
+
"6": 32,
|
| 80151 |
+
"5": 32
|
| 80152 |
},
|
| 80153 |
"bits": [
|
| 80154 |
+
6,
|
| 80155 |
+
5
|
| 80156 |
],
|
| 80157 |
"bits_prop": [
|
| 80158 |
+
0.1,
|
| 80159 |
+
0.9
|
| 80160 |
],
|
| 80161 |
"scale_bits": 4
|
| 80162 |
},
|
| 80163 |
"k_proj": {
|
| 80164 |
"group_size": {
|
| 80165 |
+
"6": 32,
|
| 80166 |
+
"5": 32
|
| 80167 |
},
|
| 80168 |
"bits": [
|
| 80169 |
+
6,
|
| 80170 |
+
5
|
| 80171 |
],
|
| 80172 |
"bits_prop": [
|
| 80173 |
+
0.1,
|
| 80174 |
+
0.9
|
| 80175 |
],
|
| 80176 |
"scale_bits": 4
|
| 80177 |
},
|
| 80178 |
"v_proj": {
|
| 80179 |
"group_size": {
|
| 80180 |
+
"6": 32
|
| 80181 |
},
|
| 80182 |
"bits": [
|
| 80183 |
+
6
|
| 80184 |
],
|
| 80185 |
"bits_prop": [
|
| 80186 |
1
|
|
|
|
| 80189 |
},
|
| 80190 |
"o_proj": {
|
| 80191 |
"group_size": {
|
| 80192 |
+
"6": 32,
|
| 80193 |
+
"5": 32
|
| 80194 |
},
|
| 80195 |
"bits": [
|
| 80196 |
+
6,
|
| 80197 |
+
5
|
| 80198 |
],
|
| 80199 |
"bits_prop": [
|
| 80200 |
+
0.1,
|
| 80201 |
+
0.9
|
| 80202 |
],
|
| 80203 |
"scale_bits": 4
|
| 80204 |
}
|
| 80205 |
},
|
| 80206 |
+
"model.layers.20.mlp": {
|
| 80207 |
+
"accuracy": 0.9920412814244628,
|
| 80208 |
"total_bits": 1670626608,
|
| 80209 |
"gate_proj": {
|
| 80210 |
"group_size": {
|
|
|
|
| 80246 |
"scale_bits": 4
|
| 80247 |
}
|
| 80248 |
},
|
| 80249 |
+
"model.layers.21.self_attn": {
|
| 80250 |
+
"accuracy": 0.9984692038269714,
|
| 80251 |
"total_bits": 526999040,
|
| 80252 |
"q_proj": {
|
| 80253 |
"group_size": {
|
|
|
|
| 80298 |
"scale_bits": 4
|
| 80299 |
}
|
| 80300 |
},
|
| 80301 |
+
"model.layers.21.mlp": {
|
| 80302 |
+
"accuracy": 0.9939599921926856,
|
| 80303 |
+
"total_bits": 1881221440,
|
| 80304 |
+
"gate_proj": {
|
| 80305 |
+
"group_size": {
|
| 80306 |
+
"8": 128,
|
| 80307 |
+
"6": 128
|
| 80308 |
+
},
|
| 80309 |
+
"bits": [
|
| 80310 |
+
8,
|
| 80311 |
+
6
|
| 80312 |
+
],
|
| 80313 |
+
"bits_prop": [
|
| 80314 |
+
0.1,
|
| 80315 |
+
0.9
|
| 80316 |
+
],
|
| 80317 |
+
"scale_bits": 4
|
| 80318 |
+
},
|
| 80319 |
+
"up_proj": {
|
| 80320 |
+
"group_size": {
|
| 80321 |
+
"8": 128,
|
| 80322 |
+
"6": 128
|
| 80323 |
+
},
|
| 80324 |
+
"bits": [
|
| 80325 |
+
8,
|
| 80326 |
+
6
|
| 80327 |
+
],
|
| 80328 |
+
"bits_prop": [
|
| 80329 |
+
0.1,
|
| 80330 |
+
0.9
|
| 80331 |
+
],
|
| 80332 |
+
"scale_bits": 4
|
| 80333 |
+
},
|
| 80334 |
+
"down_proj": {
|
| 80335 |
+
"group_size": {
|
| 80336 |
+
"8": 128
|
| 80337 |
+
},
|
| 80338 |
+
"bits": [
|
| 80339 |
+
8
|
| 80340 |
+
],
|
| 80341 |
+
"bits_prop": [
|
| 80342 |
+
1
|
| 80343 |
+
],
|
| 80344 |
+
"scale_bits": 4
|
| 80345 |
+
}
|
| 80346 |
+
},
|
| 80347 |
+
"model.layers.22.self_attn": {
|
| 80348 |
+
"accuracy": 0.9961780766025186,
|
| 80349 |
+
"total_bits": 415201280,
|
| 80350 |
+
"q_proj": {
|
| 80351 |
+
"group_size": {
|
| 80352 |
+
"6": 32
|
| 80353 |
+
},
|
| 80354 |
+
"bits": [
|
| 80355 |
+
6
|
| 80356 |
+
],
|
| 80357 |
+
"bits_prop": [
|
| 80358 |
+
1
|
| 80359 |
+
],
|
| 80360 |
+
"scale_bits": 4
|
| 80361 |
+
},
|
| 80362 |
+
"k_proj": {
|
| 80363 |
+
"group_size": {
|
| 80364 |
+
"6": 32
|
| 80365 |
+
},
|
| 80366 |
+
"bits": [
|
| 80367 |
+
6
|
| 80368 |
+
],
|
| 80369 |
+
"bits_prop": [
|
| 80370 |
+
1
|
| 80371 |
+
],
|
| 80372 |
+
"scale_bits": 4
|
| 80373 |
+
},
|
| 80374 |
+
"v_proj": {
|
| 80375 |
+
"group_size": {
|
| 80376 |
+
"8": 32
|
| 80377 |
+
},
|
| 80378 |
+
"bits": [
|
| 80379 |
+
8
|
| 80380 |
+
],
|
| 80381 |
+
"bits_prop": [
|
| 80382 |
+
1
|
| 80383 |
+
],
|
| 80384 |
+
"scale_bits": 4
|
| 80385 |
+
},
|
| 80386 |
+
"o_proj": {
|
| 80387 |
+
"group_size": {
|
| 80388 |
+
"6": 32
|
| 80389 |
+
},
|
| 80390 |
+
"bits": [
|
| 80391 |
+
6
|
| 80392 |
+
],
|
| 80393 |
+
"bits_prop": [
|
| 80394 |
+
1
|
| 80395 |
+
],
|
| 80396 |
+
"scale_bits": 4
|
| 80397 |
+
}
|
| 80398 |
+
},
|
| 80399 |
"model.layers.22.mlp": {
|
| 80400 |
"accuracy": 0.9926852826029062,
|
| 80401 |
"total_bits": 1725245760,
|
|
|
|
| 80541 |
}
|
| 80542 |
},
|
| 80543 |
"model.layers.24.self_attn": {
|
| 80544 |
+
"accuracy": 0.9965322115458548,
|
| 80545 |
+
"total_bits": 415201280,
|
| 80546 |
"q_proj": {
|
| 80547 |
"group_size": {
|
| 80548 |
+
"6": 32
|
| 80549 |
},
|
| 80550 |
"bits": [
|
| 80551 |
+
6
|
| 80552 |
],
|
| 80553 |
"bits_prop": [
|
| 80554 |
1
|
|
|
|
| 80557 |
},
|
| 80558 |
"k_proj": {
|
| 80559 |
"group_size": {
|
| 80560 |
+
"6": 32
|
| 80561 |
},
|
| 80562 |
"bits": [
|
| 80563 |
+
6
|
| 80564 |
],
|
| 80565 |
"bits_prop": [
|
| 80566 |
1
|
|
|
|
| 80569 |
},
|
| 80570 |
"v_proj": {
|
| 80571 |
"group_size": {
|
| 80572 |
+
"8": 32
|
| 80573 |
},
|
| 80574 |
"bits": [
|
| 80575 |
8
|
|
|
|
| 80581 |
},
|
| 80582 |
"o_proj": {
|
| 80583 |
"group_size": {
|
| 80584 |
+
"6": 32
|
| 80585 |
},
|
| 80586 |
"bits": [
|
| 80587 |
+
6
|
| 80588 |
],
|
| 80589 |
"bits_prop": [
|
| 80590 |
1
|
|
|
|
| 80636 |
}
|
| 80637 |
},
|
| 80638 |
"model.layers.25.self_attn": {
|
| 80639 |
+
"accuracy": 0.9966988901142031,
|
| 80640 |
+
"total_bits": 415201280,
|
| 80641 |
"q_proj": {
|
| 80642 |
"group_size": {
|
| 80643 |
+
"6": 32
|
| 80644 |
},
|
| 80645 |
"bits": [
|
| 80646 |
+
6
|
| 80647 |
],
|
| 80648 |
"bits_prop": [
|
| 80649 |
1
|
|
|
|
| 80652 |
},
|
| 80653 |
"k_proj": {
|
| 80654 |
"group_size": {
|
| 80655 |
+
"6": 32
|
| 80656 |
},
|
| 80657 |
"bits": [
|
| 80658 |
+
6
|
| 80659 |
],
|
| 80660 |
"bits_prop": [
|
| 80661 |
1
|
|
|
|
| 80664 |
},
|
| 80665 |
"v_proj": {
|
| 80666 |
"group_size": {
|
| 80667 |
+
"8": 32
|
| 80668 |
},
|
| 80669 |
"bits": [
|
| 80670 |
8
|
|
|
|
| 80675 |
"scale_bits": 4
|
| 80676 |
},
|
| 80677 |
"o_proj": {
|
| 80678 |
+
"group_size": {
|
| 80679 |
+
"6": 32
|
| 80680 |
+
},
|
| 80681 |
+
"bits": [
|
| 80682 |
+
6
|
| 80683 |
+
],
|
| 80684 |
+
"bits_prop": [
|
| 80685 |
+
1
|
| 80686 |
+
],
|
| 80687 |
+
"scale_bits": 4
|
| 80688 |
+
}
|
| 80689 |
+
},
|
| 80690 |
+
"model.layers.25.mlp": {
|
| 80691 |
+
"accuracy": 0.9943082556128502,
|
| 80692 |
+
"total_bits": 1881221440,
|
| 80693 |
+
"gate_proj": {
|
| 80694 |
+
"group_size": {
|
| 80695 |
+
"8": 128,
|
| 80696 |
+
"6": 128
|
| 80697 |
+
},
|
| 80698 |
+
"bits": [
|
| 80699 |
+
8,
|
| 80700 |
+
6
|
| 80701 |
+
],
|
| 80702 |
+
"bits_prop": [
|
| 80703 |
+
0.1,
|
| 80704 |
+
0.9
|
| 80705 |
+
],
|
| 80706 |
+
"scale_bits": 4
|
| 80707 |
+
},
|
| 80708 |
+
"up_proj": {
|
| 80709 |
+
"group_size": {
|
| 80710 |
+
"8": 128,
|
| 80711 |
+
"6": 128
|
| 80712 |
+
},
|
| 80713 |
+
"bits": [
|
| 80714 |
+
8,
|
| 80715 |
+
6
|
| 80716 |
+
],
|
| 80717 |
+
"bits_prop": [
|
| 80718 |
+
0.1,
|
| 80719 |
+
0.9
|
| 80720 |
+
],
|
| 80721 |
+
"scale_bits": 4
|
| 80722 |
+
},
|
| 80723 |
+
"down_proj": {
|
| 80724 |
"group_size": {
|
| 80725 |
"8": 128
|
| 80726 |
},
|
|
|
|
| 80733 |
"scale_bits": 4
|
| 80734 |
}
|
| 80735 |
},
|
| 80736 |
+
"model.layers.26.self_attn": {
|
| 80737 |
+
"accuracy": 0.9938132991082966,
|
| 80738 |
+
"total_bits": 349009920,
|
| 80739 |
+
"q_proj": {
|
| 80740 |
+
"group_size": {
|
| 80741 |
+
"6": 32,
|
| 80742 |
+
"5": 32
|
| 80743 |
+
},
|
| 80744 |
+
"bits": [
|
| 80745 |
+
6,
|
| 80746 |
+
5
|
| 80747 |
+
],
|
| 80748 |
+
"bits_prop": [
|
| 80749 |
+
0.1,
|
| 80750 |
+
0.9
|
| 80751 |
+
],
|
| 80752 |
+
"scale_bits": 4
|
| 80753 |
+
},
|
| 80754 |
+
"k_proj": {
|
| 80755 |
+
"group_size": {
|
| 80756 |
+
"6": 32,
|
| 80757 |
+
"5": 32
|
| 80758 |
+
},
|
| 80759 |
+
"bits": [
|
| 80760 |
+
6,
|
| 80761 |
+
5
|
| 80762 |
+
],
|
| 80763 |
+
"bits_prop": [
|
| 80764 |
+
0.1,
|
| 80765 |
+
0.9
|
| 80766 |
+
],
|
| 80767 |
+
"scale_bits": 4
|
| 80768 |
+
},
|
| 80769 |
+
"v_proj": {
|
| 80770 |
+
"group_size": {
|
| 80771 |
+
"6": 32
|
| 80772 |
+
},
|
| 80773 |
+
"bits": [
|
| 80774 |
+
6
|
| 80775 |
+
],
|
| 80776 |
+
"bits_prop": [
|
| 80777 |
+
1
|
| 80778 |
+
],
|
| 80779 |
+
"scale_bits": 4
|
| 80780 |
+
},
|
| 80781 |
+
"o_proj": {
|
| 80782 |
+
"group_size": {
|
| 80783 |
+
"6": 32,
|
| 80784 |
+
"5": 32
|
| 80785 |
+
},
|
| 80786 |
+
"bits": [
|
| 80787 |
+
6,
|
| 80788 |
+
5
|
| 80789 |
+
],
|
| 80790 |
+
"bits_prop": [
|
| 80791 |
+
0.1,
|
| 80792 |
+
0.9
|
| 80793 |
+
],
|
| 80794 |
+
"scale_bits": 4
|
| 80795 |
+
}
|
| 80796 |
+
},
|
| 80797 |
+
"model.layers.26.mlp": {
|
| 80798 |
+
"accuracy": 0.9929717490449548,
|
| 80799 |
"total_bits": 1670626608,
|
| 80800 |
"gate_proj": {
|
| 80801 |
"group_size": {
|
|
|
|
| 80837 |
"scale_bits": 4
|
| 80838 |
}
|
| 80839 |
},
|
| 80840 |
+
"model.layers.27.self_attn": {
|
| 80841 |
+
"accuracy": 0.9940781220793724,
|
| 80842 |
+
"total_bits": 349009920,
|
| 80843 |
"q_proj": {
|
| 80844 |
"group_size": {
|
| 80845 |
+
"6": 32,
|
| 80846 |
+
"5": 32
|
| 80847 |
},
|
| 80848 |
"bits": [
|
| 80849 |
+
6,
|
| 80850 |
+
5
|
| 80851 |
],
|
| 80852 |
"bits_prop": [
|
| 80853 |
+
0.1,
|
| 80854 |
+
0.9
|
| 80855 |
],
|
| 80856 |
"scale_bits": 4
|
| 80857 |
},
|
| 80858 |
"k_proj": {
|
| 80859 |
"group_size": {
|
| 80860 |
+
"6": 32,
|
| 80861 |
+
"5": 32
|
| 80862 |
},
|
| 80863 |
"bits": [
|
| 80864 |
+
6,
|
| 80865 |
+
5
|
| 80866 |
],
|
| 80867 |
"bits_prop": [
|
| 80868 |
+
0.1,
|
| 80869 |
+
0.9
|
| 80870 |
],
|
| 80871 |
"scale_bits": 4
|
| 80872 |
},
|
| 80873 |
"v_proj": {
|
| 80874 |
"group_size": {
|
| 80875 |
+
"6": 32
|
| 80876 |
},
|
| 80877 |
"bits": [
|
| 80878 |
+
6
|
| 80879 |
],
|
| 80880 |
"bits_prop": [
|
| 80881 |
1
|
|
|
|
| 80884 |
},
|
| 80885 |
"o_proj": {
|
| 80886 |
"group_size": {
|
| 80887 |
+
"6": 32,
|
| 80888 |
+
"5": 32
|
| 80889 |
},
|
| 80890 |
"bits": [
|
| 80891 |
+
6,
|
| 80892 |
+
5
|
| 80893 |
],
|
| 80894 |
"bits_prop": [
|
| 80895 |
+
0.1,
|
| 80896 |
+
0.9
|
| 80897 |
],
|
| 80898 |
"scale_bits": 4
|
| 80899 |
}
|
| 80900 |
},
|
| 80901 |
+
"model.layers.27.mlp": {
|
| 80902 |
+
"accuracy": 0.9933040356263518,
|
| 80903 |
"total_bits": 1670626608,
|
| 80904 |
"gate_proj": {
|
| 80905 |
"group_size": {
|
|
|
|
| 80941 |
"scale_bits": 4
|
| 80942 |
}
|
| 80943 |
},
|
| 80944 |
+
"model.layers.28.self_attn": {
|
| 80945 |
+
"accuracy": 0.9957931926473975,
|
| 80946 |
+
"total_bits": 395927040,
|
| 80947 |
"q_proj": {
|
| 80948 |
"group_size": {
|
| 80949 |
+
"6": 128
|
| 80950 |
},
|
| 80951 |
"bits": [
|
| 80952 |
6
|
|
|
|
| 80958 |
},
|
| 80959 |
"k_proj": {
|
| 80960 |
"group_size": {
|
| 80961 |
+
"6": 128
|
| 80962 |
},
|
| 80963 |
"bits": [
|
| 80964 |
6
|
|
|
|
| 80970 |
},
|
| 80971 |
"v_proj": {
|
| 80972 |
"group_size": {
|
| 80973 |
+
"6": 128
|
| 80974 |
},
|
| 80975 |
"bits": [
|
| 80976 |
+
6
|
| 80977 |
],
|
| 80978 |
"bits_prop": [
|
| 80979 |
1
|
|
|
|
| 80982 |
},
|
| 80983 |
"o_proj": {
|
| 80984 |
"group_size": {
|
| 80985 |
+
"6": 128
|
| 80986 |
},
|
| 80987 |
"bits": [
|
| 80988 |
6
|
|
|
|
| 80993 |
"scale_bits": 4
|
| 80994 |
}
|
| 80995 |
},
|
| 80996 |
+
"model.layers.28.mlp": {
|
| 80997 |
+
"accuracy": 0.9936611945740879,
|
| 80998 |
+
"total_bits": 1670626608,
|
| 80999 |
"gate_proj": {
|
| 81000 |
"group_size": {
|
|
|
|
| 81001 |
"6": 128
|
| 81002 |
},
|
| 81003 |
"bits": [
|
|
|
|
| 81004 |
6
|
| 81005 |
],
|
| 81006 |
"bits_prop": [
|
| 81007 |
+
1
|
|
|
|
| 81008 |
],
|
| 81009 |
"scale_bits": 4
|
| 81010 |
},
|
| 81011 |
"up_proj": {
|
| 81012 |
"group_size": {
|
|
|
|
| 81013 |
"6": 128
|
| 81014 |
},
|
| 81015 |
"bits": [
|
|
|
|
| 81016 |
6
|
| 81017 |
],
|
| 81018 |
"bits_prop": [
|
| 81019 |
+
1
|
|
|
|
| 81020 |
],
|
| 81021 |
"scale_bits": 4
|
| 81022 |
},
|
| 81023 |
"down_proj": {
|
| 81024 |
"group_size": {
|
| 81025 |
+
"8": 32,
|
| 81026 |
+
"6": 128
|
| 81027 |
},
|
| 81028 |
"bits": [
|
| 81029 |
+
8,
|
| 81030 |
+
6
|
| 81031 |
],
|
| 81032 |
"bits_prop": [
|
| 81033 |
+
0.05,
|
| 81034 |
+
0.95
|
| 81035 |
],
|
| 81036 |
"scale_bits": 4
|
| 81037 |
}
|
| 81038 |
},
|
| 81039 |
+
"model.layers.29.self_attn": {
|
| 81040 |
+
"accuracy": 0.997671986464411,
|
| 81041 |
"total_bits": 415201280,
|
| 81042 |
"q_proj": {
|
| 81043 |
"group_size": {
|
|
|
|
| 81088 |
"scale_bits": 4
|
| 81089 |
}
|
| 81090 |
},
|
| 81091 |
+
"model.layers.29.mlp": {
|
| 81092 |
+
"accuracy": 0.9884004453197122,
|
| 81093 |
+
"total_bits": 1441250608,
|
| 81094 |
"gate_proj": {
|
| 81095 |
"group_size": {
|
| 81096 |
+
"6": 128,
|
| 81097 |
+
"5": 128
|
| 81098 |
},
|
| 81099 |
"bits": [
|
| 81100 |
+
6,
|
| 81101 |
+
5
|
| 81102 |
],
|
| 81103 |
"bits_prop": [
|
| 81104 |
+
0.1,
|
| 81105 |
+
0.9
|
| 81106 |
],
|
| 81107 |
"scale_bits": 4
|
| 81108 |
},
|
| 81109 |
"up_proj": {
|
| 81110 |
"group_size": {
|
| 81111 |
+
"6": 128,
|
| 81112 |
+
"5": 128
|
| 81113 |
},
|
| 81114 |
"bits": [
|
| 81115 |
+
6,
|
| 81116 |
+
5
|
| 81117 |
],
|
| 81118 |
"bits_prop": [
|
| 81119 |
+
0.25,
|
| 81120 |
+
0.75
|
| 81121 |
],
|
| 81122 |
"scale_bits": 4
|
| 81123 |
},
|
| 81124 |
"down_proj": {
|
| 81125 |
"group_size": {
|
| 81126 |
"8": 32,
|
| 81127 |
+
"6": 128,
|
| 81128 |
+
"5": 128
|
| 81129 |
},
|
| 81130 |
"bits": [
|
| 81131 |
8,
|
| 81132 |
+
6,
|
| 81133 |
+
5
|
| 81134 |
],
|
| 81135 |
"bits_prop": [
|
| 81136 |
0.05,
|
| 81137 |
+
0.1,
|
| 81138 |
+
0.85
|
| 81139 |
],
|
| 81140 |
"scale_bits": 4
|
| 81141 |
}
|
| 81142 |
},
|
| 81143 |
+
"model.layers.30.self_attn": {
|
| 81144 |
+
"accuracy": 0.9975796616636217,
|
| 81145 |
"total_bits": 415201280,
|
| 81146 |
"q_proj": {
|
| 81147 |
"group_size": {
|
|
|
|
| 81192 |
"scale_bits": 4
|
| 81193 |
}
|
| 81194 |
},
|
| 81195 |
+
"model.layers.30.mlp": {
|
| 81196 |
+
"accuracy": 0.9939606585539877,
|
| 81197 |
+
"total_bits": 1670626608,
|
| 81198 |
"gate_proj": {
|
| 81199 |
"group_size": {
|
| 81200 |
+
"6": 128
|
|
|
|
| 81201 |
},
|
| 81202 |
"bits": [
|
| 81203 |
+
6
|
|
|
|
| 81204 |
],
|
| 81205 |
"bits_prop": [
|
| 81206 |
+
1
|
|
|
|
| 81207 |
],
|
| 81208 |
"scale_bits": 4
|
| 81209 |
},
|
| 81210 |
"up_proj": {
|
| 81211 |
"group_size": {
|
| 81212 |
+
"6": 128
|
|
|
|
| 81213 |
},
|
| 81214 |
"bits": [
|
| 81215 |
+
6
|
|
|
|
| 81216 |
],
|
| 81217 |
"bits_prop": [
|
| 81218 |
+
1
|
|
|
|
| 81219 |
],
|
| 81220 |
"scale_bits": 4
|
| 81221 |
},
|
| 81222 |
"down_proj": {
|
| 81223 |
"group_size": {
|
| 81224 |
"8": 32,
|
| 81225 |
+
"6": 128
|
|
|
|
| 81226 |
},
|
| 81227 |
"bits": [
|
| 81228 |
8,
|
| 81229 |
+
6
|
|
|
|
| 81230 |
],
|
| 81231 |
"bits_prop": [
|
| 81232 |
0.05,
|
| 81233 |
+
0.95
|
|
|
|
| 81234 |
],
|
| 81235 |
"scale_bits": 4
|
| 81236 |
}
|
| 81237 |
},
|
| 81238 |
+
"model.layers.31.self_attn": {
|
| 81239 |
+
"accuracy": 0.997897554654628,
|
| 81240 |
+
"total_bits": 415201280,
|
| 81241 |
"q_proj": {
|
| 81242 |
"group_size": {
|
| 81243 |
+
"6": 32
|
| 81244 |
},
|
| 81245 |
"bits": [
|
| 81246 |
6
|
|
|
|
| 81252 |
},
|
| 81253 |
"k_proj": {
|
| 81254 |
"group_size": {
|
| 81255 |
+
"6": 32
|
| 81256 |
},
|
| 81257 |
"bits": [
|
| 81258 |
6
|
|
|
|
| 81264 |
},
|
| 81265 |
"v_proj": {
|
| 81266 |
"group_size": {
|
| 81267 |
+
"8": 32
|
| 81268 |
},
|
| 81269 |
"bits": [
|
| 81270 |
+
8
|
| 81271 |
],
|
| 81272 |
"bits_prop": [
|
| 81273 |
1
|
|
|
|
| 81276 |
},
|
| 81277 |
"o_proj": {
|
| 81278 |
"group_size": {
|
| 81279 |
+
"6": 32
|
| 81280 |
},
|
| 81281 |
"bits": [
|
| 81282 |
6
|
|
|
|
| 81287 |
"scale_bits": 4
|
| 81288 |
}
|
| 81289 |
},
|
| 81290 |
+
"model.layers.31.mlp": {
|
| 81291 |
+
"accuracy": 0.9940744298510253,
|
| 81292 |
+
"total_bits": 1670626608,
|
| 81293 |
"gate_proj": {
|
| 81294 |
"group_size": {
|
| 81295 |
+
"6": 128
|
|
|
|
| 81296 |
},
|
| 81297 |
"bits": [
|
| 81298 |
+
6
|
|
|
|
| 81299 |
],
|
| 81300 |
"bits_prop": [
|
| 81301 |
+
1
|
|
|
|
| 81302 |
],
|
| 81303 |
"scale_bits": 4
|
| 81304 |
},
|
| 81305 |
"up_proj": {
|
| 81306 |
"group_size": {
|
| 81307 |
+
"6": 128
|
|
|
|
| 81308 |
},
|
| 81309 |
"bits": [
|
| 81310 |
+
6
|
|
|
|
| 81311 |
],
|
| 81312 |
"bits_prop": [
|
| 81313 |
+
1
|
|
|
|
| 81314 |
],
|
| 81315 |
"scale_bits": 4
|
| 81316 |
},
|
| 81317 |
"down_proj": {
|
| 81318 |
"group_size": {
|
| 81319 |
"8": 32,
|
| 81320 |
+
"6": 128
|
|
|
|
| 81321 |
},
|
| 81322 |
"bits": [
|
| 81323 |
8,
|
| 81324 |
+
6
|
|
|
|
| 81325 |
],
|
| 81326 |
"bits_prop": [
|
| 81327 |
0.05,
|
| 81328 |
+
0.95
|
|
|
|
| 81329 |
],
|
| 81330 |
"scale_bits": 4
|
| 81331 |
}
|
| 81332 |
},
|
| 81333 |
+
"model.layers.32.self_attn": {
|
| 81334 |
+
"accuracy": 0.9947752063162625,
|
| 81335 |
+
"total_bits": 342842880,
|
| 81336 |
"q_proj": {
|
| 81337 |
"group_size": {
|
| 81338 |
+
"6": 128,
|
| 81339 |
+
"5": 128
|
| 81340 |
},
|
| 81341 |
"bits": [
|
| 81342 |
+
6,
|
| 81343 |
+
5
|
| 81344 |
],
|
| 81345 |
"bits_prop": [
|
| 81346 |
+
0.1,
|
| 81347 |
+
0.9
|
| 81348 |
],
|
| 81349 |
"scale_bits": 4
|
| 81350 |
},
|
| 81351 |
"k_proj": {
|
| 81352 |
"group_size": {
|
| 81353 |
+
"6": 128,
|
| 81354 |
+
"5": 128
|
| 81355 |
},
|
| 81356 |
"bits": [
|
| 81357 |
+
6,
|
| 81358 |
+
5
|
| 81359 |
],
|
| 81360 |
"bits_prop": [
|
| 81361 |
+
0.1,
|
| 81362 |
+
0.9
|
| 81363 |
],
|
| 81364 |
"scale_bits": 4
|
| 81365 |
},
|
| 81366 |
"v_proj": {
|
| 81367 |
"group_size": {
|
| 81368 |
+
"6": 128
|
| 81369 |
},
|
| 81370 |
"bits": [
|
| 81371 |
+
6
|
| 81372 |
],
|
| 81373 |
"bits_prop": [
|
| 81374 |
1
|
|
|
|
| 81377 |
},
|
| 81378 |
"o_proj": {
|
| 81379 |
"group_size": {
|
| 81380 |
+
"6": 128,
|
| 81381 |
+
"5": 128
|
| 81382 |
},
|
| 81383 |
"bits": [
|
| 81384 |
+
6,
|
| 81385 |
+
5
|
| 81386 |
],
|
| 81387 |
"bits_prop": [
|
| 81388 |
+
0.1,
|
| 81389 |
+
0.9
|
| 81390 |
],
|
| 81391 |
"scale_bits": 4
|
| 81392 |
}
|
| 81393 |
},
|
| 81394 |
+
"model.layers.32.mlp": {
|
| 81395 |
+
"accuracy": 0.9891814850270748,
|
| 81396 |
+
"total_bits": 1441250608,
|
| 81397 |
"gate_proj": {
|
| 81398 |
"group_size": {
|
| 81399 |
+
"6": 128,
|
| 81400 |
+
"5": 128
|
| 81401 |
},
|
| 81402 |
"bits": [
|
| 81403 |
+
6,
|
| 81404 |
+
5
|
| 81405 |
],
|
| 81406 |
"bits_prop": [
|
| 81407 |
+
0.1,
|
| 81408 |
+
0.9
|
| 81409 |
],
|
| 81410 |
"scale_bits": 4
|
| 81411 |
},
|
| 81412 |
"up_proj": {
|
| 81413 |
"group_size": {
|
| 81414 |
+
"6": 128,
|
| 81415 |
+
"5": 128
|
| 81416 |
},
|
| 81417 |
"bits": [
|
| 81418 |
+
6,
|
| 81419 |
+
5
|
| 81420 |
],
|
| 81421 |
"bits_prop": [
|
| 81422 |
+
0.25,
|
| 81423 |
+
0.75
|
| 81424 |
],
|
| 81425 |
"scale_bits": 4
|
| 81426 |
},
|
| 81427 |
"down_proj": {
|
| 81428 |
"group_size": {
|
| 81429 |
"8": 32,
|
| 81430 |
+
"6": 128,
|
| 81431 |
+
"5": 128
|
| 81432 |
},
|
| 81433 |
"bits": [
|
| 81434 |
8,
|
| 81435 |
+
6,
|
| 81436 |
+
5
|
| 81437 |
],
|
| 81438 |
"bits_prop": [
|
| 81439 |
0.05,
|
| 81440 |
+
0.1,
|
| 81441 |
+
0.85
|
| 81442 |
],
|
| 81443 |
"scale_bits": 4
|
| 81444 |
}
|
| 81445 |
},
|
| 81446 |
+
"model.layers.33.self_attn": {
|
| 81447 |
+
"accuracy": 0.999251619039569,
|
| 81448 |
"total_bits": 526999040,
|
| 81449 |
"q_proj": {
|
| 81450 |
"group_size": {
|
|
|
|
| 81495 |
"scale_bits": 4
|
| 81496 |
}
|
| 81497 |
},
|
| 81498 |
+
"model.layers.33.mlp": {
|
| 81499 |
+
"accuracy": 0.9945772625505924,
|
| 81500 |
+
"total_bits": 1725245760,
|
| 81501 |
"gate_proj": {
|
| 81502 |
"group_size": {
|
| 81503 |
+
"8": 128,
|
| 81504 |
"6": 128
|
| 81505 |
},
|
| 81506 |
"bits": [
|
| 81507 |
+
8,
|
| 81508 |
6
|
| 81509 |
],
|
| 81510 |
"bits_prop": [
|
| 81511 |
+
0.1,
|
| 81512 |
+
0.9
|
| 81513 |
],
|
| 81514 |
"scale_bits": 4
|
| 81515 |
},
|
| 81516 |
"up_proj": {
|
| 81517 |
"group_size": {
|
| 81518 |
+
"8": 128,
|
| 81519 |
"6": 128
|
| 81520 |
},
|
| 81521 |
"bits": [
|
| 81522 |
+
8,
|
| 81523 |
6
|
| 81524 |
],
|
| 81525 |
"bits_prop": [
|
| 81526 |
+
0.1,
|
| 81527 |
+
0.9
|
| 81528 |
],
|
| 81529 |
"scale_bits": 4
|
| 81530 |
},
|
| 81531 |
"down_proj": {
|
| 81532 |
"group_size": {
|
| 81533 |
+
"8": 128,
|
| 81534 |
"6": 128
|
| 81535 |
},
|
| 81536 |
"bits": [
|
|
|
|
| 81538 |
6
|
| 81539 |
],
|
| 81540 |
"bits_prop": [
|
| 81541 |
+
0.15,
|
| 81542 |
+
0.85
|
| 81543 |
],
|
| 81544 |
"scale_bits": 4
|
| 81545 |
}
|
| 81546 |
},
|
| 81547 |
+
"model.layers.34.self_attn": {
|
| 81548 |
+
"accuracy": 0.9960626754909754,
|
| 81549 |
+
"total_bits": 349009920,
|
| 81550 |
"q_proj": {
|
| 81551 |
"group_size": {
|
| 81552 |
+
"6": 32,
|
| 81553 |
+
"5": 32
|
| 81554 |
},
|
| 81555 |
"bits": [
|
| 81556 |
+
6,
|
| 81557 |
+
5
|
| 81558 |
],
|
| 81559 |
"bits_prop": [
|
| 81560 |
+
0.1,
|
| 81561 |
+
0.9
|
| 81562 |
],
|
| 81563 |
"scale_bits": 4
|
| 81564 |
},
|
| 81565 |
"k_proj": {
|
| 81566 |
"group_size": {
|
| 81567 |
+
"6": 32,
|
| 81568 |
+
"5": 32
|
| 81569 |
},
|
| 81570 |
"bits": [
|
| 81571 |
+
6,
|
| 81572 |
+
5
|
| 81573 |
],
|
| 81574 |
"bits_prop": [
|
| 81575 |
+
0.1,
|
| 81576 |
+
0.9
|
| 81577 |
],
|
| 81578 |
"scale_bits": 4
|
| 81579 |
},
|
| 81580 |
"v_proj": {
|
| 81581 |
"group_size": {
|
| 81582 |
+
"6": 32
|
| 81583 |
},
|
| 81584 |
"bits": [
|
| 81585 |
+
6
|
| 81586 |
],
|
| 81587 |
"bits_prop": [
|
| 81588 |
1
|
|
|
|
| 81591 |
},
|
| 81592 |
"o_proj": {
|
| 81593 |
"group_size": {
|
| 81594 |
+
"6": 32,
|
| 81595 |
+
"5": 32
|
| 81596 |
},
|
| 81597 |
"bits": [
|
| 81598 |
+
6,
|
| 81599 |
+
5
|
| 81600 |
],
|
| 81601 |
"bits_prop": [
|
| 81602 |
+
0.1,
|
| 81603 |
+
0.9
|
| 81604 |
],
|
| 81605 |
"scale_bits": 4
|
| 81606 |
}
|
| 81607 |
},
|
| 81608 |
+
"model.layers.34.mlp": {
|
| 81609 |
+
"accuracy": 0.9943122416734695,
|
| 81610 |
+
"total_bits": 1670626608,
|
| 81611 |
"gate_proj": {
|
| 81612 |
"group_size": {
|
| 81613 |
+
"6": 128
|
|
|
|
| 81614 |
},
|
| 81615 |
"bits": [
|
| 81616 |
+
6
|
|
|
|
| 81617 |
],
|
| 81618 |
"bits_prop": [
|
| 81619 |
+
1
|
|
|
|
| 81620 |
],
|
| 81621 |
"scale_bits": 4
|
| 81622 |
},
|
| 81623 |
"up_proj": {
|
| 81624 |
"group_size": {
|
| 81625 |
+
"6": 128
|
|
|
|
| 81626 |
},
|
| 81627 |
"bits": [
|
| 81628 |
+
6
|
|
|
|
| 81629 |
],
|
| 81630 |
"bits_prop": [
|
| 81631 |
+
1
|
|
|
|
| 81632 |
],
|
| 81633 |
"scale_bits": 4
|
| 81634 |
},
|
| 81635 |
"down_proj": {
|
| 81636 |
"group_size": {
|
| 81637 |
"8": 32,
|
| 81638 |
+
"6": 128
|
|
|
|
| 81639 |
},
|
| 81640 |
"bits": [
|
| 81641 |
8,
|
| 81642 |
+
6
|
|
|
|
| 81643 |
],
|
| 81644 |
"bits_prop": [
|
| 81645 |
0.05,
|
| 81646 |
+
0.95
|
|
|
|
| 81647 |
],
|
| 81648 |
"scale_bits": 4
|
| 81649 |
}
|
| 81650 |
},
|
| 81651 |
+
"model.layers.35.self_attn": {
|
| 81652 |
+
"accuracy": 0.9989852329017594,
|
| 81653 |
+
"total_bits": 526999040,
|
| 81654 |
"q_proj": {
|
| 81655 |
"group_size": {
|
| 81656 |
+
"8": 128
|
|
|
|
| 81657 |
},
|
| 81658 |
"bits": [
|
| 81659 |
+
8
|
|
|
|
| 81660 |
],
|
| 81661 |
"bits_prop": [
|
| 81662 |
+
1
|
|
|
|
| 81663 |
],
|
| 81664 |
"scale_bits": 4
|
| 81665 |
},
|
| 81666 |
"k_proj": {
|
| 81667 |
"group_size": {
|
| 81668 |
+
"8": 128
|
|
|
|
| 81669 |
},
|
| 81670 |
"bits": [
|
| 81671 |
+
8
|
|
|
|
| 81672 |
],
|
| 81673 |
"bits_prop": [
|
| 81674 |
+
1
|
|
|
|
| 81675 |
],
|
| 81676 |
"scale_bits": 4
|
| 81677 |
},
|
| 81678 |
"v_proj": {
|
| 81679 |
"group_size": {
|
| 81680 |
+
"8": 128
|
| 81681 |
},
|
| 81682 |
"bits": [
|
| 81683 |
+
8
|
| 81684 |
],
|
| 81685 |
"bits_prop": [
|
| 81686 |
1
|
|
|
|
| 81689 |
},
|
| 81690 |
"o_proj": {
|
| 81691 |
"group_size": {
|
| 81692 |
+
"8": 128
|
|
|
|
| 81693 |
},
|
| 81694 |
"bits": [
|
| 81695 |
+
8
|
|
|
|
| 81696 |
],
|
| 81697 |
"bits_prop": [
|
| 81698 |
+
1
|
|
|
|
| 81699 |
],
|
| 81700 |
"scale_bits": 4
|
| 81701 |
}
|
| 81702 |
},
|
| 81703 |
+
"model.layers.35.mlp": {
|
| 81704 |
+
"accuracy": 0.9912119247019291,
|
| 81705 |
+
"total_bits": 1466656000,
|
| 81706 |
"gate_proj": {
|
| 81707 |
"group_size": {
|
| 81708 |
+
"6": 32,
|
| 81709 |
+
"5": 32
|
| 81710 |
},
|
| 81711 |
"bits": [
|
| 81712 |
6,
|
|
|
|
| 81720 |
},
|
| 81721 |
"up_proj": {
|
| 81722 |
"group_size": {
|
| 81723 |
+
"6": 32,
|
| 81724 |
+
"5": 32
|
| 81725 |
},
|
| 81726 |
"bits": [
|
| 81727 |
6,
|
|
|
|
| 81736 |
"down_proj": {
|
| 81737 |
"group_size": {
|
| 81738 |
"8": 32,
|
| 81739 |
+
"6": 32,
|
| 81740 |
+
"5": 32
|
| 81741 |
},
|
| 81742 |
"bits": [
|
| 81743 |
8,
|
|
|
|
| 81752 |
"scale_bits": 4
|
| 81753 |
}
|
| 81754 |
},
|
| 81755 |
+
"model.layers.36.self_attn": {
|
| 81756 |
+
"accuracy": 0.9965386835392565,
|
| 81757 |
"total_bits": 349009920,
|
| 81758 |
"q_proj": {
|
| 81759 |
"group_size": {
|
|
|
|
| 81813 |
"scale_bits": 4
|
| 81814 |
}
|
| 81815 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81816 |
"model.layers.36.mlp": {
|
| 81817 |
+
"accuracy": 0.9901178376749158,
|
| 81818 |
+
"total_bits": 1441250608,
|
| 81819 |
"gate_proj": {
|
| 81820 |
"group_size": {
|
| 81821 |
+
"6": 128,
|
| 81822 |
+
"5": 128
|
| 81823 |
},
|
| 81824 |
"bits": [
|
| 81825 |
6,
|
|
|
|
| 81833 |
},
|
| 81834 |
"up_proj": {
|
| 81835 |
"group_size": {
|
| 81836 |
+
"6": 128,
|
| 81837 |
+
"5": 128
|
| 81838 |
},
|
| 81839 |
"bits": [
|
| 81840 |
6,
|
|
|
|
| 81849 |
"down_proj": {
|
| 81850 |
"group_size": {
|
| 81851 |
"8": 32,
|
| 81852 |
+
"6": 128,
|
| 81853 |
+
"5": 128
|
| 81854 |
},
|
| 81855 |
"bits": [
|
| 81856 |
8,
|
|
|
|
| 81927 |
}
|
| 81928 |
},
|
| 81929 |
"model.layers.37.mlp": {
|
| 81930 |
+
"accuracy": 0.9958977806381881,
|
| 81931 |
+
"total_bits": 1881221440,
|
| 81932 |
"gate_proj": {
|
| 81933 |
"group_size": {
|
| 81934 |
"8": 128,
|
|
|
|
| 81961 |
},
|
| 81962 |
"down_proj": {
|
| 81963 |
"group_size": {
|
| 81964 |
+
"8": 128
|
|
|
|
| 81965 |
},
|
| 81966 |
"bits": [
|
| 81967 |
+
8
|
|
|
|
| 81968 |
],
|
| 81969 |
"bits_prop": [
|
| 81970 |
+
1
|
|
|
|
| 81971 |
],
|
| 81972 |
"scale_bits": 4
|
| 81973 |
}
|
| 81974 |
},
|
| 81975 |
"model.layers.38.self_attn": {
|
| 81976 |
+
"accuracy": 0.9993991161463782,
|
| 81977 |
+
"total_bits": 526999040,
|
| 81978 |
"q_proj": {
|
| 81979 |
"group_size": {
|
| 81980 |
+
"8": 128
|
| 81981 |
},
|
| 81982 |
"bits": [
|
| 81983 |
+
8
|
| 81984 |
],
|
| 81985 |
"bits_prop": [
|
| 81986 |
1
|
|
|
|
| 81989 |
},
|
| 81990 |
"k_proj": {
|
| 81991 |
"group_size": {
|
| 81992 |
+
"8": 128
|
| 81993 |
},
|
| 81994 |
"bits": [
|
| 81995 |
+
8
|
| 81996 |
],
|
| 81997 |
"bits_prop": [
|
| 81998 |
1
|
|
|
|
| 82001 |
},
|
| 82002 |
"v_proj": {
|
| 82003 |
"group_size": {
|
| 82004 |
+
"8": 128
|
| 82005 |
},
|
| 82006 |
"bits": [
|
| 82007 |
8
|
|
|
|
| 82013 |
},
|
| 82014 |
"o_proj": {
|
| 82015 |
"group_size": {
|
| 82016 |
+
"8": 128
|
| 82017 |
},
|
| 82018 |
"bits": [
|
| 82019 |
+
8
|
| 82020 |
],
|
| 82021 |
"bits_prop": [
|
| 82022 |
1
|
|
|
|
| 82025 |
}
|
| 82026 |
},
|
| 82027 |
"model.layers.38.mlp": {
|
| 82028 |
+
"accuracy": 0.9934721081517637,
|
| 82029 |
+
"total_bits": 1441250608,
|
| 82030 |
"gate_proj": {
|
| 82031 |
"group_size": {
|
| 82032 |
+
"6": 128,
|
| 82033 |
+
"5": 128
|
| 82034 |
},
|
| 82035 |
"bits": [
|
| 82036 |
6,
|
|
|
|
| 82044 |
},
|
| 82045 |
"up_proj": {
|
| 82046 |
"group_size": {
|
| 82047 |
+
"6": 128,
|
| 82048 |
+
"5": 128
|
| 82049 |
},
|
| 82050 |
"bits": [
|
| 82051 |
6,
|
|
|
|
| 82060 |
"down_proj": {
|
| 82061 |
"group_size": {
|
| 82062 |
"8": 32,
|
| 82063 |
+
"6": 128,
|
| 82064 |
+
"5": 128
|
| 82065 |
},
|
| 82066 |
"bits": [
|
| 82067 |
8,
|
|
|
|
| 82077 |
}
|
| 82078 |
},
|
| 82079 |
"model.layers.39.self_attn": {
|
| 82080 |
+
"accuracy": 0.9963345830328763,
|
| 82081 |
+
"total_bits": 271615360,
|
| 82082 |
"q_proj": {
|
| 82083 |
"group_size": {
|
| 82084 |
+
"5": 128,
|
| 82085 |
"4": 128
|
| 82086 |
},
|
| 82087 |
"bits": [
|
| 82088 |
+
5,
|
| 82089 |
4
|
| 82090 |
],
|
| 82091 |
"bits_prop": [
|
| 82092 |
+
0.1,
|
| 82093 |
+
0.9
|
| 82094 |
],
|
| 82095 |
"scale_bits": 4
|
| 82096 |
},
|
| 82097 |
"k_proj": {
|
| 82098 |
"group_size": {
|
| 82099 |
+
"5": 128,
|
| 82100 |
"4": 128
|
| 82101 |
},
|
| 82102 |
"bits": [
|
| 82103 |
+
5,
|
| 82104 |
4
|
| 82105 |
],
|
| 82106 |
"bits_prop": [
|
| 82107 |
+
0.1,
|
| 82108 |
+
0.9
|
| 82109 |
],
|
| 82110 |
"scale_bits": 4
|
| 82111 |
},
|
| 82112 |
"v_proj": {
|
| 82113 |
"group_size": {
|
| 82114 |
+
"5": 64,
|
| 82115 |
+
"4": 64
|
| 82116 |
},
|
| 82117 |
"bits": [
|
| 82118 |
+
5,
|
| 82119 |
4
|
| 82120 |
],
|
| 82121 |
"bits_prop": [
|
| 82122 |
+
0.1,
|
| 82123 |
+
0.9
|
| 82124 |
],
|
| 82125 |
"scale_bits": 4
|
| 82126 |
},
|
| 82127 |
"o_proj": {
|
| 82128 |
"group_size": {
|
| 82129 |
+
"5": 128,
|
| 82130 |
"4": 128
|
| 82131 |
},
|
| 82132 |
"bits": [
|
| 82133 |
+
5,
|
| 82134 |
4
|
| 82135 |
],
|
| 82136 |
"bits_prop": [
|
| 82137 |
+
0.1,
|
| 82138 |
+
0.9
|
| 82139 |
],
|
| 82140 |
"scale_bits": 4
|
| 82141 |
}
|
measurement.json
CHANGED
|
@@ -16934,7 +16934,7 @@
|
|
| 16934 |
}
|
| 16935 |
},
|
| 16936 |
{
|
| 16937 |
-
"accuracy": 0.
|
| 16938 |
"total_bits": 899923248,
|
| 16939 |
"gate_proj": {
|
| 16940 |
"group_size": {
|
|
@@ -17038,7 +17038,7 @@
|
|
| 17038 |
}
|
| 17039 |
},
|
| 17040 |
{
|
| 17041 |
-
"accuracy": 0.
|
| 17042 |
"total_bits": 995125808,
|
| 17043 |
"gate_proj": {
|
| 17044 |
"group_size": {
|
|
@@ -17087,7 +17087,7 @@
|
|
| 17087 |
}
|
| 17088 |
},
|
| 17089 |
{
|
| 17090 |
-
"accuracy": 0.
|
| 17091 |
"total_bits": 1137906608,
|
| 17092 |
"gate_proj": {
|
| 17093 |
"group_size": {
|
|
@@ -17881,7 +17881,7 @@
|
|
| 17881 |
}
|
| 17882 |
},
|
| 17883 |
{
|
| 17884 |
-
"accuracy": 0.
|
| 17885 |
"total_bits": 206079360,
|
| 17886 |
"q_proj": {
|
| 17887 |
"group_size": {
|
|
@@ -21527,7 +21527,7 @@
|
|
| 21527 |
}
|
| 21528 |
},
|
| 21529 |
{
|
| 21530 |
-
"accuracy": 0.
|
| 21531 |
"total_bits": 143375360,
|
| 21532 |
"q_proj": {
|
| 21533 |
"group_size": {
|
|
@@ -24536,7 +24536,7 @@
|
|
| 24536 |
],
|
| 24537 |
"model.layers.12.mlp": [
|
| 24538 |
{
|
| 24539 |
-
"accuracy": 0.
|
| 24540 |
"total_bits": 614790432,
|
| 24541 |
"gate_proj": {
|
| 24542 |
"group_size": {
|
|
@@ -24588,7 +24588,7 @@
|
|
| 24588 |
}
|
| 24589 |
},
|
| 24590 |
{
|
| 24591 |
-
"accuracy": 0.
|
| 24592 |
"total_bits": 637728032,
|
| 24593 |
"gate_proj": {
|
| 24594 |
"group_size": {
|
|
@@ -46251,7 +46251,7 @@
|
|
| 46251 |
}
|
| 46252 |
},
|
| 46253 |
{
|
| 46254 |
-
"accuracy": 0.
|
| 46255 |
"total_bits": 925328640,
|
| 46256 |
"gate_proj": {
|
| 46257 |
"group_size": {
|
|
@@ -51850,7 +51850,7 @@
|
|
| 51850 |
],
|
| 51851 |
"model.layers.26.mlp": [
|
| 51852 |
{
|
| 51853 |
-
"accuracy": 0.
|
| 51854 |
"total_bits": 614790432,
|
| 51855 |
"gate_proj": {
|
| 51856 |
"group_size": {
|
|
@@ -55118,7 +55118,7 @@
|
|
| 55118 |
}
|
| 55119 |
},
|
| 55120 |
{
|
| 55121 |
-
"accuracy": 0.
|
| 55122 |
"total_bits": 267324160,
|
| 55123 |
"q_proj": {
|
| 55124 |
"group_size": {
|
|
@@ -55170,7 +55170,7 @@
|
|
| 55170 |
}
|
| 55171 |
},
|
| 55172 |
{
|
| 55173 |
-
"accuracy": 0.
|
| 55174 |
"total_bits": 271022080,
|
| 55175 |
"q_proj": {
|
| 55176 |
"group_size": {
|
|
@@ -62974,7 +62974,7 @@
|
|
| 62974 |
}
|
| 62975 |
},
|
| 62976 |
{
|
| 62977 |
-
"accuracy": 0.
|
| 62978 |
"total_bits": 271022080,
|
| 62979 |
"q_proj": {
|
| 62980 |
"group_size": {
|
|
@@ -64385,7 +64385,7 @@
|
|
| 64385 |
],
|
| 64386 |
"model.layers.33.self_attn": [
|
| 64387 |
{
|
| 64388 |
-
"accuracy": 0.
|
| 64389 |
"total_bits": 139115520,
|
| 64390 |
"q_proj": {
|
| 64391 |
"group_size": {
|
|
|
|
| 16934 |
}
|
| 16935 |
},
|
| 16936 |
{
|
| 16937 |
+
"accuracy": 0.9644066467881203,
|
| 16938 |
"total_bits": 899923248,
|
| 16939 |
"gate_proj": {
|
| 16940 |
"group_size": {
|
|
|
|
| 17038 |
}
|
| 17039 |
},
|
| 17040 |
{
|
| 17041 |
+
"accuracy": 0.9716205019503832,
|
| 17042 |
"total_bits": 995125808,
|
| 17043 |
"gate_proj": {
|
| 17044 |
"group_size": {
|
|
|
|
| 17087 |
}
|
| 17088 |
},
|
| 17089 |
{
|
| 17090 |
+
"accuracy": 0.9818738773465157,
|
| 17091 |
"total_bits": 1137906608,
|
| 17092 |
"gate_proj": {
|
| 17093 |
"group_size": {
|
|
|
|
| 17881 |
}
|
| 17882 |
},
|
| 17883 |
{
|
| 17884 |
+
"accuracy": 0.9735957626253366,
|
| 17885 |
"total_bits": 206079360,
|
| 17886 |
"q_proj": {
|
| 17887 |
"group_size": {
|
|
|
|
| 21527 |
}
|
| 21528 |
},
|
| 21529 |
{
|
| 21530 |
+
"accuracy": 0.9386431761085987,
|
| 21531 |
"total_bits": 143375360,
|
| 21532 |
"q_proj": {
|
| 21533 |
"group_size": {
|
|
|
|
| 24536 |
],
|
| 24537 |
"model.layers.12.mlp": [
|
| 24538 |
{
|
| 24539 |
+
"accuracy": 0.9032091200351715,
|
| 24540 |
"total_bits": 614790432,
|
| 24541 |
"gate_proj": {
|
| 24542 |
"group_size": {
|
|
|
|
| 24588 |
}
|
| 24589 |
},
|
| 24590 |
{
|
| 24591 |
+
"accuracy": 0.9056077301502228,
|
| 24592 |
"total_bits": 637728032,
|
| 24593 |
"gate_proj": {
|
| 24594 |
"group_size": {
|
|
|
|
| 46251 |
}
|
| 46252 |
},
|
| 46253 |
{
|
| 46254 |
+
"accuracy": 0.949563205242157,
|
| 46255 |
"total_bits": 925328640,
|
| 46256 |
"gate_proj": {
|
| 46257 |
"group_size": {
|
|
|
|
| 51850 |
],
|
| 51851 |
"model.layers.26.mlp": [
|
| 51852 |
{
|
| 51853 |
+
"accuracy": 0.8996819406747818,
|
| 51854 |
"total_bits": 614790432,
|
| 51855 |
"gate_proj": {
|
| 51856 |
"group_size": {
|
|
|
|
| 55118 |
}
|
| 55119 |
},
|
| 55120 |
{
|
| 55121 |
+
"accuracy": 0.9861949309706688,
|
| 55122 |
"total_bits": 267324160,
|
| 55123 |
"q_proj": {
|
| 55124 |
"group_size": {
|
|
|
|
| 55170 |
}
|
| 55171 |
},
|
| 55172 |
{
|
| 55173 |
+
"accuracy": 0.9868358941748738,
|
| 55174 |
"total_bits": 271022080,
|
| 55175 |
"q_proj": {
|
| 55176 |
"group_size": {
|
|
|
|
| 62974 |
}
|
| 62975 |
},
|
| 62976 |
{
|
| 62977 |
+
"accuracy": 0.9891389207914472,
|
| 62978 |
"total_bits": 271022080,
|
| 62979 |
"q_proj": {
|
| 62980 |
"group_size": {
|
|
|
|
| 64385 |
],
|
| 64386 |
"model.layers.33.self_attn": [
|
| 64387 |
{
|
| 64388 |
+
"accuracy": 0.9657053500413895,
|
| 64389 |
"total_bits": 139115520,
|
| 64390 |
"q_proj": {
|
| 64391 |
"group_size": {
|
output-00001-of-00002.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8e0c887792f86c4d84ed160ca44b61acc544f14f134bd426199bee481452e6b1
|
| 3 |
+
size 8576095656
|
output-00002-of-00002.safetensors
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c5cd11fa6af7d82bb788473a9206c3c9919c231d69a4830d815c60c1b38f84df
|
| 3 |
+
size 3079355660
|