Chua, Vui Seng
commited on
Commit
•
ac8897e
1
Parent(s):
ffc90e9
Update readme and model analysis
Browse files- .gitattributes +1 -0
- README.md +11 -5
- XP_layer_wise_sparsity_global_rate_15.41.csv +200 -0
- XP_layer_wise_sparsity_global_rate_15.41.md +201 -0
- XP_linear_layer_sparsity_20M_params_33.64_sparsity.csv +73 -0
- XP_linear_layer_sparsity_20M_params_33.64_sparsity.md +74 -0
- all_results.json +5 -0
- eval_XP_results.json +5 -0
- eval_nbest_predictions.json +3 -0
- eval_predictions.json +0 -0
.gitattributes
CHANGED
@@ -25,3 +25,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
25 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
26 |
*.zstandard filter=lfs diff=lfs merge=lfs -text
|
27 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
eval_nbest_predictions.json filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -8,13 +8,14 @@ This model is a replication of [block pruning paper](https://arxiv.org/abs/2109.
|
|
8 |
To reproduce this model, pls follow [documentation here](https://github.com/vuiseng9/nn_pruning/blob/reproduce-evaluation/reproduce-eval/readme.md) until step 2.
|
9 |
|
10 |
# Eval
|
11 |
-
The model can be evaluated out-of-the-box with HF QA example. Note that only pruned self-attention heads are discarded where pruned ffn dimension are sparsified instead of removal. Verified in v4.13
|
12 |
```bash
|
13 |
export CUDA_VISIBLE_DEVICES=0
|
14 |
|
15 |
OUTDIR=eval-bert-base-squadv1-block-pruning-hybrid
|
16 |
WORKDIR=transformers/examples/pytorch/question-answering
|
17 |
cd $WORKDIR
|
|
|
18 |
|
19 |
nohup python run_qa.py \
|
20 |
--model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid \
|
@@ -34,17 +35,21 @@ git clone https://github.com/vuiseng9/nncf && cd nncf
|
|
34 |
git checkout tld-poc
|
35 |
git reset --hard 1dec7afe7a4b567c059fcf287ea2c234980fded2
|
36 |
python setup.py develop
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
|
38 |
# Huggingface Transformers
|
39 |
git clone https://github.com/vuiseng9/transformers && cd transformers
|
40 |
git checkout tld-poc
|
41 |
git reset --hard 10a1e29d84484e48fd106f58957d9ffc89dc43c5
|
42 |
pip install -e .
|
|
|
43 |
|
44 |
-
# Huggingface nn_pruning
|
45 |
-
git clone https://github.com/vuiseng9/nn_pruning && cd nn_pruning
|
46 |
-
git checkout reproduce-evaluation
|
47 |
-
git reset --hard 2d4e196d694c465e43e5fbce6c3836d0a60e1446
|
48 |
```
|
49 |
Add ```--optimize_model_before_eval``` during evaluation.
|
50 |
```bash
|
@@ -53,6 +58,7 @@ export CUDA_VISIBLE_DEVICES=0
|
|
53 |
OUTDIR=eval-bert-base-squadv1-block-pruning-hybrid-cropped
|
54 |
WORKDIR=transformers/examples/pytorch/question-answering
|
55 |
cd $WORKDIR
|
|
|
56 |
|
57 |
nohup python run_qa.py \
|
58 |
--model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid \
|
|
|
8 |
To reproduce this model, pls follow [documentation here](https://github.com/vuiseng9/nn_pruning/blob/reproduce-evaluation/reproduce-eval/readme.md) until step 2.
|
9 |
|
10 |
# Eval
|
11 |
+
The model can be evaluated out-of-the-box with HF QA example. Note that only pruned self-attention heads are discarded where pruned ffn dimension are sparsified instead of removal. Verified in v4.13.0, v4.9.1.
|
12 |
```bash
|
13 |
export CUDA_VISIBLE_DEVICES=0
|
14 |
|
15 |
OUTDIR=eval-bert-base-squadv1-block-pruning-hybrid
|
16 |
WORKDIR=transformers/examples/pytorch/question-answering
|
17 |
cd $WORKDIR
|
18 |
+
mkdir $OUTDIR
|
19 |
|
20 |
nohup python run_qa.py \
|
21 |
--model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid \
|
|
|
35 |
git checkout tld-poc
|
36 |
git reset --hard 1dec7afe7a4b567c059fcf287ea2c234980fded2
|
37 |
python setup.py develop
|
38 |
+
pip install -r examples/torch/requirements.txt
|
39 |
+
|
40 |
+
# Huggingface nn_pruning
|
41 |
+
git clone https://github.com/vuiseng9/nn_pruning && cd nn_pruning
|
42 |
+
git checkout reproduce-evaluation
|
43 |
+
git reset --hard 2d4e196d694c465e43e5fbce6c3836d0a60e1446
|
44 |
+
pip install -e ".[dev]"
|
45 |
|
46 |
# Huggingface Transformers
|
47 |
git clone https://github.com/vuiseng9/transformers && cd transformers
|
48 |
git checkout tld-poc
|
49 |
git reset --hard 10a1e29d84484e48fd106f58957d9ffc89dc43c5
|
50 |
pip install -e .
|
51 |
+
head -n 1 examples/pytorch/question-answering/requirements.txt | xargs -i pip install {}
|
52 |
|
|
|
|
|
|
|
|
|
53 |
```
|
54 |
Add ```--optimize_model_before_eval``` during evaluation.
|
55 |
```bash
|
|
|
58 |
OUTDIR=eval-bert-base-squadv1-block-pruning-hybrid-cropped
|
59 |
WORKDIR=transformers/examples/pytorch/question-answering
|
60 |
cd $WORKDIR
|
61 |
+
mkdir $OUTDIR
|
62 |
|
63 |
nohup python run_qa.py \
|
64 |
--model_name_or_path vuiseng9/bert-base-squadv1-block-pruning-hybrid \
|
XP_layer_wise_sparsity_global_rate_15.41.csv
ADDED
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
,layer_id,layer_type,param_type,shape,nparam,nnz,sparsity
|
2 |
+
0,bert.embeddings.word_embeddings,Embedding,weight,"[30522, 768]",23440896,23440896,0.0
|
3 |
+
1,bert.embeddings.position_embeddings,Embedding,weight,"[512, 768]",393216,393216,0.0
|
4 |
+
2,bert.embeddings.token_type_embeddings,Embedding,weight,"[2, 768]",1536,1536,0.0
|
5 |
+
3,bert.embeddings.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
6 |
+
4,bert.embeddings.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
7 |
+
5,bert.encoder.layer.0.attention.self.query,Linear,weight,"[320, 768]",245760,135168,0.44999998807907104
|
8 |
+
6,bert.encoder.layer.0.attention.self.query,Linear,bias,[320],320,256,0.19999998807907104
|
9 |
+
7,bert.encoder.layer.0.attention.self.key,Linear,weight,"[320, 768]",245760,149504,0.3916666507720947
|
10 |
+
8,bert.encoder.layer.0.attention.self.key,Linear,bias,[320],320,256,0.19999998807907104
|
11 |
+
9,bert.encoder.layer.0.attention.self.value,Linear,weight,"[320, 768]",245760,173056,0.2958332896232605
|
12 |
+
10,bert.encoder.layer.0.attention.self.value,Linear,bias,[320],320,256,0.19999998807907104
|
13 |
+
11,bert.encoder.layer.0.attention.output.dense,Linear,weight,"[768, 320]",245760,181248,0.26249998807907104
|
14 |
+
12,bert.encoder.layer.0.attention.output.dense,Linear,bias,[768],768,768,0.0
|
15 |
+
13,bert.encoder.layer.0.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
16 |
+
14,bert.encoder.layer.0.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
17 |
+
15,bert.encoder.layer.0.intermediate.dense,Linear,weight,"[185, 768]",142080,142080,0.0
|
18 |
+
16,bert.encoder.layer.0.intermediate.dense,Linear,bias,[185],185,185,0.0
|
19 |
+
17,bert.encoder.layer.0.output.dense,Linear,weight,"[768, 185]",142080,142080,0.0
|
20 |
+
18,bert.encoder.layer.0.output.dense,Linear,bias,[768],768,768,0.0
|
21 |
+
19,bert.encoder.layer.0.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
22 |
+
20,bert.encoder.layer.0.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
23 |
+
21,bert.encoder.layer.1.attention.self.query,Linear,weight,"[320, 768]",245760,175104,0.28749996423721313
|
24 |
+
22,bert.encoder.layer.1.attention.self.query,Linear,bias,[320],320,288,0.09999996423721313
|
25 |
+
23,bert.encoder.layer.1.attention.self.key,Linear,weight,"[320, 768]",245760,177152,0.27916663885116577
|
26 |
+
24,bert.encoder.layer.1.attention.self.key,Linear,bias,[320],320,288,0.09999996423721313
|
27 |
+
25,bert.encoder.layer.1.attention.self.value,Linear,weight,"[320, 768]",245760,166912,0.32083332538604736
|
28 |
+
26,bert.encoder.layer.1.attention.self.value,Linear,bias,[320],320,288,0.09999996423721313
|
29 |
+
27,bert.encoder.layer.1.attention.output.dense,Linear,weight,"[768, 320]",245760,167936,0.3166666030883789
|
30 |
+
28,bert.encoder.layer.1.attention.output.dense,Linear,bias,[768],768,768,0.0
|
31 |
+
29,bert.encoder.layer.1.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
32 |
+
30,bert.encoder.layer.1.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
33 |
+
31,bert.encoder.layer.1.intermediate.dense,Linear,weight,"[315, 768]",241920,241920,0.0
|
34 |
+
32,bert.encoder.layer.1.intermediate.dense,Linear,bias,[315],315,315,0.0
|
35 |
+
33,bert.encoder.layer.1.output.dense,Linear,weight,"[768, 315]",241920,241920,0.0
|
36 |
+
34,bert.encoder.layer.1.output.dense,Linear,bias,[768],768,768,0.0
|
37 |
+
35,bert.encoder.layer.1.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
38 |
+
36,bert.encoder.layer.1.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
39 |
+
37,bert.encoder.layer.2.attention.self.query,Linear,weight,"[576, 768]",442368,285696,0.3541666865348816
|
40 |
+
38,bert.encoder.layer.2.attention.self.query,Linear,bias,[576],576,480,0.1666666865348816
|
41 |
+
39,bert.encoder.layer.2.attention.self.key,Linear,weight,"[576, 768]",442368,297984,0.3263888955116272
|
42 |
+
40,bert.encoder.layer.2.attention.self.key,Linear,bias,[576],576,480,0.1666666865348816
|
43 |
+
41,bert.encoder.layer.2.attention.self.value,Linear,weight,"[576, 768]",442368,226304,0.4884259104728699
|
44 |
+
42,bert.encoder.layer.2.attention.self.value,Linear,bias,[576],576,384,0.3333333134651184
|
45 |
+
43,bert.encoder.layer.2.attention.output.dense,Linear,weight,"[768, 576]",442368,237568,0.4629629850387573
|
46 |
+
44,bert.encoder.layer.2.attention.output.dense,Linear,bias,[768],768,768,0.0
|
47 |
+
45,bert.encoder.layer.2.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
48 |
+
46,bert.encoder.layer.2.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
49 |
+
47,bert.encoder.layer.2.intermediate.dense,Linear,weight,"[339, 768]",260352,260352,0.0
|
50 |
+
48,bert.encoder.layer.2.intermediate.dense,Linear,bias,[339],339,339,0.0
|
51 |
+
49,bert.encoder.layer.2.output.dense,Linear,weight,"[768, 339]",260352,260352,0.0
|
52 |
+
50,bert.encoder.layer.2.output.dense,Linear,bias,[768],768,768,0.0
|
53 |
+
51,bert.encoder.layer.2.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
54 |
+
52,bert.encoder.layer.2.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
55 |
+
53,bert.encoder.layer.3.attention.self.query,Linear,weight,"[576, 768]",442368,277504,0.37268519401550293
|
56 |
+
54,bert.encoder.layer.3.attention.self.query,Linear,bias,[576],576,512,0.1111111044883728
|
57 |
+
55,bert.encoder.layer.3.attention.self.key,Linear,weight,"[576, 768]",442368,303104,0.31481480598449707
|
58 |
+
56,bert.encoder.layer.3.attention.self.key,Linear,bias,[576],576,512,0.1111111044883728
|
59 |
+
57,bert.encoder.layer.3.attention.self.value,Linear,weight,"[576, 768]",442368,297984,0.3263888955116272
|
60 |
+
58,bert.encoder.layer.3.attention.self.value,Linear,bias,[576],576,512,0.1111111044883728
|
61 |
+
59,bert.encoder.layer.3.attention.output.dense,Linear,weight,"[768, 576]",442368,308224,0.30324071645736694
|
62 |
+
60,bert.encoder.layer.3.attention.output.dense,Linear,bias,[768],768,768,0.0
|
63 |
+
61,bert.encoder.layer.3.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
64 |
+
62,bert.encoder.layer.3.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
65 |
+
63,bert.encoder.layer.3.intermediate.dense,Linear,weight,"[368, 768]",282624,282624,0.0
|
66 |
+
64,bert.encoder.layer.3.intermediate.dense,Linear,bias,[368],368,368,0.0
|
67 |
+
65,bert.encoder.layer.3.output.dense,Linear,weight,"[768, 368]",282624,282624,0.0
|
68 |
+
66,bert.encoder.layer.3.output.dense,Linear,bias,[768],768,768,0.0
|
69 |
+
67,bert.encoder.layer.3.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
70 |
+
68,bert.encoder.layer.3.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
71 |
+
69,bert.encoder.layer.4.attention.self.query,Linear,weight,"[576, 768]",442368,291840,0.3402777910232544
|
72 |
+
70,bert.encoder.layer.4.attention.self.query,Linear,bias,[576],576,544,0.055555522441864014
|
73 |
+
71,bert.encoder.layer.4.attention.self.key,Linear,weight,"[576, 768]",442368,310272,0.2986111044883728
|
74 |
+
72,bert.encoder.layer.4.attention.self.key,Linear,bias,[576],576,544,0.055555522441864014
|
75 |
+
73,bert.encoder.layer.4.attention.self.value,Linear,weight,"[576, 768]",442368,272384,0.38425928354263306
|
76 |
+
74,bert.encoder.layer.4.attention.self.value,Linear,bias,[576],576,480,0.1666666865348816
|
77 |
+
75,bert.encoder.layer.4.attention.output.dense,Linear,weight,"[768, 576]",442368,263168,0.40509259700775146
|
78 |
+
76,bert.encoder.layer.4.attention.output.dense,Linear,bias,[768],768,768,0.0
|
79 |
+
77,bert.encoder.layer.4.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
80 |
+
78,bert.encoder.layer.4.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
81 |
+
79,bert.encoder.layer.4.intermediate.dense,Linear,weight,"[386, 768]",296448,296448,0.0
|
82 |
+
80,bert.encoder.layer.4.intermediate.dense,Linear,bias,[386],386,386,0.0
|
83 |
+
81,bert.encoder.layer.4.output.dense,Linear,weight,"[768, 386]",296448,296448,0.0
|
84 |
+
82,bert.encoder.layer.4.output.dense,Linear,bias,[768],768,768,0.0
|
85 |
+
83,bert.encoder.layer.4.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
86 |
+
84,bert.encoder.layer.4.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
87 |
+
85,bert.encoder.layer.5.attention.self.query,Linear,weight,"[384, 768]",294912,171008,0.4201388955116272
|
88 |
+
86,bert.encoder.layer.5.attention.self.query,Linear,bias,[384],384,352,0.08333331346511841
|
89 |
+
87,bert.encoder.layer.5.attention.self.key,Linear,weight,"[384, 768]",294912,205824,0.3020833134651184
|
90 |
+
88,bert.encoder.layer.5.attention.self.key,Linear,bias,[384],384,352,0.08333331346511841
|
91 |
+
89,bert.encoder.layer.5.attention.self.value,Linear,weight,"[384, 768]",294912,217088,0.2638888955116272
|
92 |
+
90,bert.encoder.layer.5.attention.self.value,Linear,bias,[384],384,384,0.0
|
93 |
+
91,bert.encoder.layer.5.attention.output.dense,Linear,weight,"[768, 384]",294912,223232,0.243055522441864
|
94 |
+
92,bert.encoder.layer.5.attention.output.dense,Linear,bias,[768],768,768,0.0
|
95 |
+
93,bert.encoder.layer.5.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
96 |
+
94,bert.encoder.layer.5.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
97 |
+
95,bert.encoder.layer.5.intermediate.dense,Linear,weight,"[336, 768]",258048,258048,0.0
|
98 |
+
96,bert.encoder.layer.5.intermediate.dense,Linear,bias,[336],336,336,0.0
|
99 |
+
97,bert.encoder.layer.5.output.dense,Linear,weight,"[768, 336]",258048,258048,0.0
|
100 |
+
98,bert.encoder.layer.5.output.dense,Linear,bias,[768],768,768,0.0
|
101 |
+
99,bert.encoder.layer.5.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
102 |
+
100,bert.encoder.layer.5.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
103 |
+
101,bert.encoder.layer.6.attention.self.query,Linear,weight,"[448, 768]",344064,192512,0.4404761791229248
|
104 |
+
102,bert.encoder.layer.6.attention.self.query,Linear,bias,[448],448,416,0.07142853736877441
|
105 |
+
103,bert.encoder.layer.6.attention.self.key,Linear,weight,"[448, 768]",344064,224256,0.3482142686843872
|
106 |
+
104,bert.encoder.layer.6.attention.self.key,Linear,bias,[448],448,416,0.07142853736877441
|
107 |
+
105,bert.encoder.layer.6.attention.self.value,Linear,weight,"[448, 768]",344064,209920,0.3898809552192688
|
108 |
+
106,bert.encoder.layer.6.attention.self.value,Linear,bias,[448],448,352,0.21428567171096802
|
109 |
+
107,bert.encoder.layer.6.attention.output.dense,Linear,weight,"[768, 448]",344064,199680,0.4196428656578064
|
110 |
+
108,bert.encoder.layer.6.attention.output.dense,Linear,bias,[768],768,768,0.0
|
111 |
+
109,bert.encoder.layer.6.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
112 |
+
110,bert.encoder.layer.6.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
113 |
+
111,bert.encoder.layer.6.intermediate.dense,Linear,weight,"[280, 768]",215040,215040,0.0
|
114 |
+
112,bert.encoder.layer.6.intermediate.dense,Linear,bias,[280],280,280,0.0
|
115 |
+
113,bert.encoder.layer.6.output.dense,Linear,weight,"[768, 280]",215040,215040,0.0
|
116 |
+
114,bert.encoder.layer.6.output.dense,Linear,bias,[768],768,768,0.0
|
117 |
+
115,bert.encoder.layer.6.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
118 |
+
116,bert.encoder.layer.6.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
119 |
+
117,bert.encoder.layer.7.attention.self.query,Linear,weight,"[448, 768]",344064,201728,0.413690447807312
|
120 |
+
118,bert.encoder.layer.7.attention.self.query,Linear,bias,[448],448,416,0.07142853736877441
|
121 |
+
119,bert.encoder.layer.7.attention.self.key,Linear,weight,"[448, 768]",344064,237568,0.3095238208770752
|
122 |
+
120,bert.encoder.layer.7.attention.self.key,Linear,bias,[448],448,416,0.07142853736877441
|
123 |
+
121,bert.encoder.layer.7.attention.self.value,Linear,weight,"[448, 768]",344064,218112,0.3660714030265808
|
124 |
+
122,bert.encoder.layer.7.attention.self.value,Linear,bias,[448],448,352,0.21428567171096802
|
125 |
+
123,bert.encoder.layer.7.attention.output.dense,Linear,weight,"[768, 448]",344064,202752,0.4107142686843872
|
126 |
+
124,bert.encoder.layer.7.attention.output.dense,Linear,bias,[768],768,768,0.0
|
127 |
+
125,bert.encoder.layer.7.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
128 |
+
126,bert.encoder.layer.7.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
129 |
+
127,bert.encoder.layer.7.intermediate.dense,Linear,weight,"[211, 768]",162048,162048,0.0
|
130 |
+
128,bert.encoder.layer.7.intermediate.dense,Linear,bias,[211],211,211,0.0
|
131 |
+
129,bert.encoder.layer.7.output.dense,Linear,weight,"[768, 211]",162048,162048,0.0
|
132 |
+
130,bert.encoder.layer.7.output.dense,Linear,bias,[768],768,768,0.0
|
133 |
+
131,bert.encoder.layer.7.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
134 |
+
132,bert.encoder.layer.7.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
135 |
+
133,bert.encoder.layer.8.attention.self.query,Linear,weight,"[448, 768]",344064,186368,0.4583333134651184
|
136 |
+
134,bert.encoder.layer.8.attention.self.query,Linear,bias,[448],448,416,0.07142853736877441
|
137 |
+
135,bert.encoder.layer.8.attention.self.key,Linear,weight,"[448, 768]",344064,197632,0.425595223903656
|
138 |
+
136,bert.encoder.layer.8.attention.self.key,Linear,bias,[448],448,416,0.07142853736877441
|
139 |
+
137,bert.encoder.layer.8.attention.self.value,Linear,weight,"[448, 768]",344064,154624,0.550595223903656
|
140 |
+
138,bert.encoder.layer.8.attention.self.value,Linear,bias,[448],448,288,0.3571428060531616
|
141 |
+
139,bert.encoder.layer.8.attention.output.dense,Linear,weight,"[768, 448]",344064,148480,0.5684523582458496
|
142 |
+
140,bert.encoder.layer.8.attention.output.dense,Linear,bias,[768],768,768,0.0
|
143 |
+
141,bert.encoder.layer.8.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
144 |
+
142,bert.encoder.layer.8.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
145 |
+
143,bert.encoder.layer.8.intermediate.dense,Linear,weight,"[108, 768]",82944,82944,0.0
|
146 |
+
144,bert.encoder.layer.8.intermediate.dense,Linear,bias,[108],108,108,0.0
|
147 |
+
145,bert.encoder.layer.8.output.dense,Linear,weight,"[768, 108]",82944,82944,0.0
|
148 |
+
146,bert.encoder.layer.8.output.dense,Linear,bias,[768],768,768,0.0
|
149 |
+
147,bert.encoder.layer.8.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
150 |
+
148,bert.encoder.layer.8.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
151 |
+
149,bert.encoder.layer.9.attention.self.query,Linear,weight,"[320, 768]",245760,144384,0.41249996423721313
|
152 |
+
150,bert.encoder.layer.9.attention.self.query,Linear,bias,[320],320,288,0.09999996423721313
|
153 |
+
151,bert.encoder.layer.9.attention.self.key,Linear,weight,"[320, 768]",245760,155648,0.36666661500930786
|
154 |
+
152,bert.encoder.layer.9.attention.self.key,Linear,bias,[320],320,288,0.09999996423721313
|
155 |
+
153,bert.encoder.layer.9.attention.self.value,Linear,weight,"[320, 768]",245760,63488,0.7416666746139526
|
156 |
+
154,bert.encoder.layer.9.attention.self.value,Linear,bias,[320],320,160,0.5
|
157 |
+
155,bert.encoder.layer.9.attention.output.dense,Linear,weight,"[768, 320]",245760,65536,0.7333333492279053
|
158 |
+
156,bert.encoder.layer.9.attention.output.dense,Linear,bias,[768],768,704,0.08333331346511841
|
159 |
+
157,bert.encoder.layer.9.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
160 |
+
158,bert.encoder.layer.9.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
161 |
+
159,bert.encoder.layer.9.intermediate.dense,Linear,weight,"[53, 768]",40704,40704,5.960464477539063e-08
|
162 |
+
160,bert.encoder.layer.9.intermediate.dense,Linear,bias,[53],53,53,0.0
|
163 |
+
161,bert.encoder.layer.9.output.dense,Linear,weight,"[768, 53]",40704,40704,5.960464477539063e-08
|
164 |
+
162,bert.encoder.layer.9.output.dense,Linear,bias,[768],768,768,0.0
|
165 |
+
163,bert.encoder.layer.9.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
166 |
+
164,bert.encoder.layer.9.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
167 |
+
165,bert.encoder.layer.10.attention.self.query,Linear,weight,"[384, 768]",294912,158720,0.461805522441864
|
168 |
+
166,bert.encoder.layer.10.attention.self.query,Linear,bias,[384],384,320,0.16666662693023682
|
169 |
+
167,bert.encoder.layer.10.attention.self.key,Linear,weight,"[384, 768]",294912,158720,0.461805522441864
|
170 |
+
168,bert.encoder.layer.10.attention.self.key,Linear,bias,[384],384,320,0.16666662693023682
|
171 |
+
169,bert.encoder.layer.10.attention.self.value,Linear,weight,"[384, 768]",294912,77824,0.7361111044883728
|
172 |
+
170,bert.encoder.layer.10.attention.self.value,Linear,bias,[384],384,192,0.5
|
173 |
+
171,bert.encoder.layer.10.attention.output.dense,Linear,weight,"[768, 384]",294912,78848,0.7326388955116272
|
174 |
+
172,bert.encoder.layer.10.attention.output.dense,Linear,bias,[768],768,736,0.041666626930236816
|
175 |
+
173,bert.encoder.layer.10.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
176 |
+
174,bert.encoder.layer.10.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
177 |
+
175,bert.encoder.layer.10.intermediate.dense,Linear,weight,"[86, 768]",66048,66048,0.0
|
178 |
+
176,bert.encoder.layer.10.intermediate.dense,Linear,bias,[86],86,86,0.0
|
179 |
+
177,bert.encoder.layer.10.output.dense,Linear,weight,"[768, 86]",66048,66048,0.0
|
180 |
+
178,bert.encoder.layer.10.output.dense,Linear,bias,[768],768,768,0.0
|
181 |
+
179,bert.encoder.layer.10.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
182 |
+
180,bert.encoder.layer.10.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
183 |
+
181,bert.encoder.layer.11.attention.self.query,Linear,weight,"[384, 768]",294912,107520,0.6354166269302368
|
184 |
+
182,bert.encoder.layer.11.attention.self.query,Linear,bias,[384],384,256,0.3333333134651184
|
185 |
+
183,bert.encoder.layer.11.attention.self.key,Linear,weight,"[384, 768]",294912,118784,0.5972222089767456
|
186 |
+
184,bert.encoder.layer.11.attention.self.key,Linear,bias,[384],384,256,0.3333333134651184
|
187 |
+
185,bert.encoder.layer.11.attention.self.value,Linear,weight,"[384, 768]",294912,62464,0.7881944179534912
|
188 |
+
186,bert.encoder.layer.11.attention.self.value,Linear,bias,[384],384,192,0.5
|
189 |
+
187,bert.encoder.layer.11.attention.output.dense,Linear,weight,"[768, 384]",294912,54272,0.8159722089767456
|
190 |
+
188,bert.encoder.layer.11.attention.output.dense,Linear,bias,[768],768,672,0.125
|
191 |
+
189,bert.encoder.layer.11.attention.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
192 |
+
190,bert.encoder.layer.11.attention.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
193 |
+
191,bert.encoder.layer.11.intermediate.dense,Linear,weight,"[105, 768]",80640,80640,0.0
|
194 |
+
192,bert.encoder.layer.11.intermediate.dense,Linear,bias,[105],105,105,0.0
|
195 |
+
193,bert.encoder.layer.11.output.dense,Linear,weight,"[768, 105]",80640,80640,0.0
|
196 |
+
194,bert.encoder.layer.11.output.dense,Linear,bias,[768],768,768,0.0
|
197 |
+
195,bert.encoder.layer.11.output.LayerNorm,LayerNorm,weight,[768],768,768,0.0
|
198 |
+
196,bert.encoder.layer.11.output.LayerNorm,LayerNorm,bias,[768],768,768,0.0
|
199 |
+
197,qa_outputs,Linear,weight,"[2, 768]",1536,1536,0.0
|
200 |
+
198,qa_outputs,Linear,bias,[2],2,2,0.0
|
XP_layer_wise_sparsity_global_rate_15.41.md
ADDED
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
| | layer_id | layer_type | param_type | shape | nparam | nnz | sparsity |
|
2 |
+
|----:|:-------------------------------------------------|:-------------|:-------------|:-------------|---------:|---------:|------------:|
|
3 |
+
| 0 | bert.embeddings.word_embeddings | Embedding | weight | [30522, 768] | 23440896 | 23440896 | 0 |
|
4 |
+
| 1 | bert.embeddings.position_embeddings | Embedding | weight | [512, 768] | 393216 | 393216 | 0 |
|
5 |
+
| 2 | bert.embeddings.token_type_embeddings | Embedding | weight | [2, 768] | 1536 | 1536 | 0 |
|
6 |
+
| 3 | bert.embeddings.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
7 |
+
| 4 | bert.embeddings.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
8 |
+
| 5 | bert.encoder.layer.0.attention.self.query | Linear | weight | [320, 768] | 245760 | 135168 | 0.45 |
|
9 |
+
| 6 | bert.encoder.layer.0.attention.self.query | Linear | bias | [320] | 320 | 256 | 0.2 |
|
10 |
+
| 7 | bert.encoder.layer.0.attention.self.key | Linear | weight | [320, 768] | 245760 | 149504 | 0.391667 |
|
11 |
+
| 8 | bert.encoder.layer.0.attention.self.key | Linear | bias | [320] | 320 | 256 | 0.2 |
|
12 |
+
| 9 | bert.encoder.layer.0.attention.self.value | Linear | weight | [320, 768] | 245760 | 173056 | 0.295833 |
|
13 |
+
| 10 | bert.encoder.layer.0.attention.self.value | Linear | bias | [320] | 320 | 256 | 0.2 |
|
14 |
+
| 11 | bert.encoder.layer.0.attention.output.dense | Linear | weight | [768, 320] | 245760 | 181248 | 0.2625 |
|
15 |
+
| 12 | bert.encoder.layer.0.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
16 |
+
| 13 | bert.encoder.layer.0.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
17 |
+
| 14 | bert.encoder.layer.0.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
18 |
+
| 15 | bert.encoder.layer.0.intermediate.dense | Linear | weight | [185, 768] | 142080 | 142080 | 0 |
|
19 |
+
| 16 | bert.encoder.layer.0.intermediate.dense | Linear | bias | [185] | 185 | 185 | 0 |
|
20 |
+
| 17 | bert.encoder.layer.0.output.dense | Linear | weight | [768, 185] | 142080 | 142080 | 0 |
|
21 |
+
| 18 | bert.encoder.layer.0.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
22 |
+
| 19 | bert.encoder.layer.0.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
23 |
+
| 20 | bert.encoder.layer.0.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
24 |
+
| 21 | bert.encoder.layer.1.attention.self.query | Linear | weight | [320, 768] | 245760 | 175104 | 0.2875 |
|
25 |
+
| 22 | bert.encoder.layer.1.attention.self.query | Linear | bias | [320] | 320 | 288 | 0.1 |
|
26 |
+
| 23 | bert.encoder.layer.1.attention.self.key | Linear | weight | [320, 768] | 245760 | 177152 | 0.279167 |
|
27 |
+
| 24 | bert.encoder.layer.1.attention.self.key | Linear | bias | [320] | 320 | 288 | 0.1 |
|
28 |
+
| 25 | bert.encoder.layer.1.attention.self.value | Linear | weight | [320, 768] | 245760 | 166912 | 0.320833 |
|
29 |
+
| 26 | bert.encoder.layer.1.attention.self.value | Linear | bias | [320] | 320 | 288 | 0.1 |
|
30 |
+
| 27 | bert.encoder.layer.1.attention.output.dense | Linear | weight | [768, 320] | 245760 | 167936 | 0.316667 |
|
31 |
+
| 28 | bert.encoder.layer.1.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
32 |
+
| 29 | bert.encoder.layer.1.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
33 |
+
| 30 | bert.encoder.layer.1.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
34 |
+
| 31 | bert.encoder.layer.1.intermediate.dense | Linear | weight | [315, 768] | 241920 | 241920 | 0 |
|
35 |
+
| 32 | bert.encoder.layer.1.intermediate.dense | Linear | bias | [315] | 315 | 315 | 0 |
|
36 |
+
| 33 | bert.encoder.layer.1.output.dense | Linear | weight | [768, 315] | 241920 | 241920 | 0 |
|
37 |
+
| 34 | bert.encoder.layer.1.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
38 |
+
| 35 | bert.encoder.layer.1.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
39 |
+
| 36 | bert.encoder.layer.1.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
40 |
+
| 37 | bert.encoder.layer.2.attention.self.query | Linear | weight | [576, 768] | 442368 | 285696 | 0.354167 |
|
41 |
+
| 38 | bert.encoder.layer.2.attention.self.query | Linear | bias | [576] | 576 | 480 | 0.166667 |
|
42 |
+
| 39 | bert.encoder.layer.2.attention.self.key | Linear | weight | [576, 768] | 442368 | 297984 | 0.326389 |
|
43 |
+
| 40 | bert.encoder.layer.2.attention.self.key | Linear | bias | [576] | 576 | 480 | 0.166667 |
|
44 |
+
| 41 | bert.encoder.layer.2.attention.self.value | Linear | weight | [576, 768] | 442368 | 226304 | 0.488426 |
|
45 |
+
| 42 | bert.encoder.layer.2.attention.self.value | Linear | bias | [576] | 576 | 384 | 0.333333 |
|
46 |
+
| 43 | bert.encoder.layer.2.attention.output.dense | Linear | weight | [768, 576] | 442368 | 237568 | 0.462963 |
|
47 |
+
| 44 | bert.encoder.layer.2.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
48 |
+
| 45 | bert.encoder.layer.2.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
49 |
+
| 46 | bert.encoder.layer.2.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
50 |
+
| 47 | bert.encoder.layer.2.intermediate.dense | Linear | weight | [339, 768] | 260352 | 260352 | 0 |
|
51 |
+
| 48 | bert.encoder.layer.2.intermediate.dense | Linear | bias | [339] | 339 | 339 | 0 |
|
52 |
+
| 49 | bert.encoder.layer.2.output.dense | Linear | weight | [768, 339] | 260352 | 260352 | 0 |
|
53 |
+
| 50 | bert.encoder.layer.2.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
54 |
+
| 51 | bert.encoder.layer.2.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
55 |
+
| 52 | bert.encoder.layer.2.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
56 |
+
| 53 | bert.encoder.layer.3.attention.self.query | Linear | weight | [576, 768] | 442368 | 277504 | 0.372685 |
|
57 |
+
| 54 | bert.encoder.layer.3.attention.self.query | Linear | bias | [576] | 576 | 512 | 0.111111 |
|
58 |
+
| 55 | bert.encoder.layer.3.attention.self.key | Linear | weight | [576, 768] | 442368 | 303104 | 0.314815 |
|
59 |
+
| 56 | bert.encoder.layer.3.attention.self.key | Linear | bias | [576] | 576 | 512 | 0.111111 |
|
60 |
+
| 57 | bert.encoder.layer.3.attention.self.value | Linear | weight | [576, 768] | 442368 | 297984 | 0.326389 |
|
61 |
+
| 58 | bert.encoder.layer.3.attention.self.value | Linear | bias | [576] | 576 | 512 | 0.111111 |
|
62 |
+
| 59 | bert.encoder.layer.3.attention.output.dense | Linear | weight | [768, 576] | 442368 | 308224 | 0.303241 |
|
63 |
+
| 60 | bert.encoder.layer.3.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
64 |
+
| 61 | bert.encoder.layer.3.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
65 |
+
| 62 | bert.encoder.layer.3.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
66 |
+
| 63 | bert.encoder.layer.3.intermediate.dense | Linear | weight | [368, 768] | 282624 | 282624 | 0 |
|
67 |
+
| 64 | bert.encoder.layer.3.intermediate.dense | Linear | bias | [368] | 368 | 368 | 0 |
|
68 |
+
| 65 | bert.encoder.layer.3.output.dense | Linear | weight | [768, 368] | 282624 | 282624 | 0 |
|
69 |
+
| 66 | bert.encoder.layer.3.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
70 |
+
| 67 | bert.encoder.layer.3.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
71 |
+
| 68 | bert.encoder.layer.3.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
72 |
+
| 69 | bert.encoder.layer.4.attention.self.query | Linear | weight | [576, 768] | 442368 | 291840 | 0.340278 |
|
73 |
+
| 70 | bert.encoder.layer.4.attention.self.query | Linear | bias | [576] | 576 | 544 | 0.0555555 |
|
74 |
+
| 71 | bert.encoder.layer.4.attention.self.key | Linear | weight | [576, 768] | 442368 | 310272 | 0.298611 |
|
75 |
+
| 72 | bert.encoder.layer.4.attention.self.key | Linear | bias | [576] | 576 | 544 | 0.0555555 |
|
76 |
+
| 73 | bert.encoder.layer.4.attention.self.value | Linear | weight | [576, 768] | 442368 | 272384 | 0.384259 |
|
77 |
+
| 74 | bert.encoder.layer.4.attention.self.value | Linear | bias | [576] | 576 | 480 | 0.166667 |
|
78 |
+
| 75 | bert.encoder.layer.4.attention.output.dense | Linear | weight | [768, 576] | 442368 | 263168 | 0.405093 |
|
79 |
+
| 76 | bert.encoder.layer.4.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
80 |
+
| 77 | bert.encoder.layer.4.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
81 |
+
| 78 | bert.encoder.layer.4.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
82 |
+
| 79 | bert.encoder.layer.4.intermediate.dense | Linear | weight | [386, 768] | 296448 | 296448 | 0 |
|
83 |
+
| 80 | bert.encoder.layer.4.intermediate.dense | Linear | bias | [386] | 386 | 386 | 0 |
|
84 |
+
| 81 | bert.encoder.layer.4.output.dense | Linear | weight | [768, 386] | 296448 | 296448 | 0 |
|
85 |
+
| 82 | bert.encoder.layer.4.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
86 |
+
| 83 | bert.encoder.layer.4.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
87 |
+
| 84 | bert.encoder.layer.4.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
88 |
+
| 85 | bert.encoder.layer.5.attention.self.query | Linear | weight | [384, 768] | 294912 | 171008 | 0.420139 |
|
89 |
+
| 86 | bert.encoder.layer.5.attention.self.query | Linear | bias | [384] | 384 | 352 | 0.0833333 |
|
90 |
+
| 87 | bert.encoder.layer.5.attention.self.key | Linear | weight | [384, 768] | 294912 | 205824 | 0.302083 |
|
91 |
+
| 88 | bert.encoder.layer.5.attention.self.key | Linear | bias | [384] | 384 | 352 | 0.0833333 |
|
92 |
+
| 89 | bert.encoder.layer.5.attention.self.value | Linear | weight | [384, 768] | 294912 | 217088 | 0.263889 |
|
93 |
+
| 90 | bert.encoder.layer.5.attention.self.value | Linear | bias | [384] | 384 | 384 | 0 |
|
94 |
+
| 91 | bert.encoder.layer.5.attention.output.dense | Linear | weight | [768, 384] | 294912 | 223232 | 0.243056 |
|
95 |
+
| 92 | bert.encoder.layer.5.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
96 |
+
| 93 | bert.encoder.layer.5.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
97 |
+
| 94 | bert.encoder.layer.5.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
98 |
+
| 95 | bert.encoder.layer.5.intermediate.dense | Linear | weight | [336, 768] | 258048 | 258048 | 0 |
|
99 |
+
| 96 | bert.encoder.layer.5.intermediate.dense | Linear | bias | [336] | 336 | 336 | 0 |
|
100 |
+
| 97 | bert.encoder.layer.5.output.dense | Linear | weight | [768, 336] | 258048 | 258048 | 0 |
|
101 |
+
| 98 | bert.encoder.layer.5.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
102 |
+
| 99 | bert.encoder.layer.5.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
103 |
+
| 100 | bert.encoder.layer.5.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
104 |
+
| 101 | bert.encoder.layer.6.attention.self.query | Linear | weight | [448, 768] | 344064 | 192512 | 0.440476 |
|
105 |
+
| 102 | bert.encoder.layer.6.attention.self.query | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
106 |
+
| 103 | bert.encoder.layer.6.attention.self.key | Linear | weight | [448, 768] | 344064 | 224256 | 0.348214 |
|
107 |
+
| 104 | bert.encoder.layer.6.attention.self.key | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
108 |
+
| 105 | bert.encoder.layer.6.attention.self.value | Linear | weight | [448, 768] | 344064 | 209920 | 0.389881 |
|
109 |
+
| 106 | bert.encoder.layer.6.attention.self.value | Linear | bias | [448] | 448 | 352 | 0.214286 |
|
110 |
+
| 107 | bert.encoder.layer.6.attention.output.dense | Linear | weight | [768, 448] | 344064 | 199680 | 0.419643 |
|
111 |
+
| 108 | bert.encoder.layer.6.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
112 |
+
| 109 | bert.encoder.layer.6.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
113 |
+
| 110 | bert.encoder.layer.6.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
114 |
+
| 111 | bert.encoder.layer.6.intermediate.dense | Linear | weight | [280, 768] | 215040 | 215040 | 0 |
|
115 |
+
| 112 | bert.encoder.layer.6.intermediate.dense | Linear | bias | [280] | 280 | 280 | 0 |
|
116 |
+
| 113 | bert.encoder.layer.6.output.dense | Linear | weight | [768, 280] | 215040 | 215040 | 0 |
|
117 |
+
| 114 | bert.encoder.layer.6.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
118 |
+
| 115 | bert.encoder.layer.6.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
119 |
+
| 116 | bert.encoder.layer.6.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
120 |
+
| 117 | bert.encoder.layer.7.attention.self.query | Linear | weight | [448, 768] | 344064 | 201728 | 0.41369 |
|
121 |
+
| 118 | bert.encoder.layer.7.attention.self.query | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
122 |
+
| 119 | bert.encoder.layer.7.attention.self.key | Linear | weight | [448, 768] | 344064 | 237568 | 0.309524 |
|
123 |
+
| 120 | bert.encoder.layer.7.attention.self.key | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
124 |
+
| 121 | bert.encoder.layer.7.attention.self.value | Linear | weight | [448, 768] | 344064 | 218112 | 0.366071 |
|
125 |
+
| 122 | bert.encoder.layer.7.attention.self.value | Linear | bias | [448] | 448 | 352 | 0.214286 |
|
126 |
+
| 123 | bert.encoder.layer.7.attention.output.dense | Linear | weight | [768, 448] | 344064 | 202752 | 0.410714 |
|
127 |
+
| 124 | bert.encoder.layer.7.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
128 |
+
| 125 | bert.encoder.layer.7.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
129 |
+
| 126 | bert.encoder.layer.7.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
130 |
+
| 127 | bert.encoder.layer.7.intermediate.dense | Linear | weight | [211, 768] | 162048 | 162048 | 0 |
|
131 |
+
| 128 | bert.encoder.layer.7.intermediate.dense | Linear | bias | [211] | 211 | 211 | 0 |
|
132 |
+
| 129 | bert.encoder.layer.7.output.dense | Linear | weight | [768, 211] | 162048 | 162048 | 0 |
|
133 |
+
| 130 | bert.encoder.layer.7.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
134 |
+
| 131 | bert.encoder.layer.7.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
135 |
+
| 132 | bert.encoder.layer.7.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
136 |
+
| 133 | bert.encoder.layer.8.attention.self.query | Linear | weight | [448, 768] | 344064 | 186368 | 0.458333 |
|
137 |
+
| 134 | bert.encoder.layer.8.attention.self.query | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
138 |
+
| 135 | bert.encoder.layer.8.attention.self.key | Linear | weight | [448, 768] | 344064 | 197632 | 0.425595 |
|
139 |
+
| 136 | bert.encoder.layer.8.attention.self.key | Linear | bias | [448] | 448 | 416 | 0.0714285 |
|
140 |
+
| 137 | bert.encoder.layer.8.attention.self.value | Linear | weight | [448, 768] | 344064 | 154624 | 0.550595 |
|
141 |
+
| 138 | bert.encoder.layer.8.attention.self.value | Linear | bias | [448] | 448 | 288 | 0.357143 |
|
142 |
+
| 139 | bert.encoder.layer.8.attention.output.dense | Linear | weight | [768, 448] | 344064 | 148480 | 0.568452 |
|
143 |
+
| 140 | bert.encoder.layer.8.attention.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
144 |
+
| 141 | bert.encoder.layer.8.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
145 |
+
| 142 | bert.encoder.layer.8.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
146 |
+
| 143 | bert.encoder.layer.8.intermediate.dense | Linear | weight | [108, 768] | 82944 | 82944 | 0 |
|
147 |
+
| 144 | bert.encoder.layer.8.intermediate.dense | Linear | bias | [108] | 108 | 108 | 0 |
|
148 |
+
| 145 | bert.encoder.layer.8.output.dense | Linear | weight | [768, 108] | 82944 | 82944 | 0 |
|
149 |
+
| 146 | bert.encoder.layer.8.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
150 |
+
| 147 | bert.encoder.layer.8.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
151 |
+
| 148 | bert.encoder.layer.8.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
152 |
+
| 149 | bert.encoder.layer.9.attention.self.query | Linear | weight | [320, 768] | 245760 | 144384 | 0.4125 |
|
153 |
+
| 150 | bert.encoder.layer.9.attention.self.query | Linear | bias | [320] | 320 | 288 | 0.1 |
|
154 |
+
| 151 | bert.encoder.layer.9.attention.self.key | Linear | weight | [320, 768] | 245760 | 155648 | 0.366667 |
|
155 |
+
| 152 | bert.encoder.layer.9.attention.self.key | Linear | bias | [320] | 320 | 288 | 0.1 |
|
156 |
+
| 153 | bert.encoder.layer.9.attention.self.value | Linear | weight | [320, 768] | 245760 | 63488 | 0.741667 |
|
157 |
+
| 154 | bert.encoder.layer.9.attention.self.value | Linear | bias | [320] | 320 | 160 | 0.5 |
|
158 |
+
| 155 | bert.encoder.layer.9.attention.output.dense | Linear | weight | [768, 320] | 245760 | 65536 | 0.733333 |
|
159 |
+
| 156 | bert.encoder.layer.9.attention.output.dense | Linear | bias | [768] | 768 | 704 | 0.0833333 |
|
160 |
+
| 157 | bert.encoder.layer.9.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
161 |
+
| 158 | bert.encoder.layer.9.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
162 |
+
| 159 | bert.encoder.layer.9.intermediate.dense | Linear | weight | [53, 768] | 40704 | 40704 | 5.96046e-08 |
|
163 |
+
| 160 | bert.encoder.layer.9.intermediate.dense | Linear | bias | [53] | 53 | 53 | 0 |
|
164 |
+
| 161 | bert.encoder.layer.9.output.dense | Linear | weight | [768, 53] | 40704 | 40704 | 5.96046e-08 |
|
165 |
+
| 162 | bert.encoder.layer.9.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
166 |
+
| 163 | bert.encoder.layer.9.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
167 |
+
| 164 | bert.encoder.layer.9.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
168 |
+
| 165 | bert.encoder.layer.10.attention.self.query | Linear | weight | [384, 768] | 294912 | 158720 | 0.461806 |
|
169 |
+
| 166 | bert.encoder.layer.10.attention.self.query | Linear | bias | [384] | 384 | 320 | 0.166667 |
|
170 |
+
| 167 | bert.encoder.layer.10.attention.self.key | Linear | weight | [384, 768] | 294912 | 158720 | 0.461806 |
|
171 |
+
| 168 | bert.encoder.layer.10.attention.self.key | Linear | bias | [384] | 384 | 320 | 0.166667 |
|
172 |
+
| 169 | bert.encoder.layer.10.attention.self.value | Linear | weight | [384, 768] | 294912 | 77824 | 0.736111 |
|
173 |
+
| 170 | bert.encoder.layer.10.attention.self.value | Linear | bias | [384] | 384 | 192 | 0.5 |
|
174 |
+
| 171 | bert.encoder.layer.10.attention.output.dense | Linear | weight | [768, 384] | 294912 | 78848 | 0.732639 |
|
175 |
+
| 172 | bert.encoder.layer.10.attention.output.dense | Linear | bias | [768] | 768 | 736 | 0.0416666 |
|
176 |
+
| 173 | bert.encoder.layer.10.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
177 |
+
| 174 | bert.encoder.layer.10.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
178 |
+
| 175 | bert.encoder.layer.10.intermediate.dense | Linear | weight | [86, 768] | 66048 | 66048 | 0 |
|
179 |
+
| 176 | bert.encoder.layer.10.intermediate.dense | Linear | bias | [86] | 86 | 86 | 0 |
|
180 |
+
| 177 | bert.encoder.layer.10.output.dense | Linear | weight | [768, 86] | 66048 | 66048 | 0 |
|
181 |
+
| 178 | bert.encoder.layer.10.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
182 |
+
| 179 | bert.encoder.layer.10.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
183 |
+
| 180 | bert.encoder.layer.10.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
184 |
+
| 181 | bert.encoder.layer.11.attention.self.query | Linear | weight | [384, 768] | 294912 | 107520 | 0.635417 |
|
185 |
+
| 182 | bert.encoder.layer.11.attention.self.query | Linear | bias | [384] | 384 | 256 | 0.333333 |
|
186 |
+
| 183 | bert.encoder.layer.11.attention.self.key | Linear | weight | [384, 768] | 294912 | 118784 | 0.597222 |
|
187 |
+
| 184 | bert.encoder.layer.11.attention.self.key | Linear | bias | [384] | 384 | 256 | 0.333333 |
|
188 |
+
| 185 | bert.encoder.layer.11.attention.self.value | Linear | weight | [384, 768] | 294912 | 62464 | 0.788194 |
|
189 |
+
| 186 | bert.encoder.layer.11.attention.self.value | Linear | bias | [384] | 384 | 192 | 0.5 |
|
190 |
+
| 187 | bert.encoder.layer.11.attention.output.dense | Linear | weight | [768, 384] | 294912 | 54272 | 0.815972 |
|
191 |
+
| 188 | bert.encoder.layer.11.attention.output.dense | Linear | bias | [768] | 768 | 672 | 0.125 |
|
192 |
+
| 189 | bert.encoder.layer.11.attention.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
193 |
+
| 190 | bert.encoder.layer.11.attention.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
194 |
+
| 191 | bert.encoder.layer.11.intermediate.dense | Linear | weight | [105, 768] | 80640 | 80640 | 0 |
|
195 |
+
| 192 | bert.encoder.layer.11.intermediate.dense | Linear | bias | [105] | 105 | 105 | 0 |
|
196 |
+
| 193 | bert.encoder.layer.11.output.dense | Linear | weight | [768, 105] | 80640 | 80640 | 0 |
|
197 |
+
| 194 | bert.encoder.layer.11.output.dense | Linear | bias | [768] | 768 | 768 | 0 |
|
198 |
+
| 195 | bert.encoder.layer.11.output.LayerNorm | LayerNorm | weight | [768] | 768 | 768 | 0 |
|
199 |
+
| 196 | bert.encoder.layer.11.output.LayerNorm | LayerNorm | bias | [768] | 768 | 768 | 0 |
|
200 |
+
| 197 | qa_outputs | Linear | weight | [2, 768] | 1536 | 1536 | 0 |
|
201 |
+
| 198 | qa_outputs | Linear | bias | [2] | 2 | 2 | 0 |
|
XP_linear_layer_sparsity_20M_params_33.64_sparsity.csv
ADDED
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
,layer_id,layer_type,param_type,shape,nparam,nnz,sparsity
|
2 |
+
5,bert.encoder.layer.0.attention.self.query,Linear,weight,"[320, 768]",245760,135168,0.44999998807907104
|
3 |
+
7,bert.encoder.layer.0.attention.self.key,Linear,weight,"[320, 768]",245760,149504,0.3916666507720947
|
4 |
+
9,bert.encoder.layer.0.attention.self.value,Linear,weight,"[320, 768]",245760,173056,0.2958332896232605
|
5 |
+
11,bert.encoder.layer.0.attention.output.dense,Linear,weight,"[768, 320]",245760,181248,0.26249998807907104
|
6 |
+
15,bert.encoder.layer.0.intermediate.dense,Linear,weight,"[185, 768]",142080,142080,0.0
|
7 |
+
17,bert.encoder.layer.0.output.dense,Linear,weight,"[768, 185]",142080,142080,0.0
|
8 |
+
21,bert.encoder.layer.1.attention.self.query,Linear,weight,"[320, 768]",245760,175104,0.28749996423721313
|
9 |
+
23,bert.encoder.layer.1.attention.self.key,Linear,weight,"[320, 768]",245760,177152,0.27916663885116577
|
10 |
+
25,bert.encoder.layer.1.attention.self.value,Linear,weight,"[320, 768]",245760,166912,0.32083332538604736
|
11 |
+
27,bert.encoder.layer.1.attention.output.dense,Linear,weight,"[768, 320]",245760,167936,0.3166666030883789
|
12 |
+
31,bert.encoder.layer.1.intermediate.dense,Linear,weight,"[315, 768]",241920,241920,0.0
|
13 |
+
33,bert.encoder.layer.1.output.dense,Linear,weight,"[768, 315]",241920,241920,0.0
|
14 |
+
37,bert.encoder.layer.2.attention.self.query,Linear,weight,"[576, 768]",442368,285696,0.3541666865348816
|
15 |
+
39,bert.encoder.layer.2.attention.self.key,Linear,weight,"[576, 768]",442368,297984,0.3263888955116272
|
16 |
+
41,bert.encoder.layer.2.attention.self.value,Linear,weight,"[576, 768]",442368,226304,0.4884259104728699
|
17 |
+
43,bert.encoder.layer.2.attention.output.dense,Linear,weight,"[768, 576]",442368,237568,0.4629629850387573
|
18 |
+
47,bert.encoder.layer.2.intermediate.dense,Linear,weight,"[339, 768]",260352,260352,0.0
|
19 |
+
49,bert.encoder.layer.2.output.dense,Linear,weight,"[768, 339]",260352,260352,0.0
|
20 |
+
53,bert.encoder.layer.3.attention.self.query,Linear,weight,"[576, 768]",442368,277504,0.37268519401550293
|
21 |
+
55,bert.encoder.layer.3.attention.self.key,Linear,weight,"[576, 768]",442368,303104,0.31481480598449707
|
22 |
+
57,bert.encoder.layer.3.attention.self.value,Linear,weight,"[576, 768]",442368,297984,0.3263888955116272
|
23 |
+
59,bert.encoder.layer.3.attention.output.dense,Linear,weight,"[768, 576]",442368,308224,0.30324071645736694
|
24 |
+
63,bert.encoder.layer.3.intermediate.dense,Linear,weight,"[368, 768]",282624,282624,0.0
|
25 |
+
65,bert.encoder.layer.3.output.dense,Linear,weight,"[768, 368]",282624,282624,0.0
|
26 |
+
69,bert.encoder.layer.4.attention.self.query,Linear,weight,"[576, 768]",442368,291840,0.3402777910232544
|
27 |
+
71,bert.encoder.layer.4.attention.self.key,Linear,weight,"[576, 768]",442368,310272,0.2986111044883728
|
28 |
+
73,bert.encoder.layer.4.attention.self.value,Linear,weight,"[576, 768]",442368,272384,0.38425928354263306
|
29 |
+
75,bert.encoder.layer.4.attention.output.dense,Linear,weight,"[768, 576]",442368,263168,0.40509259700775146
|
30 |
+
79,bert.encoder.layer.4.intermediate.dense,Linear,weight,"[386, 768]",296448,296448,0.0
|
31 |
+
81,bert.encoder.layer.4.output.dense,Linear,weight,"[768, 386]",296448,296448,0.0
|
32 |
+
85,bert.encoder.layer.5.attention.self.query,Linear,weight,"[384, 768]",294912,171008,0.4201388955116272
|
33 |
+
87,bert.encoder.layer.5.attention.self.key,Linear,weight,"[384, 768]",294912,205824,0.3020833134651184
|
34 |
+
89,bert.encoder.layer.5.attention.self.value,Linear,weight,"[384, 768]",294912,217088,0.2638888955116272
|
35 |
+
91,bert.encoder.layer.5.attention.output.dense,Linear,weight,"[768, 384]",294912,223232,0.243055522441864
|
36 |
+
95,bert.encoder.layer.5.intermediate.dense,Linear,weight,"[336, 768]",258048,258048,0.0
|
37 |
+
97,bert.encoder.layer.5.output.dense,Linear,weight,"[768, 336]",258048,258048,0.0
|
38 |
+
101,bert.encoder.layer.6.attention.self.query,Linear,weight,"[448, 768]",344064,192512,0.4404761791229248
|
39 |
+
103,bert.encoder.layer.6.attention.self.key,Linear,weight,"[448, 768]",344064,224256,0.3482142686843872
|
40 |
+
105,bert.encoder.layer.6.attention.self.value,Linear,weight,"[448, 768]",344064,209920,0.3898809552192688
|
41 |
+
107,bert.encoder.layer.6.attention.output.dense,Linear,weight,"[768, 448]",344064,199680,0.4196428656578064
|
42 |
+
111,bert.encoder.layer.6.intermediate.dense,Linear,weight,"[280, 768]",215040,215040,0.0
|
43 |
+
113,bert.encoder.layer.6.output.dense,Linear,weight,"[768, 280]",215040,215040,0.0
|
44 |
+
117,bert.encoder.layer.7.attention.self.query,Linear,weight,"[448, 768]",344064,201728,0.413690447807312
|
45 |
+
119,bert.encoder.layer.7.attention.self.key,Linear,weight,"[448, 768]",344064,237568,0.3095238208770752
|
46 |
+
121,bert.encoder.layer.7.attention.self.value,Linear,weight,"[448, 768]",344064,218112,0.3660714030265808
|
47 |
+
123,bert.encoder.layer.7.attention.output.dense,Linear,weight,"[768, 448]",344064,202752,0.4107142686843872
|
48 |
+
127,bert.encoder.layer.7.intermediate.dense,Linear,weight,"[211, 768]",162048,162048,0.0
|
49 |
+
129,bert.encoder.layer.7.output.dense,Linear,weight,"[768, 211]",162048,162048,0.0
|
50 |
+
133,bert.encoder.layer.8.attention.self.query,Linear,weight,"[448, 768]",344064,186368,0.4583333134651184
|
51 |
+
135,bert.encoder.layer.8.attention.self.key,Linear,weight,"[448, 768]",344064,197632,0.425595223903656
|
52 |
+
137,bert.encoder.layer.8.attention.self.value,Linear,weight,"[448, 768]",344064,154624,0.550595223903656
|
53 |
+
139,bert.encoder.layer.8.attention.output.dense,Linear,weight,"[768, 448]",344064,148480,0.5684523582458496
|
54 |
+
143,bert.encoder.layer.8.intermediate.dense,Linear,weight,"[108, 768]",82944,82944,0.0
|
55 |
+
145,bert.encoder.layer.8.output.dense,Linear,weight,"[768, 108]",82944,82944,0.0
|
56 |
+
149,bert.encoder.layer.9.attention.self.query,Linear,weight,"[320, 768]",245760,144384,0.41249996423721313
|
57 |
+
151,bert.encoder.layer.9.attention.self.key,Linear,weight,"[320, 768]",245760,155648,0.36666661500930786
|
58 |
+
153,bert.encoder.layer.9.attention.self.value,Linear,weight,"[320, 768]",245760,63488,0.7416666746139526
|
59 |
+
155,bert.encoder.layer.9.attention.output.dense,Linear,weight,"[768, 320]",245760,65536,0.7333333492279053
|
60 |
+
159,bert.encoder.layer.9.intermediate.dense,Linear,weight,"[53, 768]",40704,40704,5.960464477539063e-08
|
61 |
+
161,bert.encoder.layer.9.output.dense,Linear,weight,"[768, 53]",40704,40704,5.960464477539063e-08
|
62 |
+
165,bert.encoder.layer.10.attention.self.query,Linear,weight,"[384, 768]",294912,158720,0.461805522441864
|
63 |
+
167,bert.encoder.layer.10.attention.self.key,Linear,weight,"[384, 768]",294912,158720,0.461805522441864
|
64 |
+
169,bert.encoder.layer.10.attention.self.value,Linear,weight,"[384, 768]",294912,77824,0.7361111044883728
|
65 |
+
171,bert.encoder.layer.10.attention.output.dense,Linear,weight,"[768, 384]",294912,78848,0.7326388955116272
|
66 |
+
175,bert.encoder.layer.10.intermediate.dense,Linear,weight,"[86, 768]",66048,66048,0.0
|
67 |
+
177,bert.encoder.layer.10.output.dense,Linear,weight,"[768, 86]",66048,66048,0.0
|
68 |
+
181,bert.encoder.layer.11.attention.self.query,Linear,weight,"[384, 768]",294912,107520,0.6354166269302368
|
69 |
+
183,bert.encoder.layer.11.attention.self.key,Linear,weight,"[384, 768]",294912,118784,0.5972222089767456
|
70 |
+
185,bert.encoder.layer.11.attention.self.value,Linear,weight,"[384, 768]",294912,62464,0.7881944179534912
|
71 |
+
187,bert.encoder.layer.11.attention.output.dense,Linear,weight,"[768, 384]",294912,54272,0.8159722089767456
|
72 |
+
191,bert.encoder.layer.11.intermediate.dense,Linear,weight,"[105, 768]",80640,80640,0.0
|
73 |
+
193,bert.encoder.layer.11.output.dense,Linear,weight,"[768, 105]",80640,80640,0.0
|
XP_linear_layer_sparsity_20M_params_33.64_sparsity.md
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
| | layer_id | layer_type | param_type | shape | nparam | nnz | sparsity |
|
2 |
+
|----:|:---------------------------------------------|:-------------|:-------------|:-----------|---------:|-------:|------------:|
|
3 |
+
| 5 | bert.encoder.layer.0.attention.self.query | Linear | weight | [320, 768] | 245760 | 135168 | 0.45 |
|
4 |
+
| 7 | bert.encoder.layer.0.attention.self.key | Linear | weight | [320, 768] | 245760 | 149504 | 0.391667 |
|
5 |
+
| 9 | bert.encoder.layer.0.attention.self.value | Linear | weight | [320, 768] | 245760 | 173056 | 0.295833 |
|
6 |
+
| 11 | bert.encoder.layer.0.attention.output.dense | Linear | weight | [768, 320] | 245760 | 181248 | 0.2625 |
|
7 |
+
| 15 | bert.encoder.layer.0.intermediate.dense | Linear | weight | [185, 768] | 142080 | 142080 | 0 |
|
8 |
+
| 17 | bert.encoder.layer.0.output.dense | Linear | weight | [768, 185] | 142080 | 142080 | 0 |
|
9 |
+
| 21 | bert.encoder.layer.1.attention.self.query | Linear | weight | [320, 768] | 245760 | 175104 | 0.2875 |
|
10 |
+
| 23 | bert.encoder.layer.1.attention.self.key | Linear | weight | [320, 768] | 245760 | 177152 | 0.279167 |
|
11 |
+
| 25 | bert.encoder.layer.1.attention.self.value | Linear | weight | [320, 768] | 245760 | 166912 | 0.320833 |
|
12 |
+
| 27 | bert.encoder.layer.1.attention.output.dense | Linear | weight | [768, 320] | 245760 | 167936 | 0.316667 |
|
13 |
+
| 31 | bert.encoder.layer.1.intermediate.dense | Linear | weight | [315, 768] | 241920 | 241920 | 0 |
|
14 |
+
| 33 | bert.encoder.layer.1.output.dense | Linear | weight | [768, 315] | 241920 | 241920 | 0 |
|
15 |
+
| 37 | bert.encoder.layer.2.attention.self.query | Linear | weight | [576, 768] | 442368 | 285696 | 0.354167 |
|
16 |
+
| 39 | bert.encoder.layer.2.attention.self.key | Linear | weight | [576, 768] | 442368 | 297984 | 0.326389 |
|
17 |
+
| 41 | bert.encoder.layer.2.attention.self.value | Linear | weight | [576, 768] | 442368 | 226304 | 0.488426 |
|
18 |
+
| 43 | bert.encoder.layer.2.attention.output.dense | Linear | weight | [768, 576] | 442368 | 237568 | 0.462963 |
|
19 |
+
| 47 | bert.encoder.layer.2.intermediate.dense | Linear | weight | [339, 768] | 260352 | 260352 | 0 |
|
20 |
+
| 49 | bert.encoder.layer.2.output.dense | Linear | weight | [768, 339] | 260352 | 260352 | 0 |
|
21 |
+
| 53 | bert.encoder.layer.3.attention.self.query | Linear | weight | [576, 768] | 442368 | 277504 | 0.372685 |
|
22 |
+
| 55 | bert.encoder.layer.3.attention.self.key | Linear | weight | [576, 768] | 442368 | 303104 | 0.314815 |
|
23 |
+
| 57 | bert.encoder.layer.3.attention.self.value | Linear | weight | [576, 768] | 442368 | 297984 | 0.326389 |
|
24 |
+
| 59 | bert.encoder.layer.3.attention.output.dense | Linear | weight | [768, 576] | 442368 | 308224 | 0.303241 |
|
25 |
+
| 63 | bert.encoder.layer.3.intermediate.dense | Linear | weight | [368, 768] | 282624 | 282624 | 0 |
|
26 |
+
| 65 | bert.encoder.layer.3.output.dense | Linear | weight | [768, 368] | 282624 | 282624 | 0 |
|
27 |
+
| 69 | bert.encoder.layer.4.attention.self.query | Linear | weight | [576, 768] | 442368 | 291840 | 0.340278 |
|
28 |
+
| 71 | bert.encoder.layer.4.attention.self.key | Linear | weight | [576, 768] | 442368 | 310272 | 0.298611 |
|
29 |
+
| 73 | bert.encoder.layer.4.attention.self.value | Linear | weight | [576, 768] | 442368 | 272384 | 0.384259 |
|
30 |
+
| 75 | bert.encoder.layer.4.attention.output.dense | Linear | weight | [768, 576] | 442368 | 263168 | 0.405093 |
|
31 |
+
| 79 | bert.encoder.layer.4.intermediate.dense | Linear | weight | [386, 768] | 296448 | 296448 | 0 |
|
32 |
+
| 81 | bert.encoder.layer.4.output.dense | Linear | weight | [768, 386] | 296448 | 296448 | 0 |
|
33 |
+
| 85 | bert.encoder.layer.5.attention.self.query | Linear | weight | [384, 768] | 294912 | 171008 | 0.420139 |
|
34 |
+
| 87 | bert.encoder.layer.5.attention.self.key | Linear | weight | [384, 768] | 294912 | 205824 | 0.302083 |
|
35 |
+
| 89 | bert.encoder.layer.5.attention.self.value | Linear | weight | [384, 768] | 294912 | 217088 | 0.263889 |
|
36 |
+
| 91 | bert.encoder.layer.5.attention.output.dense | Linear | weight | [768, 384] | 294912 | 223232 | 0.243056 |
|
37 |
+
| 95 | bert.encoder.layer.5.intermediate.dense | Linear | weight | [336, 768] | 258048 | 258048 | 0 |
|
38 |
+
| 97 | bert.encoder.layer.5.output.dense | Linear | weight | [768, 336] | 258048 | 258048 | 0 |
|
39 |
+
| 101 | bert.encoder.layer.6.attention.self.query | Linear | weight | [448, 768] | 344064 | 192512 | 0.440476 |
|
40 |
+
| 103 | bert.encoder.layer.6.attention.self.key | Linear | weight | [448, 768] | 344064 | 224256 | 0.348214 |
|
41 |
+
| 105 | bert.encoder.layer.6.attention.self.value | Linear | weight | [448, 768] | 344064 | 209920 | 0.389881 |
|
42 |
+
| 107 | bert.encoder.layer.6.attention.output.dense | Linear | weight | [768, 448] | 344064 | 199680 | 0.419643 |
|
43 |
+
| 111 | bert.encoder.layer.6.intermediate.dense | Linear | weight | [280, 768] | 215040 | 215040 | 0 |
|
44 |
+
| 113 | bert.encoder.layer.6.output.dense | Linear | weight | [768, 280] | 215040 | 215040 | 0 |
|
45 |
+
| 117 | bert.encoder.layer.7.attention.self.query | Linear | weight | [448, 768] | 344064 | 201728 | 0.41369 |
|
46 |
+
| 119 | bert.encoder.layer.7.attention.self.key | Linear | weight | [448, 768] | 344064 | 237568 | 0.309524 |
|
47 |
+
| 121 | bert.encoder.layer.7.attention.self.value | Linear | weight | [448, 768] | 344064 | 218112 | 0.366071 |
|
48 |
+
| 123 | bert.encoder.layer.7.attention.output.dense | Linear | weight | [768, 448] | 344064 | 202752 | 0.410714 |
|
49 |
+
| 127 | bert.encoder.layer.7.intermediate.dense | Linear | weight | [211, 768] | 162048 | 162048 | 0 |
|
50 |
+
| 129 | bert.encoder.layer.7.output.dense | Linear | weight | [768, 211] | 162048 | 162048 | 0 |
|
51 |
+
| 133 | bert.encoder.layer.8.attention.self.query | Linear | weight | [448, 768] | 344064 | 186368 | 0.458333 |
|
52 |
+
| 135 | bert.encoder.layer.8.attention.self.key | Linear | weight | [448, 768] | 344064 | 197632 | 0.425595 |
|
53 |
+
| 137 | bert.encoder.layer.8.attention.self.value | Linear | weight | [448, 768] | 344064 | 154624 | 0.550595 |
|
54 |
+
| 139 | bert.encoder.layer.8.attention.output.dense | Linear | weight | [768, 448] | 344064 | 148480 | 0.568452 |
|
55 |
+
| 143 | bert.encoder.layer.8.intermediate.dense | Linear | weight | [108, 768] | 82944 | 82944 | 0 |
|
56 |
+
| 145 | bert.encoder.layer.8.output.dense | Linear | weight | [768, 108] | 82944 | 82944 | 0 |
|
57 |
+
| 149 | bert.encoder.layer.9.attention.self.query | Linear | weight | [320, 768] | 245760 | 144384 | 0.4125 |
|
58 |
+
| 151 | bert.encoder.layer.9.attention.self.key | Linear | weight | [320, 768] | 245760 | 155648 | 0.366667 |
|
59 |
+
| 153 | bert.encoder.layer.9.attention.self.value | Linear | weight | [320, 768] | 245760 | 63488 | 0.741667 |
|
60 |
+
| 155 | bert.encoder.layer.9.attention.output.dense | Linear | weight | [768, 320] | 245760 | 65536 | 0.733333 |
|
61 |
+
| 159 | bert.encoder.layer.9.intermediate.dense | Linear | weight | [53, 768] | 40704 | 40704 | 5.96046e-08 |
|
62 |
+
| 161 | bert.encoder.layer.9.output.dense | Linear | weight | [768, 53] | 40704 | 40704 | 5.96046e-08 |
|
63 |
+
| 165 | bert.encoder.layer.10.attention.self.query | Linear | weight | [384, 768] | 294912 | 158720 | 0.461806 |
|
64 |
+
| 167 | bert.encoder.layer.10.attention.self.key | Linear | weight | [384, 768] | 294912 | 158720 | 0.461806 |
|
65 |
+
| 169 | bert.encoder.layer.10.attention.self.value | Linear | weight | [384, 768] | 294912 | 77824 | 0.736111 |
|
66 |
+
| 171 | bert.encoder.layer.10.attention.output.dense | Linear | weight | [768, 384] | 294912 | 78848 | 0.732639 |
|
67 |
+
| 175 | bert.encoder.layer.10.intermediate.dense | Linear | weight | [86, 768] | 66048 | 66048 | 0 |
|
68 |
+
| 177 | bert.encoder.layer.10.output.dense | Linear | weight | [768, 86] | 66048 | 66048 | 0 |
|
69 |
+
| 181 | bert.encoder.layer.11.attention.self.query | Linear | weight | [384, 768] | 294912 | 107520 | 0.635417 |
|
70 |
+
| 183 | bert.encoder.layer.11.attention.self.key | Linear | weight | [384, 768] | 294912 | 118784 | 0.597222 |
|
71 |
+
| 185 | bert.encoder.layer.11.attention.self.value | Linear | weight | [384, 768] | 294912 | 62464 | 0.788194 |
|
72 |
+
| 187 | bert.encoder.layer.11.attention.output.dense | Linear | weight | [768, 384] | 294912 | 54272 | 0.815972 |
|
73 |
+
| 191 | bert.encoder.layer.11.intermediate.dense | Linear | weight | [105, 768] | 80640 | 80640 | 0 |
|
74 |
+
| 193 | bert.encoder.layer.11.output.dense | Linear | weight | [768, 105] | 80640 | 80640 | 0 |
|
all_results.json
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"eval_exact_match": 78.52412488174078,
|
3 |
+
"eval_f1": 86.41375972267356,
|
4 |
+
"eval_samples": 10784
|
5 |
+
}
|
eval_XP_results.json
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"eval_exact_match": 78.52412488174078,
|
3 |
+
"eval_f1": 86.41375972267356,
|
4 |
+
"eval_samples": 10784
|
5 |
+
}
|
eval_nbest_predictions.json
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c20dd1cd87aef1f85546fe8ca84d29cc07b2c058d81cc83b86fb02cee38c166b
|
3 |
+
size 48940269
|
eval_predictions.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|