-
Notifications
You must be signed in to change notification settings - Fork 82
/
benchmarks.yml
184 lines (167 loc) · 5.25 KB
/
benchmarks.yml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
---
# --- Pretraining ---
pretrain_options: &pretrain_options
env:
TF_POPLAR_FLAGS: '--executable_cache_path=./tf_cache/'
data:
throughput:
regexp: 'throughput: *(.*?) samples\/sec'
skip: 1
mlm_loss:
regexp: 'mlm_loss: *(\d*\.\d*)'
reduction_type: "final"
nsp_loss:
regexp: 'nsp_loss: *(\d*\.\d*)'
reduction_type: "final"
output:
- [samples/sec, 'throughput']
- [mlm_loss, 'mlm_loss']
- [nsp_loss, 'nsp_loss']
config_options: &config_options
requirements_path: requirements.txt
# POD4
tf2_bert_large_pretrain_real_pod4:
<<: [*pretrain_options, *config_options]
description: |
This benchmark measures the throughput of Bert Large phase 1
and phase 2 pretraining.
parameters:
- [seqlen, phase]
- ['128', '1']
- ['384', '2']
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_large_{seqlen}_phase{phase}_pod4.json
--dataset-dir $DATASETS_DIR/wikipedia/{seqlen}/
--total-num-train-samples 300000
--enable-wandb false
tf2_bert_base_pretrain_real_pod4:
<<: [*pretrain_options, *config_options]
description: |
This benchmark measures the throughput of Bert Base phase 1
and phase 2 pretraining.
parameters:
- [seqlen, phase]
- ['128', '1']
- ['384', '2']
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_base_{seqlen}_phase{phase}_pod4.json
--dataset-dir $DATASETS_DIR/wikipedia/{seqlen}/
--total-num-train-samples 300000
--enable-wandb false
# POD16
tf2_bert_large_pretrain_real_pod16:
<<: [*pretrain_options, *config_options]
description: |
This benchmark measures the throughput of Bert Large phase 1
and phase 2 pretraining.
parameters:
- [seqlen, phase]
- ['128', '1']
- ['384', '2']
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_large_{seqlen}_phase{phase}.json
--dataset-dir $DATASETS_DIR/wikipedia/{seqlen}/
--total-num-train-samples 300000
--enable-wandb false
tf2_bert_base_pretrain_real_pod16:
<<: [*pretrain_options, *config_options]
description: |
This benchmark measures the throughput of Bert Base phase 1
and phase 2 pretraining.
parameters:
- [seqlen, phase]
- ['128', '1']
- ['384', '2']
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_base_{seqlen}_phase{phase}.json
--dataset-dir $DATASETS_DIR/wikipedia/{seqlen}/
--total-num-train-samples 300000
--enable-wandb false
# POD64
tf2_bert_large_sl128_pretrain_real_pod64_conv:
<<: [*pretrain_options, *config_options]
description: |
Tests convergence of Bert Large reference config for
phase 1 pretraining on 64 IPUs.
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_large_128_phase1_POD64.json
--dataset-dir $DATASETS_DIR/wikipedia/128/
--save-ckpt-path checkpoint/phase1/
--enable-wandb true
--wandb-entity apps-benchmarking
tf2_bert_large_sl384_pretrain_real_pod64_conv:
<<: [*pretrain_options, *config_options]
description: |
Tests convergence of Bert Large reference config for
phase 2 pretraining on 64 IPUs.
cmd: >-
python3 run_pretraining.py
--config configs/pretrain_large_384_phase2_POD64.json
--dataset-dir $DATASETS_DIR/wikipedia/384/
--save-ckpt-path checkpoint/phase2/
--pretrained-ckpt-path checkpoint/phase1/
--enable-wandb true
--wandb-entity-name apps-benchmarking
# --- SQuAD training ---
squad_options: &squad_options
data:
throughput:
regexp: 'throughput: *(.*?) samples\/sec'
skip: 2
output:
- [samples/sec, 'throughput']
tf2_bert_squad_base_train_real_pod16:
<<: [*squad_options, *config_options]
description:
This benchmark measures the throughput of Bert Base SQuAD fine tuning.
cmd: >-
python3 run_squad.py
--config configs/squad_base_checkpoint_path.json
--do-validation false
--total-num-train-samples 4000
tf2_bert_squad_large_train_real_pod16:
<<: [*squad_options, *config_options]
description:
This benchmark measures the throughput of Bert Large SQuAD fine tuning.
cmd: >-
python3 run_squad.py
--config configs/squad_large_checkpoint_path.json
--do-validation false
--total-num-train-samples 2000
tf2_bert_squad_large_train_real_pod16_conv:
<<: [*squad_options, *config_options]
description:
This benchmark trains Bert Large SQuAD fine tuning to convergence.
cmd: >-
python3 run_squad.py
--config configs/squad_large_checkpoint_path.json
--do-validation true
--pretrained-ckpt-path ./checkpoint/phase2/
--enable-wandb true
--wandb-entity apps-benchmarking
# --- GLUE training ---
glue_options: &glue_options
data:
throughput:
regexp: 'throughput: *(.*?) samples\/sec'
skip: 2
output:
- [samples/sec, 'throughput']
tf2_glue_base_train_real_pod16:
<<: [*glue_options, *config_options]
description:
This benchmark measures the throughput of Bert Base GLUE fine tuning.
cmd: >-
python3 run_seq_classification.py
--config configs/glue_base_checkpoint_path.json
--glue-task mrpc
--do-validation false
--do-test false
--total-num-train-samples 300
--enable-wandb true
--wandb-entity apps-benchmarking