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
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
|
# Some models have large dataset that doesn't fit in memory. Lower the batch
# size to test the accuracy.
batch_size:
training:
demucs: 4
dlrm: 1024
densenet121: 4
hf_Reformer: 4
hf_T5_base: 4
timm_efficientdet: 1
llama_v2_7b_16h: 1
# reduced from 16 due to cudagraphs OOM in TorchInductor dashboard
yolov3: 8
inference:
timm_efficientdet: 32
dont_change_batch_size:
- demucs
- pytorch_struct
- pyhpc_turbulent_kinetic_energy
# https://github.com/pytorch/benchmark/pull/1656
- vision_maskrcnn
tolerance:
# Need lower tolerance on GPU. GPU kernels have non deterministic kernels for these models.
higher:
- alexnet
- attention_is_all_you_need_pytorch
- densenet121
- hf_Albert
- vgg16
- mobilenet_v3_large
- nvidia_deeprecommender
- timm_efficientdet
# These models need >1e-3 tolerance
even_higher:
- soft_actor_critic
- tacotron2
- yolov3
- timm_efficientdet
- squeezenet1_1
higher_fp16:
- doctr_reco_predictor
- drq
- hf_Whisper
higher_bf16:
- doctr_reco_predictor
- drq
- hf_Whisper
cosine: []
require_larger_multiplier_for_smaller_tensor:
- yolov3
# These benchmarks took >600s on an i9-11900K CPU
very_slow: &VERY_SLOW_MODELS
# 3339s
- hf_BigBird
# 3062s
- hf_Longformer
# 930s
- hf_T5
# These benchmarks took >60s on an i9-11900K CPU
slow:
- *VERY_SLOW_MODELS
# 137s
- BERT_pytorch
# 116s
- demucs
# 242s
- fastNLP_Bert
# 221s
- hf_Albert
# 400s
- hf_Bart
# 334s
- hf_Bert
# 187s
- hf_DistilBert
# 470s
- hf_GPT2
# 141s
- hf_Reformer
# 317s
- speech_transformer
# 99s
- vision_maskrcnn
non_deterministic:
# https://github.com/pytorch/pytorch/issues/98355
- mobilenet_v3_large
- sam_fast
dtype:
force_amp_for_fp16_bf16_models:
- DALLE2_pytorch
- doctr_det_predictor
- doctr_reco_predictor
- Super_SloMo
- tts_angular
- pyhpc_turbulent_kinetic_energy
- detectron2_fcos_r_50_fpn
force_fp16_for_bf16_models:
- vision_maskrcnn
# models in canary_models that we should run anyway
canary_models:
- torchrec_dlrm
detectron2_models: &DETECTRON2_MODELS
- detectron2_fasterrcnn_r_101_c4
- detectron2_fasterrcnn_r_101_dc5
- detectron2_fasterrcnn_r_101_fpn
- detectron2_fasterrcnn_r_50_c4
- detectron2_fasterrcnn_r_50_dc5
- detectron2_fasterrcnn_r_50_fpn
- detectron2_maskrcnn_r_101_c4
- detectron2_maskrcnn_r_101_fpn
- detectron2_maskrcnn_r_50_fpn
# These models support only train mode. So accuracy checking can't be done in
# eval mode.
only_training:
- *DETECTRON2_MODELS
- tts_angular
- tacotron2
- demucs
- hf_Reformer
- pytorch_struct
- yolov3
trt_not_yet_working:
- alexnet
- resnet18
- resnet50
- mobilenet_v2
- mnasnet1_0
- squeezenet1_1
- shufflenetv2_x1_0
- vgg16
- resnext50_32x4d
skip:
all:
# OOMs (A100 40G)
- detectron2_maskrcnn
# TIMEOUT, https://github.com/pytorch/pytorch/issues/98467
- tacotron2
# Failing in eager mode
- hf_clip
# multi gpu not always available in benchmark runners
- simple_gpt_tp_manual
device:
cpu:
# OOMs
- hf_T5_generate
# model is CUDA only
- cm3leon_generate
# timeout
- nanogpt
# timeout
- sam
# model is CUDA only
- sam_fast
# model is CUDA only
- llama_v2_7b_16h
# flaky
- stable_diffusion
# requires FBGEMM, CUDA only
- torchrec_dlrm
- simple_gpt
# works on cuda, accuracy failure on cpu
- hf_Whisper
- stable_diffusion_text_encoder
- llava
- moco
cuda: []
test:
training:
- *DETECTRON2_MODELS
# not designed for training
- pyhpc_equation_of_state
- pyhpc_isoneutral_mixing
- pyhpc_turbulent_kinetic_energy
- maml
- llama
- llama_v2_7b_16h
- simple_gpt
- sam_fast
# Model's DEFAULT_TRAIN_BSIZE is not implemented
- cm3leon_generate
- hf_T5_generate
- doctr_det_predictor
- doctr_reco_predictor
- moondream
# doesnt fit in memory
- phi_1_5
- detectron2_fcos_r_50_fpn
control_flow:
- cm3leon_generate
- detectron2_fcos_r_50_fpn
- fastNLP_Bert
- hf_Longformer
- hf_Reformer
- hf_T5_generate
- opacus_cifar10
- speech_transformer
# Models that should only run in --multiprocess mode
multiprocess:
- simple_gpt
# for these models, conv-batchnorm fusing causes big numerical churn.
# Skip them
# mnasnet1_0 and shufflenet_v2_x1_0 can pass on cpu, moco cuda only.
freezing:
cuda:
- mnasnet1_0
- moco
- shufflenet_v2_x1_0
cpu: []
accuracy:
skip:
large_models:
# Models too large to have eager, dynamo and fp64_numbers simultaneosuly
# even for 40 GB machine. We have tested accuracy for smaller version of
# these models
- hf_GPT2_large
- hf_T5_large
- timm_vision_transformer_large
# accuracy https://github.com/pytorch/pytorch/issues/93847
- maml
- llama_v2_7b_16h
- Background_Matting
- stable_diffusion_unet
eager_not_deterministic:
# Models that deterministic algorithms can not be turned on for eager mode.
- Background_Matting
- pytorch_unet
max_batch_size:
hf_GPT2: 2
pytorch_unet: 2
|
