Skip to content

Commit 1b4c319

Browse files
huydhnfacebook-github-bot
authored andcommitted
Allow higher fp16 tolerance for phlippe_resnet on CUDA 12.8 (#154109)
Summary: After pytorch/pytorch#154004, one of the model `phlippe_resnet` needs higher tolerance for fp16 on CUDA 12.8. I can reproduce it locally with: ``` python benchmarks/dynamo/torchbench.py --accuracy --timing --explain --print-compilation-time --inductor --device cuda --training --amp --only phlippe_resnet E0522 02:47:12.392000 2130213 site-packages/torch/_dynamo/utils.py:2949] RMSE (res-fp64): 0.00144, (ref-fp64): 0.00036 and shape=torch.Size([]). res.dtype: torch.float32, multiplier: 3.000000, tol: 0.001000, use_larger_multiplier_for_smaller_tensor: 0 ``` I'm not sure what exactly happens behind the scene, but this should help fix the CI failure. Also remove some left over expected accuracy results for CUDA 12.4 which we are not using anymore on CI for benchmark jobs. X-link: pytorch/pytorch#154109 Approved by: https://github.com/Skylion007, https://github.com/malfet Reviewed By: yangw-dev Differential Revision: D75251772 fbshipit-source-id: 1cbf629f60f84bb5d2a51ad884370edbb923c388
1 parent 316981e commit 1b4c319

File tree

1 file changed

+1
-0
lines changed

1 file changed

+1
-0
lines changed

userbenchmark/dynamo/dynamobench/torchbench.yaml

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -48,6 +48,7 @@ tolerance:
4848
- doctr_reco_predictor
4949
- drq
5050
- hf_Whisper
51+
- phlippe_resnet
5152

5253
higher_bf16:
5354
- doctr_reco_predictor

0 commit comments

Comments
 (0)