fix(tests): dynamic tolerance for cuDNN TF32 precision on Ampere+ GPUs#771
Conversation
Zhaoxian-Wu
force-pushed
the
fix/test-numerical-precision
branch
from
c55b024 to
76a8e81
Compare
|
Thanks @Zhaoxian-Wu for your work! Alongside this one, look at the others ones and sync them with master, so the actions triggers again and should be working now since I added the fix to some of the lint errors were arising in the past. Also pass in the linting tool to address this: tests/helpers/testcases.py:56:53: E261 at least two spaces before inline comment
tests/helpers/testcases.py:58:55: E261 at least two spaces before inline comment |
Zhaoxian-Wu
force-pushed
the
fix/test-numerical-precision
branch
from
76a8e81 to
0078e37
Compare
|
Got it. I forgot to test the pycodestyle. This commit should be okay. Could you please trigger the test again to see whether it's correct? @PabloCarmona |
PabloCarmona
left a comment
PabloCarmona
left a comment
There was a problem hiding this comment.
Seems ok for me, please update rest of PRs with the tests accordantly and we can start to finally merge them. I supposed this doesn't need to be merged once the other PRs has its test cases updated.
cuDNN defaults to TF32 Tensor Cores on Ampere+ GPUs (sm>=80), causing ~1e-3 divergence vs the RPU backend's FP32 CUBLAS path. The existing hard-coded tolerances (decimal=4/6) fail on H100 (RNN) and Blackwell (RNN + Conv3d). Add hardware-adaptive probes that measure the actual cuDNN-vs-non-cuDNN divergence at test session start, then derive tolerances from the measured value. CPU tests remain at decimal=6; CUDA tests relax only as much as the current GPU requires. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: Zhaoxian Wu <wuzhaoxian97@gmail.com>
Zhaoxian-Wu
force-pushed
the
fix/test-numerical-precision
branch
from
0078e37 to
dc10cef
Compare
|
Sure, I've updated |
|
Hello @Zhaoxian-Wu, do we need to merge this finally? Or we wait until you made the update in the correspondant PR? I already merge some of them that have tests attached to it, but the #764 don't have any. Thanks for your incredible work! |
2 participants
Summary
This PR addresses issue #766, the CUDA test failures on Ampere+ GPUs caused by the precision mismatch between cuDNN's default TF32 Tensor Core path and the RPU backend's FP32 CUBLAS path.
Instead of globally disabling TF32, this follows the hardware-adaptive direction discussed in the issue:
assert_array_almost_equaldecimal tolerances from the measured divergencedecimal=6Testing
The affected CUDA tests now pass on my setup, including the previously failing RNN/LSTM and Conv3d cases on TF32-capable GPUs.
Hi @PabloCarmona, this PR helps me pass the full test suite locally. If the tests on the Git CLI pass as well, I will update the code in my other PRs accordingly. Thanks for your effort in maintaining the library.