Fix cuda sanitizer and as_subclass calls #138218
Conversation
albanD
added
the
release notes: python_frontend
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138218
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 3d6e927 with merge base 20af56d ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
| # different state of already cleaned up. | ||
| # Similarly other imports might already have been cleaned up so `sys` might | ||
| # be already gone as well. | ||
| # Skip exiting the mode if it outlived the runtime. |
There was a problem hiding this comment.
If we don't exit the mode, do we expect anything bad to happen? Or is it more like: the runtime is gone, sys is gone, exiting the mode is not a concern as everything is gone?
| csan.enable_cuda_sanitizer() | ||
| t = TwoTensor(torch.rand(2), torch.rand(2)) | ||
|
|
||
| t = MyT(torch.rand(2)) |
There was a problem hiding this comment.
I don't understand how this test case would have failed before. It looks like we expected a race for creating two different subclasses ?
There was a problem hiding this comment.
No just two different kind of subclass should work.
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
5 participants
This fixes 4 main issues:
Tensor.as_sublass()to properly disable modes when creating the new Tensor object just like we already do inmake_subclassandmake_wrapper_subclass. The change here is just to apply the exact same treatment to it.FixWe have test that check that this behavior happen so I guess this is not an obvious bugfix and expected behavior. Reverted that change.Tensor.as_subclass()not to propagate autograd as there is no valid backward associated here.