Fix profiling with record_shapes=True and nested tensor#82854
Fix profiling with record_shapes=True and nested tensor#82854
Conversation
…ref/masked_fill
🔗 Helpful links
✅ No Failures (1 Pending)As of commit 4ccd3b8 (more details on the Dr. CI page): Expand to see more💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
|
Is the goal to actually get shapes out, or just to not break? Because if it's the latter I think we should just change to The reason is that we're looking at better support in #81824 (currently being refactored to use the underlying data for better overhead) and then we could handle |
|
Yeah the goal is to just not break, looking forward to proper support. |
robieta
left a comment
There was a problem hiding this comment.
LGTM. (You'll need to rip out the unit test?)
|
Unit test passes as it should (nested tensors call regular mm underneath, and regular mm has correct sizes), and I'd rather not rip it out because otherwise #81824 will break nested tensor profiling again. |
Gotcha. Thanks for the explanation. |
|
@pytorchbot merge -g |
|
@pytorchbot successfully started a merge job. Check the current status here |
|
Hey @ngimel. |
) Summary: We probably want more systematic fix for other tensorImpls that don't support sizes (and also there are other places in profiler that call sizes unchecked, which should be fixed), but this is to unblock Transformer profiling. Pull Request resolved: #82854 Approved by: https://github.com/robieta Test Plan: contbuild & OSS CI, see https://hud.pytorch.org/commit/pytorch/pytorch/bdb0abb234e4914d68a99abf06184b3926b0abf0 Reviewed By: kit1980 Differential Revision: D38478641 Pulled By: ngimel fbshipit-source-id: e4e774042029d206ebf7fb6a92c84182fae8640e
We probably want more systematic fix for other tensorImpls that don't support sizes (and also there are other places in profiler that call sizes unchecked, which should be fixed), but this is to unblock Transformer profiling.