Fix SDPA decomp problem #4851
Merged
Fix SDPA decomp problem #4851
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/4851
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 66c39b4 with merge base 87b38cf ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D61639074 |
hsharma35
approved these changes
Aug 22, 2024
Summary: As titled. The new `_safe_softmax` function is meant to avoid NaN issues mostly in training. For inference, we shouldn't need it so we swap with the regular softmax, which will prevent the decomposition that introduces the unsupported ops (`eq`, `logical_not` and `any`). See https://www.internalfb.com/code/fbsource/fbcode/caffe2/torch/_decomp/decompositions.py?lines=425. Reviewed By: hsharma35 Differential Revision: D61639074
facebook-github-bot
force-pushed
the
export-D61639074
branch
from
244ebc0 to
66c39b4
Compare
|
This pull request was exported from Phabricator. Differential Revision: D61639074 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
As titled. The new
_safe_softmaxfunction is meant to avoid NaN issues mostly in training. For inference, we shouldn't need it so we swap with the regular softmax, which will prevent the decomposition that introduces the unsupported ops (eq,logical_notandany). See https://www.internalfb.com/code/fbsource/fbcode/caffe2/torch/_decomp/decompositions.py?lines=425.Note that it needed some changes to
run_and_verifysince we now need some aten IR changes. I will fix it in another diff, whererun_and_verifywill use a nop quantizer instead. This way the code path will be the same for fp32 and quantized. But let's make CI green first!We will also need to formalize better how to apply passes on the initial graph module (aten IR passes as opposed to edge IR passes). Seems like lifted constants and other things like that can create issues, but unless we see errors, let's wait until the IR changes from PT/ET are in first.
Reviewed By: hsharma35
Differential Revision: D61639074