[flex attention][triton pin] use new TMA API#155771
Closed
davidberard98 wants to merge 3 commits intogh/davidberard98/374/basefrom
Closed
[flex attention][triton pin] use new TMA API#155771davidberard98 wants to merge 3 commits intogh/davidberard98/374/basefrom
davidberard98 wants to merge 3 commits intogh/davidberard98/374/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/155771
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 9c7b915 with merge base 132babe ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Triton 3.4 will remove the experimental TMA APIs: triton-lang/triton#6488. Ahead of this, we are **replacing the experimental TMA API usage with the stable TMA API** in flex attention. This means that **flex attention TMA will stop working with Triton 3.2 or Triton 3.3/3.3.1** for now (but it should work for Triton 3.4 in the PyTorch 2.8 release, and Meta-internal triton 3.3.1fb, which have the new TMA API). This PR does the following: * replace the experimental TMA APIs with the stable TMA APIs * remove the workspace args. Testing: I ran test/inductor/test_flex_attention.py on a H100, [TODO confirm results] TODO: When #153662 lands, turning on TMA support by default, it should be modified slightly to check specifically for stable TMA API support. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
Triton 3.4 will remove the experimental TMA APIs: triton-lang/triton#6488. Ahead of this, we are **replacing the experimental TMA API usage with the stable TMA API** in flex attention. This means that **flex attention TMA will stop working with Triton 3.2 or Triton 3.3/3.3.1** for now (but it should work for Triton 3.4 in the PyTorch 2.8 release, and Meta-internal triton 3.3.1fb, which have the new TMA API). This PR does the following: * replace the experimental TMA APIs with the stable TMA APIs * remove the workspace args. Testing: I ran test/inductor/test_flex_attention.py on a H100, [TODO confirm results] Note: When #153662 lands, turning on TMA support by default, it should be checking specifically for stable TMA API support (commented on PR) cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov [ghstack-poisoned]
mandroid6
approved these changes
Jun 13, 2025
Contributor
Author
|
@pytorchbot merge |
Collaborator
Merge failedReason: Approvers from one of the following sets are needed:
|
Contributor
Author
|
@pytorchbot merge |
Collaborator
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 |
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Stack from ghstack (oldest at bottom):
Triton 3.4 will remove the experimental TMA APIs: triton-lang/triton#6488. Ahead of this, we are replacing the experimental TMA API usage with the stable TMA API in flex attention. This means that flex attention TMA will stop working with Triton 3.2 or Triton 3.3/3.3.1 for now (but it should work for Triton 3.4 in the PyTorch 2.8 release, and Meta-internal triton 3.3.1fb, which have the new TMA API).
This PR does the following:
Testing: I ran test/inductor/test_flex_attention.py on a H100 with @mandroid6's PR #153662 patched in to turn on TMA [TODO: confirm results once all the local tests pass, but from the first 100 tests I ran locally, all the failing tests were also failing on #153662 alone]
Note: When #153662 lands, turning on TMA support by default, it should be checking specifically for stable TMA API support (commented on PR)
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov