[Executorch][SDPA] Fix bug in sdpa #9105
Merged
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
This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/) [ghstack-poisoned]
kimishpatel
requested review from
iseeyuan,
jackzhxng and
larryliu0820
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9105
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 16c1fce with merge base 6daff83 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
kimishpatel
added a commit
that referenced
this pull request
Mar 10, 2025
This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/) ghstack-source-id: 270854238 Pull Request resolved: #9105
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D70922039 |
larryliu0820
approved these changes
Mar 10, 2025
iseeyuan
approved these changes
Mar 10, 2025
| /* dst */ qSplitSize * headSize; | ||
|
|
||
| int64_t size_bytes = size_per_thread * num_thread * query.element_size(); | ||
| int64_t size_bytes = size_per_thread * num_thread * query.element_size() * 4; |
There was a problem hiding this comment.
oh really good catch. Thats was left over from local debug. need to remove
There was a problem hiding this comment.
please finalize so that @spalatinate has the correct version
There was a problem hiding this comment.
oh totally forgot about this
There was a problem hiding this comment.
here is the fix pr #9492
kimishpatel
added
the
module: llm
This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/) cc larryliu0820 mergennachin cccclai helunwencser jackzhxng [ghstack-poisoned]
kimishpatel
added a commit
that referenced
this pull request
Mar 18, 2025
Pull Request resolved: #9105 This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/) ghstack-source-id: 272375653
|
This pull request was exported from Phabricator. Differential Revision: D70922039 |
This was referenced Mar 18, 2025
This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/) cc larryliu0820 mergennachin cccclai helunwencser jackzhxng [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D70922039 |
kimishpatel
added
the
release notes: ops & kernels
facebook-github-bot
merged commit 76e3fd9
into
gh/kimishpatel/159/base
82 of 83 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
Pull Request resolved: pytorch/executorch#9105 This diff fixes two bugs 1. When doing flash attention, the partical q @ k block may contain some entries that needs to be masked out. This logic had a bug. Maybe this bug also exist in PT core. I will look into that to add test and see if I can prove it. 2. Due to special handling via start_pos in SDPA it also exposed the bug in 1 when doing really long sequence prefill in chunked manner. It is probably better to just use mask though. Code has detail comments on the issue and fix. ghstack-source-id: 272776939 Differential Revision: [D70922039](https://our.internmc.facebook.com/intern/diff/D70922039/)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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):
This diff fixes two bugs
It is probably better to just use mask though.
Code has detail comments on the issue and fix.
Differential Revision: D70922039
cc @larryliu0820 @mergennachin @cccclai @helunwencser @jackzhxng