[MPS] Fix SDPA output shape when value head dim differs#176843
Closed
hvaara wants to merge 1 commit intopytorch:mainfrom
Closed
[MPS] Fix SDPA output shape when value head dim differs#176843hvaara wants to merge 1 commit intopytorch:mainfrom
hvaara wants to merge 1 commit intopytorch:mainfrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/176843
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit cf60249 with merge base 7643509 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
malfet
added
ciflow/mps
malfet
approved these changes
Mar 9, 2026
Contributor
|
@pytorchbot merge -f "Lint + MPS is green" |
Collaborator
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
4 tasks
pytorchmergebot
pushed a commit
that referenced
this pull request
Mar 17, 2026
Fix proposed by @mergennachin in #177603. The issue was introduced in #176843. Remove data-dependent branching in the MPS SDPA meta kernel so export supports dynamic seq. Update meta-dispatch test to compare only the first output and add an export regression test. @angelayi, you wrote the original meta registration and tests in #159695. Does this LGTY? Fixes #177603 Pull Request resolved: #177620 Approved by: https://github.com/malfet
mergennachin
added a commit
that referenced
this pull request
Mar 17, 2026
Summary Add float mask test coverage to test_sdpa_export_dynamic_seq_len, complementing the fix in #177620. The original regression (#177603, introduced by #176843) was triggered in practice by float attention masks — Metal SDPA requires mask dtype to match Q/K/V dtype, so real models use float masks, not bool. The existing test only covered bool masks. This also verifies across the <= 8 / > 8 seq_len boundary that was the branching condition in the buggy meta kernel. Test plan - python test/test_mps.py TestSDPA.test_sdpa_export_dynamic_seq_len
EmanueleCoradin
pushed a commit
to EmanueleCoradin/pytorch
that referenced
this pull request
Mar 30, 2026
This fixes MPS SDPA output shape for cases where `value.size(-1) != query.size(-1)`, so output now follows `(..., L, Ev)` as expected. I also added guards in Metal kernel paths that assume equal qkv head dims. Added the updated meta shape inference for the `sdpa_general_mps` path which seems to have been left out initially. Added regression coverage in `test/test_transformers.py` covering the shape semantics, and a similar one in `test/test_mps.py` that also checks for numerical parity with CPU. Fixes pytorch#176767 Pull Request resolved: pytorch#176843 Approved by: https://github.com/malfet
EmanueleCoradin
pushed a commit
to EmanueleCoradin/pytorch
that referenced
this pull request
Mar 30, 2026
Fix proposed by @mergennachin in pytorch#177603. The issue was introduced in pytorch#176843. Remove data-dependent branching in the MPS SDPA meta kernel so export supports dynamic seq. Update meta-dispatch test to compare only the first output and add an export regression test. @angelayi, you wrote the original meta registration and tests in pytorch#159695. Does this LGTY? Fixes pytorch#177603 Pull Request resolved: pytorch#177620 Approved by: https://github.com/malfet
AaronWang04
pushed a commit
to AaronWang04/pytorch
that referenced
this pull request
Mar 31, 2026
Fix proposed by @mergennachin in pytorch#177603. The issue was introduced in pytorch#176843. Remove data-dependent branching in the MPS SDPA meta kernel so export supports dynamic seq. Update meta-dispatch test to compare only the first output and add an export regression test. @angelayi, you wrote the original meta registration and tests in pytorch#159695. Does this LGTY? Fixes pytorch#177603 Pull Request resolved: pytorch#177620 Approved by: https://github.com/malfet
Jah-yee
pushed a commit
to Jah-yee/transformers-jy
that referenced
this pull request
Apr 16, 2026
… dim Workaround for PyTorch < 2.12 bug (pytorch/pytorch#176767, pytorch/pytorch#176843) where scaled_dot_product_attention on MPS produces incorrect output when value head dim != query head dim. DeepSeek models (MQA) are affected as they have different qk and v head dims. The fix pads v to match q's head dim before SDPA, then truncates output back to the original v size. Fixes huggingface#44554
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.
This fixes MPS SDPA output shape for cases where
value.size(-1) != query.size(-1), so output now follows(..., L, Ev)as expected. I also added guards in Metal kernel paths that assume equal qkv head dims.Added the updated meta shape inference for the
sdpa_general_mpspath which seems to have been left out initially.Added regression coverage in
test/test_transformers.pycovering the shape semantics, and a similar one intest/test_mps.pythat also checks for numerical parity with CPU.Fixes #176767