Replace custom SDPA with F.scaled_dot_product_attention#538
Open
stashuk-olek wants to merge 4 commits intofacebookresearch:mainfrom
Open
Replace custom SDPA with F.scaled_dot_product_attention#538stashuk-olek wants to merge 4 commits intofacebookresearch:mainfrom
stashuk-olek wants to merge 4 commits intofacebookresearch:mainfrom
Conversation
|
@stashuk-olek has exported this pull request. If you are a Meta employee, you can view the originating Diff in D92927084. |
meta-cla
bot
added
the
CLA Signed
stashuk-olek
force-pushed
the
export-D92927084
branch
from
59e7160 to
00ef2fe
Compare
stashuk-olek
added a commit
to stashuk-olek/multimodal
that referenced
this pull request
Feb 13, 2026
…arch#538) Summary: Replace the manual scaled dot-product attention implementation (matmul -> scale -> mask -> softmax -> dropout -> matmul) with PyTorch's `F.scaled_dot_product_attention` using the MATH backend. For my context on unified attention API, read https://docs.google.com/document/d/1XCZkhLtBNXGhxoYZ2-47XVvH7bQDOzPXBNZyL7Q7Fqo/edit?tab=t.0#heading=h.gzjhznhk1ejy Two numerical equivalence tests are added to verify the new implementation matches the manual computation path with `torch.allclose`. Reviewed By: OmarPavel Differential Revision: D92927084
stashuk-olek
force-pushed
the
export-D92927084
branch
from
00ef2fe to
3f5f7f5
Compare
stashuk-olek
added a commit
to stashuk-olek/multimodal
that referenced
this pull request
Feb 13, 2026
…arch#538) Summary: Replace the manual scaled dot-product attention implementation (matmul -> scale -> mask -> softmax -> dropout -> matmul) with PyTorch's `F.scaled_dot_product_attention` using the MATH backend. For my context on unified attention API, read https://docs.google.com/document/d/1XCZkhLtBNXGhxoYZ2-47XVvH7bQDOzPXBNZyL7Q7Fqo/edit?tab=t.0#heading=h.gzjhznhk1ejy Two numerical equivalence tests are added to verify the new implementation matches the manual computation path with `torch.allclose`. Reviewed By: OmarPavel Differential Revision: D92927084
… weights in FLAVA (facebookresearch#535) Summary: The `attentions` field on `TransformerOutput` and `return_attn_weights`/`head_mask` parameters in the FLAVA encoder stack were never used by any consumer. This diffs cleans it up. Later the intent is to simplify attention usage / use common API for them. Reviewed By: OmarPavel Differential Revision: D92927086
…h#536) Summary: Remove dead `head_mask`, `return_attn_weights`, and `attention_weights` from the VideoGPT stack. These features were never used by any consumer — `head_mask` was always `None` or all-ones, and `return_attn_weights` was always `False` except in tests that verified the feature itself. Reviewed By: OmarPavel Differential Revision: D92927089
Summary: After removing all consumers of `head_mask`, `return_attn_weights`, and `attn_probs` in the previous commits, the core attention module can be simplified. This commit: - Removes `head_mask` param from `scaled_dot_product_attention` and `SelfAttention.forward` - Changes return types from `Tuple[Tensor, Tensor]` to `Tensor` (no longer returning attention probabilities) - Removes `return_attn_weights` param and tuple unpacking logic from `MultiHeadAttention.forward` - Cleans up unused imports (`Tuple`, `Union`) No behavioral change — the attention computation itself is unchanged. Reviewed By: OmarPavel Differential Revision: D92927085
…arch#538) Summary: Replace the manual scaled dot-product attention implementation (matmul -> scale -> mask -> softmax -> dropout -> matmul) with PyTorch's `F.scaled_dot_product_attention` using the MATH backend. For my context on unified attention API, read https://docs.google.com/document/d/1XCZkhLtBNXGhxoYZ2-47XVvH7bQDOzPXBNZyL7Q7Fqo/edit?tab=t.0#heading=h.gzjhznhk1ejy Two numerical equivalence tests are added to verify the new implementation matches the manual computation path with `torch.allclose`. Reviewed By: OmarPavel Differential Revision: D92927084
stashuk-olek
force-pushed
the
export-D92927084
branch
from
3f5f7f5 to
6c2cfdf
Compare
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
1 participant
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:
Replace the manual scaled dot-product attention implementation (matmul -> scale -> mask -> softmax -> dropout -> matmul) with PyTorch's
F.scaled_dot_product_attentionusing the MATH backend.For my context on unified attention API, read https://docs.google.com/document/d/1XCZkhLtBNXGhxoYZ2-47XVvH7bQDOzPXBNZyL7Q7Fqo/edit?tab=t.0#heading=h.gzjhznhk1ejy
Two numerical equivalence tests are added to verify the new implementation matches the manual computation path with
torch.allclose.Reviewed By: OmarPavel
Differential Revision: D92927084