[CPU] Support dynamic attention by tiling K1 when needed. #23304
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
The K1 dimension (head_dim) in attention was unconditionally left untiled, which leads large stack allocation when the dimension is dynamic. K1 is typically small (64/128 per AttentionOpDetail docs), so the original heuristic to leave it untiled was reasonable. The revision sets the tile sizes if the dimension is dynamic or it is not within typical range (<= 128). An e2e test is added. Signed-off-by: hanhanW <[email protected]>
Groverkss
reviewed
Jan 28, 2026
Comment on lines
-2399
to
-2400
| // Due to the way attention works, K1 dimensions cannot be tiled. Mark k1 | ||
| // reduction dimensions not to distribute. |
Contributor
There was a problem hiding this comment.
I wonder what I was thinking when i wrote this comment, ofcourse you can serially tile it.
Groverkss
approved these changes
Jan 28, 2026
bjacob
approved these changes
Jan 28, 2026
Contributor
|
@hanhanW this PR produces a build error while generating tests for android: |
Contributor
Author
|
It looks like it triggers a bug, looking. Let's revert it for now. |
hanhanW
added a commit
that referenced
this pull request
Jan 28, 2026
…23313) Reverts #23304 It triggers a bug on android build. To repro: `iree-compile --output-format=vm-bytecode --iree-hal-target-device=local --iree-hal-local-target-device-backends=llvm-cpu --iree-llvmcpu-target-cpu=generic --iree-llvmcpu-target-triple=aarch64-none-linux-android29 tests/e2e/linalg_ext_ops/dynamic_attention.mlir`
Contributor
Author
|
The issue happens when masking is disabled. It routes back to the other old issue: #16956 I have a local fix and I'm polishing it. |
Contributor
Author
|
#23318 fixes the issue. I'll re-land the PR once the other change is landed. |
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.
The K1 dimension (head_dim) in attention was unconditionally left untiled, which leads large stack allocation when the dimension is dynamic.
K1 is typically small (64/128 per AttentionOpDetail docs), so the original heuristic to leave it untiled was reasonable. The revision sets the tile sizes if the dimension is dynamic or it is not within typical range (<= 128).
E2E tests are added, and they have the same inputs and expected outputs like attention.mlir (which is a static version). Some backends, e.g., AMDGPU, does not support dynamic attention, so we create a new file. The test is enabled on CPU and VMVX backends in the revision.
Fixes #23277