Skip to content

[CPU] Propagate the reduction tile sizes to producers because of fusion.#23660

Merged
hanhanW merged 1 commit intomainfrom
users/hanhanW/debug-attention-aarch64
Mar 5, 2026
Merged

[CPU] Propagate the reduction tile sizes to producers because of fusion.#23660
hanhanW merged 1 commit intomainfrom
users/hanhanW/debug-attention-aarch64

Conversation

@hanhanW
Copy link
Contributor

@hanhanW hanhanW commented Mar 5, 2026

The codegen pipeline is designed to fuse produers into reduction loops for less memory footprint. Thus, the tile sizes should be propagated to producers.

Previously, it triggered the vector input sizes from lowering config, which leads to numeric issues.

Fixes #23638

ci-extra: linux_arm64_clang

The codegen pipeline is designed to fuse produers into reduction loops
for less memory footprint. Thus, the tile sizes should be propagated to
producers.

Previously, it triggered the vector input sizes from lowering config,
which leads to numeric issues.

Fixes #23638

ci-extra: linux_arm64_clang

Signed-off-by: hanhanW <[email protected]>
@hanhanW
Copy link
Contributor Author

hanhanW commented Mar 5, 2026

#23661 is the fix for the wrong codegen trigger

Copy link
Contributor

@Yu-Zhewen Yu-Zhewen left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@hanhanW hanhanW merged commit a02e85f into main Mar 5, 2026
59 of 61 checks passed
@hanhanW hanhanW deleted the users/hanhanW/debug-attention-aarch64 branch March 5, 2026 19:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Incorrect numerical results in dynamic_gather_attention.mlir on arm64 targets

2 participants