Skip to content

Fix flash-attn-cute namespace conflict on arm64#1949

Merged
JannikSt merged 1 commit intofeature/arm64-supportfrom
bugfix/flash-attn-cute-namespace-arm64
Mar 4, 2026
Merged

Fix flash-attn-cute namespace conflict on arm64#1949
JannikSt merged 1 commit intofeature/arm64-supportfrom
bugfix/flash-attn-cute-namespace-arm64

Conversation

@JannikSt
Copy link
Member

@JannikSt JannikSt commented Mar 4, 2026

Reinstall flash-attn-cute after flash-attn on arm64. Installing flash-attn overwrites the flash_attn/cute/ directory with a stub, breaking the real FA4 kernels. This causes the trainer to fail validation at startup.


Note

Cursor Bugbot is generating a summary for commit 7938742. Configure here.

@JannikSt JannikSt merged commit e7269a3 into feature/arm64-support Mar 4, 2026
9 checks passed
Copy link

@cursor cursor bot left a comment

Choose a reason for hiding this comment

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

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

uv pip install flash-attn --no-build-isolation; \
uv pip install flash-attn --no-build-isolation && \
uv pip install --reinstall --no-deps \
"flash-attn-cute @ git+https://github.com/Dao-AILab/flash-attention.git@main#subdirectory=flash_attn/cute"; \
Copy link

Choose a reason for hiding this comment

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

Unpinned reinstall breaks deterministic arm64 builds

Medium Severity

uv pip install --reinstall --no-deps "flash-attn-cute @ ...@main" bypasses uv.lock and pulls whatever commit main points to at build time. This makes Dockerfile.cuda non-deterministic and can introduce runtime incompatibilities with the pinned flash-attn/dependency set even when the lockfile is unchanged.

Fix in Cursor Fix in Web

Copy link

@chatgpt-codex-connector chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

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

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7938742c79

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

uv pip install flash-attn --no-build-isolation; \
uv pip install flash-attn --no-build-isolation && \
uv pip install --reinstall --no-deps \
"flash-attn-cute @ git+https://github.com/Dao-AILab/flash-attention.git@main#subdirectory=flash_attn/cute"; \

Choose a reason for hiding this comment

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

P1 Badge Pin flash-attn-cute reinstall to a fixed commit

Reinstalling flash-attn-cute from git+...@main makes the arm64 image build non-reproducible and can silently pull incompatible upstream changes, even when uv.lock pins flash-attn-cute to a specific commit. In practice, rebuilding the same repo revision at a later date may produce a different FA4 kernel set (or fail entirely), which is especially risky for training/runtime stability on arm64.

Useful? React with 👍 / 👎.

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.

1 participant