Skip to content

Update cudnn-frontend to 1.24.0#3016

Merged
sudhakarsingh27 merged 2 commits into
NVIDIA:mainfrom
sudhakarsingh27:update_cudnn_fe_1.24
May 22, 2026
Merged

Update cudnn-frontend to 1.24.0#3016
sudhakarsingh27 merged 2 commits into
NVIDIA:mainfrom
sudhakarsingh27:update_cudnn_fe_1.24

Conversation

@sudhakarsingh27
Copy link
Copy Markdown
Member

Description

Update 3rd party submodule cudnn-frontend to 1.24.0

Fixes # (issue)

Type of change

  • Documentation change (change only to the documentation, either a fix or a new content)
  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to not work as expected)
  • Infra/Build change
  • Code refactoring

Changes

Please list the changes introduced in this PR:

  • Change A
  • Change B

Checklist:

  • I have read and followed the contributing guidelines
  • The functionality is complete
  • I have commented my code, particularly in hard-to-understand areas
  • I have made corresponding changes to the documentation
  • My changes generate no new warnings
  • I have added tests that prove my fix is effective or that my feature works
  • New and existing unit tests pass locally with my changes

Signed-off-by: Sudhakar Singh <sudhakars@nvidia.com>
@sudhakarsingh27
Copy link
Copy Markdown
Member Author

/te-ci

@greptile-apps
Copy link
Copy Markdown
Contributor

greptile-apps Bot commented May 20, 2026

Greptile Summary

This PR bumps the 3rdparty/cudnn-frontend submodule from v1.23.0 (fb682ce7) to v1.24.0 (c4a97621). No other source files are modified.

  • New in 1.24.0: RoPE is now available as a native cuDNN op (standalone or fused with SDPA; RoPE+SDPA fusion requires cuDNN ≥ 9.24.0), and SDPA backward gains support for hidden dimension d=256 (requires cuDNN ≥ 9.23.0). Both are purely additive — no existing APIs were changed or removed.
  • Scope: The change is a single-line submodule pointer update; none of the repo's cuDNN-frontend consumers (fused_attn.cpp, cudnn_utils.cpp, etc.) required modification for compatibility with the new version.

Confidence Score: 5/5

Safe to merge — the change is a single submodule pointer bump with no modifications to any source files.

This is a minimal, well-scoped dependency update. The 1.24.0 release is entirely additive (new RoPE op, wider SDPA backward dimension support), with no API removals or breaking changes. Existing TransformerEngine consumers of cudnn-frontend (fused_attn.cpp, cudnn_utils.cpp) do not require any corresponding changes.

No files require special attention.

Important Files Changed

Filename Overview
3rdparty/cudnn-frontend Submodule pointer bumped from v1.23.0 (fb682ce7) to v1.24.0 (c4a97621); adds RoPE cuDNN op and d=256 SDPA backward support — no breaking API changes.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[TransformerEngine main] -->|depends on| B[3rdparty/cudnn-frontend]
    B -->|before: v1.23.0\nfb682ce7| C[Old submodule]
    B -->|after: v1.24.0\nc4a97621| D[New submodule]
    D --> E[New: RoPE cuDNN op\nrequires cuDNN ≥ 9.24.0 for SDPA fusion]
    D --> F[New: SDPA bwd d=256\nrequires cuDNN ≥ 9.23.0]
    D --> G[Existing APIs unchanged]
Loading

Reviews (2): Last reviewed commit: "Merge branch 'main' of github.com:NVIDIA..." | Re-trigger Greptile

Copy link
Copy Markdown
Member

@ksivaman ksivaman left a comment

Choose a reason for hiding this comment

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

LGTM

@sudhakarsingh27 sudhakarsingh27 merged commit 856d075 into NVIDIA:main May 22, 2026
37 of 41 checks passed
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.

2 participants