Update cudnnFE to v1.20.0

[ksivaman]

Description

v1.20.0 of cudnn frontend includes python exposure for various GEMM fusions. Not including as a part of #2769 to be able to test in isolation.

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

• Update cudnnFE to v1.20.0.

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

[ksivaman]

/te-ci

[greptile-apps]

Greptile Summary

This PR bumps the 3rdparty/cudnn-frontend git submodule from a pre-v1.20.0 commit (8d19d31) to the v1.20.0 release commit (d33027a), released 2026-03-16. The update brings Python bindings for several new GEMM fusions (GEMM+Amax, GEMM+SwiGLU, Grouped GEMM+SwiGLU, Grouped GEMM+dSwiGLU, NSA, and Fused RMSNorm+SiLU) into TransformerEngine as a prerequisite for upcoming integration work (tracked separately in #2769).

• Single-file change: Only the submodule pointer in 3rdparty/cudnn-frontend is modified — no TransformerEngine source files are changed.
• New capabilities available: v1.20.0 exposes Python bindings for GEMM fusions that are intended to be integrated in a follow-up PR, following the author's stated goal of testing the submodule bump in isolation first.
• No breaking changes: The update is additive; existing cuDNN graph APIs and SDPA paths are preserved.
• No tests added: The checklist indicates no new tests were added for the new GEMM fusion Python bindings, which is appropriate given that this PR is intentionally scoped to the submodule bump only.

Confidence Score: 5/5

• This PR is safe to merge — it is a minimal submodule pointer update with no source code changes in TransformerEngine itself.
• The change is limited to a single submodule pointer update. v1.20.0 is a tagged, released version of cudnn-frontend (released 2026-03-16). The update is intentionally isolated from any consumer code changes (which are deferred to GEMM + Swiglu fused Grouped MLP for MXFP8 #2769), reducing the blast radius of this change to near zero. No new logic was introduced in TransformerEngine, so there is no risk of regressions beyond what the cudnn-frontend library itself may introduce.
• No files require special attention.

Important Files Changed

Filename	Overview
3rdparty/cudnn-frontend	Submodule pointer updated from commit 8d19d31 (pre-v1.20.0) to d33027a (v1.20.0), which adds Python bindings for GEMM fusions including GEMM+Amax, GEMM+SwiGLU, Grouped GEMM+SwiGLU, Grouped GEMM+dSwiGLU, NSA, and Fused RMSNorm+SiLU. No issues found.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[TransformerEngine Main Repo] --> B[3rdparty/cudnn-frontend submodule]
    
    B --> C{Before PR\ncudnn-frontend\ncommit 8d19d31}
    B --> D{After PR\ncudnn-frontend\nv1.20.0\ncommit d33027a}

    D --> E[Python GEMM Fusions Exposed]
    E --> F[GEMM + Amax]
    E --> G[GEMM + SwiGLU]
    E --> H[Grouped GEMM + SwiGLU]
    E --> I[Grouped GEMM + dSwiGLU]
    E --> J[NSA - Native Sparse Attention]
    E --> K[Fused RMSNorm + SiLU]

    style D fill:#90EE90
    style C fill:#FFD700

Loading

_{Last reviewed commit: "Update cudnnFE to v1..."}