[timmoon10]

Description

NVFP4-UE5M3 (added in #3325) previously used cuDNN Frontend GEMM kernels. This PR switches to use cuBLAS GEMM kernels so that we match NVFP4-E4M3.

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

• Expose NVFP4-UE5M3 GEMM kernels from cuBLAS.
• Replace cuDNN NVFP4-UE5M3 GEMM calls with cuBLAS calls.

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

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The changes since the previous review appear safe to merge.

Summary

The PR routes NVFP4-UE5M3 GEMMs through cuBLAS instead of the temporary cuDNN paths.

• NVFP4-UE5M3 GEMMs now use the cuBLASLt paths.
• Grouped MLP NVFP4 forward and backward use shared GEMM paths.

_{Reviews (2) · Last reviewed commit: "Fix linter warnings" · Reviewed by Greptile}

[timmoon10]

Pipeline 79760221

[kainzhong]

LGTM for the cudnn-fe workaround removal