NVFP4 per-token recipe

[YigongQin]

Description

Add NVFP4PerTokenBlockScaling recipe for MoE grouped GEMM forward pass with per-token global scaling.

• Forward: NVFP4 per-token fused grouped GEMM via cuDNN Frontend (global_scale_tensor)
• Backward: higher precision (BF16) via backward_override, matching the MXFP8 pattern from Add NVTE_BACKWARD_OVERRIDE=high_precision|dequantized #2644

Install cuDNN Frontend branch (required dependency)

pip install "git+https://github.com/YigongQin/cudnn-frontend.git@yigongq/pertoken-nvfp4-gemm#egg=nvidia-cudnn-frontend[cutedsl]"

If build fails due to multiple Python versions, try:

git clone -b yigongq/pertoken-nvfp4-gemm https://github.com/YigongQin/cudnn-frontend.git
cd cudnn-frontend
CUDAToolkit_ROOT=/path/to/cuda
CUDNN_PATH=$(python -c "import nvidia.cudnn; print(nvidia.cudnn.path[0])")
pip install -v .

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

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

[ptrendx]

Those md files should not be part of this PR, right?

[YigongQin]

yes, I will clean them up, currently there are drafts