[PyTorch] Enable NVFP4 row-scaled (per-token) backward for GroupedLinear

[cael-ling]

Description

Extend the row-scaled NVFP4 support added for dense Linear (originally #2931 and #3206) to the MoE GroupedLinear module, so the wgrad is computed in NVFP4 instead of falling back to high precision.

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

[greptile-apps]

Greptile Summary

The PR enables row-scaled NVFP4 weight-gradient computation for PyTorch GroupedLinear by routing affected grouped GEMMs through per-expert dense kernels and allocating columnwise per-column amax metadata.

• Extends grouped GEMM dispatch to accept row-scaled NVFP4 operands in either position.
• Enables row-scaled NVFP4 bulk allocation with columnwise usage.
• Makes the portable columnwise amax reduction available across compiled architectures.
• Adds bit-exact GroupedLinear-versus-Linear backward coverage.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/cpp_extensions/gemm.py	Extends the existing row-scaled NVFP4 fallback loop from B-only operands to row-scaled operands in either GEMM position.
transformer_engine/pytorch/csrc/extensions/cast.cpp	Enables columnwise bulk storage for row-scaled NVFP4 and attaches per-column amax buffers using their allocated shapes.
transformer_engine/common/cast/nvfp4/quantize_transpose_nvfp4.cuh	Removes an unnecessary SM100 compile-time guard from the architecture-independent columnwise amax reduction.
tests/pytorch/test_grouped_linear.py	Adds bit-exact forward, dgrad, and wgrad comparison against independent dense Linear experts for quantized row-scaled backward.

Sequence Diagram

sequenceDiagram
  participant GL as GroupedLinear backward
  participant GG as general_grouped_gemm
  participant DG as per-expert general_gemm
  participant Q as NVFP4 quantize/transpose
  participant W as Weight gradients
  GL->>GG: NT wgrad with expert splits
  GG->>DG: Route each row-scaled expert
  DG->>Q: Quantize activation and gradient operands
  Q-->>DG: FP4 data and row/column amax metadata
  DG-->>GG: Expert weight gradient
  GG-->>W: Combined expert gradients

Loading

_{Reviews (3): Last reviewed commit: "Remove unnecessary comment" | Re-trigger Greptile}

[timmoon10]

LGTM, pending CI.

[timmoon10]

/te-ci