[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 GroupedLinear backward by routing affected grouped GEMMs through dense per-expert operations and adding the required columnwise amax allocation.

• Supports row-scaled NVFP4 tensors in either grouped-GEMM operand.
• Allocates and exposes per-column amax metadata for transposed row-scaled tensors.
• 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	Routes grouped GEMMs containing row-scaled NVFP4 operands through compatible per-expert dense GEMMs.
transformer_engine/pytorch/csrc/extensions/cast.cpp	Enables columnwise storage for row-scaled NVFP4 bulk allocation and assigns per-column amax shapes.
transformer_engine/common/cast/nvfp4/quantize_transpose_nvfp4.cuh	Makes the general columnwise amax reduction available without an unnecessary SM100 device guard.
tests/pytorch/test_grouped_linear.py	Adds bit-exact forward, dgrad, and wgrad comparisons against independent dense Linear layers.

Sequence Diagram

sequenceDiagram
  participant GL as GroupedLinear
  participant Q as NVFP4 quantization
  participant GG as general_grouped_gemm
  participant DG as general_gemm
  participant K as CUDA kernels
  GL->>Q: Quantize per-expert activations and gradients
  Q->>Q: Allocate rowwise and columnwise amax metadata
  GL->>GG: Submit grouped forward/dgrad/wgrad
  loop Each expert
    GG->>DG: Invoke dense row-scaled GEMM
    DG->>K: Quantize, transpose, and multiply
    K-->>DG: Expert output
  end
  GG-->>GL: Combined outputs and gradients

Loading

_{Reviews (2): Last reviewed commit: "Remove unhelpful comments" | Re-trigger Greptile}

[timmoon10]

LGTM, pending CI.

[timmoon10]

/te-ci