[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

This PR enables row-scaled NVFP4 backward computation for PyTorch GroupedLinear, replacing the previous unsupported path with per-expert dense GEMM execution.

• Routes row-scaled NVFP4 grouped GEMMs through existing dense general_gemm kernels.
• Adds columnwise amax allocation and metadata support for row-scaled NVFP4 transposes.
• Allows the general columnwise amax reduction kernel to run below SM 10.0.
• Adds bit-exact backward comparisons between GroupedLinear and equivalent independent Linear layers.

Confidence Score: 5/5

The PR appears safe to merge with no concrete blocking or non-blocking defects identified in the reviewed changes.

The grouped path consistently delegates row-scaled operands to existing dense kernels, while the cast allocation changes provide the corresponding per-column amax shape and the tests compare all resulting outputs and gradients bit-for-bit.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/cpp_extensions/gemm.py	Extends the existing per-expert dense GEMM fallback to row-scaled NVFP4 tensors in either grouped operand.
transformer_engine/pytorch/csrc/extensions/cast.cpp	Enables columnwise storage for row-scaled NVFP4 and allocates per-column amax metadata using the actual tensor shape.
transformer_engine/common/cast/nvfp4/quantize_transpose_nvfp4.cuh	Removes the SM 10.0 restriction from the architecture-independent columnwise amax reduction kernel.
tests/pytorch/test_grouped_linear.py	Adds bit-exact forward, dgrad, and wgrad coverage against equivalent independent dense Linear layers.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
    A[GroupedLinear backward] --> B[Row-scaled NVFP4 activation]
    B --> C{Grouped GEMM wrapper}
    C --> D[Per-expert general_gemm loop]
    D --> E[Columnwise amax and transpose]
    E --> F[NVFP4 weight-gradient GEMM]
    F --> G[Per-expert weight gradients]

Loading

_{Reviews (1): Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." | Re-trigger Greptile}