[Common/PyTorch] Fused grouped MXFP8 requantization

[YangFei1990]

Description

Replace the group_dequantize -> group_quantize(columnwise) -> grouped_swizzle(rowwise scales) chain in group_requantize_inplace with a single kernel (NVTE_FUSED_GROUP_REQUANTIZE=0 restores the unfused path).

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

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 adds a Blackwell-only fused path for converting rowwise grouped MXFP8 tensors into GEMM-ready rowwise and columnwise representations, while retaining the existing unfused path behind an environment-variable fallback.

• Adds the fused CUDA requantization kernel and public C API.
• Integrates fused dispatch into PyTorch grouped requantization.
• Adds parameterized C++ coverage for data, scales, optional BF16 output, zero-sized groups, and capacity tails.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains within the provided follow-up-review scope.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/common/cast/fused_group_requantize.cu	Implements the Blackwell fused kernel, validation, scale-layout writes, optional BF16 materialization, and C API entry point; no eligible follow-up finding was established.
transformer_engine/pytorch/csrc/extensions/cast.cpp	Selects the fused path for supported grouped MXFP8 inputs and installs the resulting buffers and metadata, with the existing implementation retained as fallback.
tests/cpp/operator/test_fused_group_requantize_mxfp8.cu	Compares fused output against the production unfused chain across input formats, fast-math modes, zero-sized groups, capacity tails, and optional dequantization.
transformer_engine/common/include/transformer_engine/cast.h	Declares and documents the new fused grouped requantization C API and its layout and alignment contracts.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Rowwise grouped MXFP8 input] --> B{Fused path supported?}
  B -->|Yes| C[Fused grouped requantization kernel]
  C --> D[Swizzled rowwise scales]
  C --> E[Columnwise E4M3 data and scales]
  C --> F[Optional BF16 output]
  B -->|No or disabled| G[Existing dequantize, quantize, and swizzle chain]
  G --> D
  G --> E
  G --> F

Loading

_{Reviews (2): Last reviewed commit: "Merge branch 'fused_group_requantize' of..." | Re-trigger Greptile}

[Oleg-Goncharov]

Hi @YangFei1990, I’ve done some further optimization work on this implementation, and the version in this GitLab branch is ~10% faster than the current implementation in this PR:
https://gitlab-master.nvidia.com/ogoncharov/transformerengine/-/tree/pr_requantize_mxfp8

The main improvements came from using data swizzling in the TMA descriptor and the redux.sync instruction, which reduces the number of reads/writes to the shared buffer holding intermediate results.

Could you please update the PR to incorporate the corresponding changes before merging?