[xrennvidia]

A distributed (e.g. GTP) weight whose shards are not natively quantized is re-gathered as BF16. The fused grouped-MLP op assumed gathered weights were already quantized: forward quantized the shards before the gather, and backward read _columnwise_data / the NVFP4 amax straight off the gathered tensor, which fails for a module whose per-module recipe (e.g. MXFP8 experts under a global NVFP4 recipe) does not create native-quantized params.

Gather first, then quantize whatever is still unquantized with the op's per-expert weight quantizers, mirroring module/linear.py's handling of a re-gathered BF16 distributed weight. Native-quantized shards gather as quantized tensors and pass through unchanged. The FC2 backward re-gather moves ahead of the NVFP4 global-scale computation so that path no longer reads un-gathered shards.

Description

Please include a brief summary of the changes, relevant motivation and context.

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]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new blocking issues were found.

Summary

The fused grouped MLP now gathers distributed BF16 weights before quantizing them with each expert’s quantizer. Already-quantized weights pass through unchanged.

• Distributed grouped MLP weights are gathered before plain weights are quantized.
• Backward saves only the grouped input data its weight gradients use.

The three earlier, unnumbered findings are addressed: regression tests were added, CI runs both files, and both saved-input test layers set split_alignment=256. No new actionable issues were found. CUDA tests were not run.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Distributed weight shards] --> B[Gather full weights]
  B --> C{Already quantized?}
  C -->|Yes| E[Use weights unchanged]
  C -->|No| D[Quantize with per-expert quantizers]
  D --> F[Run grouped GEMM]
  E --> F
  G[Quantizers saved in forward] --> D

Loading

_{Reviews (12) · Last reviewed commit: "[PyTorch] Drop NVFP4 rowwise input copie..." · Reviewed by Greptile}

[xrennvidia]

Close because the changes have been cherry-picked into PR3545.