Single group mxfp8 grouped mlp

[sraman-rgb]

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]

Greptile Summary

Adds single-group MXFP8 support to the fused grouped-MLP path.

• Converts one-member grouped MXFP8/NVFP4 storage into ordinary quantized tensor views for dense FC2, dgrad, and wgrad GEMMs.
• Adds cuDNN frontend-gated single-group runtime-offset handling and avoids allocating offset metadata for the optimized shared-expert path.
• Preserves unit activation-scale behavior across forward and backward execution.
• Adds single-group MXFP8 tests with bias and runtime-offset capability variants.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ops/fused/grouped_mlp.py	Extends the fused grouped-MLP forward and backward paths with single-group MXFP8 tensor conversion, dense GEMM dispatch, runtime offsets, and unit-scale handling.
tests/pytorch/test_grouped_mlp.py	Adds single-group MXFP8 coverage across bias and simulated cuDNN runtime-offset support configurations.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Single-group grouped MLP input] --> B[Quantize as grouped MXFP8 storage]
  B --> C[FC1 fused GEMM and activation]
  C --> D[Convert one-member grouped activation to MXFP8 tensor view]
  D --> E[Dense FC2 GEMM]
  E --> F[MLP output]
  F --> G[Backward dactivation]
  G --> H[Dense single-group dgrad]
  G --> I[Dense single-group wgrad]

Loading

_{Reviews (9): Last reviewed commit: "Skip runtime-offset test with older cuDN..." | Re-trigger Greptile}

[timmoon10]

Mostly fine as a hacky expedient. Remember that the single-group approach is a dead end and we should not put much more effort into it. The right approach is to implement a dense MLP fused op.

[vthumbe1503]

/te-ci pytorch

[sraman-rgb]

/te-ci pytorch

[vthumbe1503]

/te-ci pytorch

[vthumbe1503]

LGTM. Although I agree with Tim, grouped_mlp has now become grouped_mlp+dense_mlp. We probably should consider migrating it to a seperate fused op soon.