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

The PR adds optimized single-group MXFP8 grouped-MLP execution.

• Converts one-member grouped MXFP8 storage into regular quantized tensor views for dense forward and backward GEMMs.
• Adds single-group split metadata handling, runtime-offset specialization, and weight-scale swizzling.
• Extends grouped-MLP tests to cover one-group MXFP8 execution with bias and runtime-offset 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	Adds single-group MXFP8 tensor conversion, dense GEMM paths, metadata specialization, and corresponding backward handling.
tests/pytorch/test_grouped_mlp.py	Adds single-group MXFP8 coverage and ensures the one-group fixture uses a nonempty token split.

Sequence Diagram

sequenceDiagram
  participant Caller
  participant GroupedMLP
  participant Quantizer
  participant Kernel
  Caller->>GroupedMLP: forward(input, one split)
  GroupedMLP->>GroupedMLP: prepare single-group metadata
  GroupedMLP->>Quantizer: quantize MXFP8 activations
  Quantizer-->>GroupedMLP: grouped quantized storage
  GroupedMLP->>Kernel: fused FC1 and activation
  Kernel-->>GroupedMLP: quantized activation
  GroupedMLP->>Kernel: dense single-group FC2 GEMM
  Kernel-->>Caller: output
  Caller->>GroupedMLP: backward(grad_output)
  GroupedMLP->>Kernel: dactivation, dgrad, and wgrad
  Kernel-->>Caller: input and parameter gradients

Loading

_{Reviews (5): Last reviewed commit: "Support older cuDNN grouped MLP APIs" | 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.