[kainzhong]

Description

Add CuTeDSL kernels for MXFP8 grouped quantization

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:

• CuTeDSL grouped MXFP8 quantization kernels

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: 4/5

[Medium impact] The change affects grouped MXFP8 quantization across supported CuTeDSL callers.

The PR appears safe to merge, with a non-blocking correction needed in the grouped PyTorch comparisons.

Findings

1.  Equal outputs fail comparison ▶

Summary

Adds CuTeDSL kernels for grouped MXFP8 quantization.

• Grouped MXFP8 kernels handle different member shapes and quantization options.
• Grouped MXFP8 calls can now use CuTeDSL when the backend supports them.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Grouped MXFP8 call] --> B{CuTeDSL enabled and supported?}
  B -- No --> C[Existing CUDA kernel]
  B -- Yes --> D{Single-tensor view eligible?}
  D -- Yes --> E[Existing single-tensor kernel]
  D -- No --> F{Grouped register kernel eligible?}
  F -- Yes --> G[Grouped register kernel]
  F -- No --> H[General grouped kernel]
  H --> I[Reduce partial bias sums when requested]

Loading

_{Reviews (1) · Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." · Reviewed by Greptile}

[greptile-apps]

Equal outputs fail comparison

extract_group_quantized_output compares raw FP8 bytes, but the reused CuTeDSL conversion helpers preserve negative zero where CUDA's grouped conversion can produce positive zero. Small negative inputs that round to zero can therefore fail these tests even though both outputs have the same value.

Fold 0x80 into 0x00 for grouped data, as extract_quantized_output already does. Keep the scale comparisons unchanged.