[sraman-rgb]

Opt-in with NVTE_NVFP4_FC1_DY_AMAX_FROM_DGLU=1. The rowwise amax is the exact dGLU epilogue amax rounded to BF16; the post-RHT columnwise amax is bounded by 4x the rowwise amax. Skips the separate grouped amax pass.

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 actionable defect was established.

Summary

Adds an opt-in path that reuses the dGLU epilogue’s maxima when quantizing FC1 gradients.

• FC1 gradient quantization can reuse the FC2 dGLU amax values.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[FC2 backward and dGLU] --> B[BF16 FC1 gradient and maxima]
  B --> C{Flag enabled and quantizer supported?}
  C -->|Yes| D[Round maximum to BF16]
  D --> E[Use rowwise maximum and columnwise bound]
  C -->|No| F[Compute maxima from gradient]
  E --> G[Quantize FC1 gradient]
  F --> G

Loading

_{Reviews (3) · Last reviewed commit: "Merge remote-tracking branch 'upstream/m..." · Reviewed by Greptile}