Deterministic MoE Router Kernel By Removing Global Atomic Ops

[zhongbozhu]

Description

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

Fixes NVIDIA/Megatron-LM#7221

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

[High impact] The PR appears safe to merge; no actionable issues were found.

Summary

Adds an optional fixed-order reduction for MoE auxiliary loss while preserving the existing atomic path by default.

• MoE aux-loss callers can now choose a fixed-order reduction.
• PyTorch's aux-loss call offers deterministic mode for eager and graph calls.
• The fused MLA module keeps its extension import with the package imports.

No actionable issues were found. CUDA tests were not run.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["fused_moe_aux_loss"] --> B{"deterministic?"}
  B -->|No| C["Existing atomic reduction"]
  B -->|Yes| D["Get workspace"]
  D --> E["Each block writes one partial sum"]
  E --> F["Sum partials in a fixed order"]
  C --> G["Loss and saved coefficient"]
  F --> G
  G --> H["Existing backward kernel"]

Loading

_{Reviews (3) · Last reviewed commit: "Align deterministic MoE aux-loss workspa..." · Reviewed by Greptile}

[ptrendx]

The way we handle the workspace in the other APIs is similar to CUB - we run the APIs with the nullptr arguments for the workspace once (which populates the other fields), then run it again with the proper workspace. Please align this API to that.

[zhongbozhu]

fixed

[ZhiyuLi-Nvidia]

Thanks a lot @zhongbozhu.

The kernel is deterministic on one GPU. For multi-node runs, does determinism also hold with multi-node job, do we need the same placement and a fixed NCCL algorithm? Good to put it in the docs if there's anything worth noting.

[zhongbozhu]

/te-ci pytorch