Skip to content

BugFix for MegaMoE - #954

Merged
coderfeli merged 11 commits into
mainfrom
mega_moe_v1
Aug 6, 2026
Merged

BugFix for MegaMoE#954
coderfeli merged 11 commits into
mainfrom
mega_moe_v1

Conversation

@GwilliamHu

@GwilliamHu GwilliamHu commented Aug 2, 2026

Copy link
Copy Markdown
Member

• ## Motivation

Clean up MegaMoEV2 after its initial integration and fix synchronization and correctness issues found during review and serving tests.

Technical Details

  • Simplify MegaMoEV2 communication and buffer-access code, removing unused protocol paths and redundant logic.
  • Fix multi-rank dispatch synchronization and completion handling to prevent hangs.
  • Add missing dtype, payload-size, scale-copy, and buffer-ABI correctness checks.
  • Preserve the existing operator interface, static configuration policy, and kernel performance behavior.

Test Plan

  • Run Python style and static configuration tests.
  • Validate v4_pro A8W4 accuracy on MI355X.
  • Exercise fixed-slot and compact paths across multiple batch sizes and multi-rank token distributions.

coderfeli
coderfeli previously approved these changes Aug 3, 2026
@coderfeli

Copy link
Copy Markdown
Collaborator

CI failed? @GwilliamHu

@GwilliamHu

Copy link
Copy Markdown
Member Author

CI failed? @GwilliamHu

好像是CI容器命名出了重复的,不像是代码问题,等复跑看下~~

@coderfeli
coderfeli merged commit 9d8edc0 into main Aug 6, 2026
13 of 14 checks passed
@coderfeli
coderfeli deleted the mega_moe_v1 branch August 6, 2026 13:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants