Skip to content

perf(moe): initialize Prefill output in the first gather - #1702

Merged
shihaobai merged 1 commit into
mainfrom
bsh/glm-moe-gather-first-write
Nov 11, 2026
Merged

shihaobai merged 1 commit into
mainfrom
bsh/glm-moe-gather-first-write

Conversation

@shihaobai

@shihaobai shihaobai commented Nov 11, 2026 •

Copy link
Copy Markdown
Collaborator

Initialize expanded Prefill output in the first gather, avoiding separate zeroing. Later chunks continue to accumulate.

CUDA correctness tests and Black/Flake8 checks pass. An H200 local MoE benchmark improved from 369.15 to 350.68 µs (~5% less time), excluding DeepEP communication.

@shihaobai
shihaobai marked this pull request as ready for review November 11, 2026 13:43
@shihaobai
shihaobai force-pushed the bsh/glm-moe-gather-first-write branch from 127a27a to debf564 Compare November 11, 2026 13:46
@shihaobai
shihaobai merged commit 2a4bfda into main Nov 11, 2026
1 check passed
@shihaobai
shihaobai deleted the bsh/glm-moe-gather-first-write branch November 11, 2026 13:48
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant