Skip to content

[Tunne] Retune kimi k2 moe configs#2625

Merged
valarLip merged 4 commits intomainfrom
wjx/retune_kimik2
Apr 6, 2026
Merged

[Tunne] Retune kimi k2 moe configs#2625
valarLip merged 4 commits intomainfrom
wjx/retune_kimik2

Conversation

@lalala-sh
Copy link
Copy Markdown
Contributor

@lalala-sh lalala-sh commented Apr 6, 2026

Motivation

Retune Kimi configs: Add the third expert group with inter_dim=385, topk=9 to provide full MoE configuration coverage for the Kimik2 model.
Fix fallback config overwrite bug in incremental tuning: result_to_csv incorrectly removes all existing flydsl_fallback rows from the CSV when writing new shape results, causing previously tuned shapes to lose their non-flydsl fallback configurations. This means those shapes cannot fall back to CK kernels when flydsl is unavailable, breaking compatibility.

Technical Details

Test Plan

Test Result

Submission Checklist

@lalala-sh lalala-sh requested a review from a team April 6, 2026 08:55
@github-actions
Copy link
Copy Markdown
Contributor

github-actions bot commented Apr 6, 2026

🏷️ CI Guide

Runs automatically on every PR:

  • ✅ Pre-checks (submodule verification, code formatting)
  • ✅ Aiter op tests (gfx942 + gfx950)
  • ✅ Triton tests (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label Tests
ci:triton-355 Run Triton tests on MI355 in addition to MI325
ci:sglang SGLang integration tests
ci:atom ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm vLLM benchmark
ci:all All of the above

Add labels via the sidebar or gh pr edit 2625 --add-label <label>

@valarLip valarLip merged commit 70f6d65 into main Apr 6, 2026
24 checks passed
@valarLip valarLip deleted the wjx/retune_kimik2 branch April 6, 2026 11:42
yzhou103 pushed a commit that referenced this pull request Apr 8, 2026
* retune kimik2 moe configs

* fix tunner fallback overwrite

* fix format
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.

2 participants