refresh Kimi K2 FP4 fused MoE tunings (TP2 / 256 CU)#2938
Closed
xaguilar-amd wants to merge 1 commit intoROCm:mainfrom
Closed
refresh Kimi K2 FP4 fused MoE tunings (TP2 / 256 CU)#2938xaguilar-amd wants to merge 1 commit intoROCm:mainfrom
xaguilar-amd wants to merge 1 commit intoROCm:mainfrom
Conversation
Contributor
🏷️ CI GuideRuns automatically on every PR:
Extended tests (opt-in via labels):
|
Contributor
Author
|
The CI is failing due to Docker Hub rate limits, not code issues: @sunway513 Could you please help resolving this? Thanks! |
2 tasks
Collaborator
|
This PR's content was bulk-merged via #3004 ([Silo] Bulk merge: tuned GEMM and FMoE configs, merged 2026-05-02 03:16 UTC). Please close this PR as superseded. Tracking issue: ROCm/AI-Frameworks-Dashboard#141 |
sunway513
added a commit
that referenced
this pull request
May 4, 2026
Squash-merged from main commit 52c4554. Includes 5 atomic Silo PRs: - #2923 GLM-4.7 FP8 tuned/untuned FMoE configs (new) - #2938 Kimi-K2.5 FP4 fused MoE tunings (TP2 / 256 CU refresh) - #2979 MiniMax-M2.5 A8W8 blockscale GEMM tunings - #2981 DeepSeek-V3.2 MI355X tuned GEMM and FMoE configs - #2982 MiniMax-M2.5 FMoE tunings Conflict in aiter/configs/model_configs/kimik2_fp4_tuned_fmoe.csv: two blocks resolved by taking theirs (Silo). Block 1 upgrades existing M=256/N=512 rows from base kernel suffixes (w3) to tuner-discovered variants (w3_xcd4, _bnt2_persist, _sbm32, _sbm64). Block 2 is purely additive: 30+ new rows for previously-uncovered N=7168/K=1024 shapes plus a flydsl_fallback section. Driver: vLLM 0.21 freeze 2026-05-08 — Silo customers need these tunings on the AITER release wheel, not nightly. Verification gate before tag: - Kernel suffix parser smoke (Kimi-K2.5-MXFP4 1-token inference, confirm new suffixes JIT-compile without falling back) - ATOM 5-model accuracy unchanged within +/- 0.005 vs v0.1.13-rc1 - Perf delta on Kimi-K2.5 / MiniMax-M2.5 / DSv3.2 (expect flat or better) (cherry picked from commit 52c4554)
2 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Updates
aiter/configs/model_configs/kimik2_fp4_tuned_fmoe.csvwith a new round of tuned fused MoE kernel selections for Kimi K2–style FP4 MoE, tuned for MI355X.What changed
inter_dim256 / 512, expert counts 384/8 and 385/9, plusinter_dim = 1024/ 385/9 where new rows were added).flydsl_fallbackrows that used pure CK two-stage GEMMs when FlyDSL was unavailable with concrete FlyDSL MoE kernels where tuning shows a win, including populated timing / TFLOPS / bandwidth metadata where applicable.Motivation
The existing table mixed strong FlyDSL choices with a large fallback-only region. This refresh aligns the shipped config with measured best kernels for the TP2-style (256 CU) layout and extends coverage for additional intermediate / routed shapes used by the model.