perf(flydsl): Kimi-K3 B1 route-sort + prepared-MoE handoff (stack)

[JohnQinAMD]

Summary

Banked change #5 — the route-sort + AttnRes integration (the campaign's largest
single jump, +20.3% to 60.909 tok/s). This branch is a 9-commit stack, not a
single independent change:

ac4ba2dfe3 perf(mla): expose split override for graph decode
2790dbfe9a [Kimi-K3] Fuse gfx950 B1 routing and MoE sort metadata   <- route-sort (essence of #5)
d1a3ef562d perf(moe): specialize Kimi-K3 B1 gate projection          (re-created; see PR gate)
c257ea6509 perf(moe): add Kimi-K3 B1 stage1 direct tile
8687dca545 perf(moe): use atomic Kimi-K3 B1 stage2 reduction
15d1d1744f perf(flydsl): fuse Kimi-K3 KDA decode                      (re-created; see PR KDA-FB)
95dfd13ee8 perf(flydsl): fuse Kimi-K3 f_b projection into KDA decode  (re-created; see PR KDA-FB)
d4fb7ab2b4 perf(flydsl): fuse Kimi-K3 B1 latent MoE tail              (re-created; see PR tail)
459601b680 perf(flydsl): consume Kimi-K3 prepared MoE routing         <- route->expert handoff (essence of #5)

The route-sort kernel (2790dbfe9a) does FP32-logit conversion + correction bias

• grouped top-k + both sort phases, feeding prepared AITER metadata into GEMM1;
  459601b680 consumes that metadata directly in fused_moe_2stages with no second
  grouped-top-k/sort. Route boundary 1.62x; replaced ~330 fragmented launches. The
  paired AttnRes+RMSNorm (RES-1) change lives on the vLLM side.

Overlap / supersession (read before merge)

• Commits d1a3ef562d, 15d1d1744f, 95dfd13ee8, d4fb7ab2b4 are re-created
  copies of the gate / KDA-decode / KDA-FB / tail changes that also ship as
  their own independent PRs. Do not merge both without de-duplication.
• Persistent-MoE (aiter#2) supersedes the route + gate ownership here — it
  removes the duplicate route ownership this stack introduces. A single integrated
  1->7 chain must resolve this supersession (route-sort is retained; the separate
  gate/stage tiles are absorbed by the persistent boundary) and be re-verified at
  the endpoint. See perf_changelog.md "Stacked-integration blocker".

Provenance

Banked endpoint 60.909 tok/s via composed overlay image
(route-attnres:candidate1). Independent-off-main PR for review; not the final
integration branch.

AI assistance (OpenAI Codex) was used; a human submitter must review every line
before merge.

Made with Cursor

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 7 --add-label <label>

[JohnQinAMD]

Superseded by the reviewed, minimal PR decomposition: MLA split scheduling is now #20; prepared route/expert ownership and B1 stage policy are replaced by the single-owner persistent A8W4 boundary in #2; gate projection is #6; KDA decode + f_b is #16; and latent MoE tail is #17 (with FP8 follow-up #18). The corresponding vLLM ownership boundary is #10. Closing this conflicting 9-commit integration stack avoids duplicate kernel ownership and duplicate commits; its 60.909 tok/s endpoint result remains valid provenance for the composed features.