[XiaobingSuper]

Problem

MiniMax-M3 writes page-shuffled FP8 KV plus per-token K/V scales. On gfx1250, the unified fallback cannot consume those scales, and sparse prefill attempts a missing FP8 pa_fwd ASM code object and aborts the worker. The verified MI455 baseline therefore has to use BF16 KV.

Change

• dispatch dense max-q-len-1 per-token FP8 decode to the general shuffled paged-decode contract
• reuse the same contract for MiniMax sparse decode and per-token-as-decode prefill
• collapse per-kv-head cache and scale views consistently
• decline gfx1250 FP8 sparse ASM before the low-level launcher aborts; BF16 and other architectures keep their existing routes
• keep unified attention rejection for unsupported multi-row/prefix-cache contracts

Depends on ATOM #2594 and AITER #6625.

Validation

MI455/gfx1250, TP4, MiniMax-M3-MXFP4 with --kv_cache_dtype fp8:

• AITER producer-to-consumer operator test: 2 passed, cosine >= 0.999
• ATOM capability/dispatch tests: 3 passed
• service health: passed
• functional request 17 * 23: 391
• GSM8K fixed first 8: 7/8; BF16 control is 8/8, so full accuracy validation is still open and this PR remains Draft
• Ruff: passed

No performance claim is made yet.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2614 --add-label <label>

[XiaobingSuper]

GPU follow-up on C5 (4×MI455, TP4, --kv_cache_dtype fp8, eager): service health and 17*23 -> 391 pass; GSM8K 5-shot first 100 completes at 0.93 flexible / 0.93 strict. This reaches the issue threshold at the 100-sample gate. Full 1319 + CUDA graph replay remain pending; the shared node was released rather than running a >2h eager full set.

[XiaobingSuper]

Final MI455 TP4 correctness gate passes under CUDA graph: GSM8K 5-shot 1319/1319, 0.9378 flexible / 0.9386 strict, within 0.1pp of the BF16-KV baseline. Graph replay and 17*23 -> 391 pass.

Performance guardrail does not pass: random 8K/1K c4 is 186.43 tok/s, TPOT 20.58 ms versus BF16 368.03 tok/s, 10.41 ms. The PR remains Draft: the general shuffled per-token kernel is correctness-complete but needs split-K/tuned integration before recipe enablement.

[XiaobingSuper]

Superseded by #2710: FlyDSL paged decode handles MiniMax-M3 per-token FP8 KV on gfx1250 (full GSM8K 0.9348) and skips pa_fwd_asm there. Closing.