[pbkowalski]

Adds a guarded MiMo-V2.6 target-verification caller for AITER MegaMoE. The path is restricted to gfx950 TP8/EP8 A8W4 H6144/I2048/E384/top-8, uses a rank-consistent token maximum, preserves ordinary-AITER fallback, and excludes draft and prefill execution. Depends on the MiMo profile extension based on ROCm/aiter#6881.

---

CI States

Latest PR Test (Base): ❌ Run #41205431536
Latest PR Test (Extra): ❌ Run #41205431468
Latest PR Test (AMD ROCm 10): ❌ Run #41205431571

[pbkowalski]

Status: keep as draft; do not promote. The target-only caller preserves ordinary A8W4 prefill/fallback and passes graph startup, but MiMo DFLASH acceptance stabilizes near 1.0-1.3 versus ~7.9 on the promoted stack. Profiles/GPQA/OOB were not run after this quality-gate failure.

[pbkowalski]

Follow-up: the acceptance collapse is fixed. MiMo needs global expert IDs at the MegaMoE call and MegaMoE output divided by EP size before MiMo’s existing FFN reduction. Full-model drift remains unsafe, so the validated opt-in is SGLANG_MIMO_MEGA_LAYERS=1,3,7 with SGLANG_MIMO_MEGA_MAX_TOKENS=32; ordinary AITER remains the fallback everywhere else.

Quality/attribution: acceptance stays ~7.9, GPQA is 0.7374, and matched profiling reduces target verify 2.09%. Same-node full OOB parents are baseline f547cf7a9fde4d32a796dd8c0fae6b36 and candidate de2fdac7fac44da1bfde4aa410002898 (12/12 FINISHED each): throughput geomean +0.56%, TTFT speed +3.97%, ITL speed -1.09%. Excluding a noisy 8k/c4 outlier leaves throughput neutral, so keep this draft opt-in rather than broadly promoting it.