… race UNRESOLVED)

Attempted fixes vllm-project#2/vllm-project#3 for the long-context non-deterministic race in the d=128
QK CoFDA emulation: IntraWGOverlap=!UseQKEmu and UsePersistentScheduler=!UseQKEmu
(route the synchronous emu through the simpler non-overlapped / single-tile path).

BOTH INEFFECTIVE — the race persists with an identical pattern, so the cause is
NOT the overlap pipeline or the persistent scheduler. Working theory: a per-gemm
smem read race inside gemm_qk_cofda_emu vs the async TMA producer. Stopped per
systematic-debugging Iron Law (no fix vllm-project#4 without re-think).

Race scales with total emu QK-gemm calls (M_blocks x KV_blocks); bit-exact for
short context. Full handoff + repro:
vllm-mma/cair/experiments/docs/qk-emu-longcontext-race-debug.md

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>