Six Whisper workers and six requested CPUs, so a batch chunk never queues a realtime turn

Порезанный

25 сентября 2026 г. в 17:36 UTC

Автор

Kamo

Обещать

b55020f

The acceptance run on k3m1 (4 realtime streams + 1 batch) measured p95 17.9 s: the batch held one of the 4 workers, so one turn waited behind a whole turn. Six workers run all five at once (Task 6 Step 7 fallback).