- Szycy
- 28 września 2026 02:41 UTC
- Autor
- Kamo
- Pochęt się
- 4dbf61c
The load probe's realtime sessions sent no language, so every turn paid Whisper's language detection (9.4 s alone at 3 threads on k3m1), a call shape no production caller uses: SP13's RelayRealtimeSTT sends session.update with **************** The probe now sends SP13's session.update byte for byte after session.created and waits for session.updated to confirm the language before any audio; a session still detecting it fails. WHISPER__CPU_THREADS and OMP_NUM_THREADS go from 3 to 8 (6 workers x 8 = 48, k3m1's thread count). Task 6's benchmark: four concurrent turns with the language given take 3.4-3.8 s at 8 threads, 4.8-5.2 s at 3. The CPU request stays 6 so the next rollout's surge pod still fits on k3m1. A send error the pacing thread records after the verdict is now reported in the detail instead of dropped (final review M2). SP12 final review I1.