| """Cancellable pre-generation arbitration for the single shared TTS worker. |
| |
| Retry only an explicit 409/503 before SSE admission, never a partially accepted |
| generation/network failure. No other session is cancelled or evicted. |
| """ |
| import asyncio,time |
| from contextlib import asynccontextmanager |
|
|
| @asynccontextmanager |
| async def admitted_turn(client,body,on_wait,timeout=45.,poll=.2): |
| started=time.monotonic();last_notice=-100. |
| while time.monotonic()-started<timeout: |
| state=await client.get('/api/state',timeout=3.) |
| ready=state.status_code==200 and not state.json().get('busy') and state.json().get('ready',True) |
| if ready: |
| async with client.stream('POST','/api/turn',json=body) as response: |
| if response.status_code not in (409,503): |
| response.raise_for_status() |
| await on_wait('Voice model available. Preparing your spoken reply…','preparing') |
| yield response |
| return |
| now=time.monotonic() |
| if now-last_notice>=2.: |
| last_notice=now |
| await on_wait(f'Waiting for the shared voice model ({now-started:.0f} s). You can stop this reply.','waiting') |
| await asyncio.sleep(poll) |
| raise TimeoutError('Shared voice model did not become available within the bounded wait') |
|
|
