Add configurable TTS and STT residency policies

audio_policy.py

@@ -0,0 +1,63 @@
+"""Persisted residency preferences for Deck-owned speech workers."""
+import json
+import threading
+import time
+from pathlib import Path
+
+MODES=('auto','warm','per_request')
+
+class AudioPolicy:
+ def __init__(self,path,profiles,tts,stt,scheduler,endpoint):
+  self.path=Path(path);self.profiles=profiles;self.tts=tts;self.stt=stt;self.scheduler=scheduler;self.endpoint=endpoint
+  self.lock=threading.RLock();self.settings={kind:{'mode':'auto','profile_id':None} for kind in ('tts','stt')};self.errors={};self.retry_after={}
+  if self.path.exists():
+   try:
+    stored=json.loads(self.path.read_text())
+    for kind in self.settings:
+     value=stored.get(kind,{})
+     if value.get('mode') in MODES and (value.get('profile_id') is None or isinstance(value.get('profile_id'),str)):
+      self.settings[kind]=dict(mode=value['mode'],profile_id=value.get('profile_id'))
+   except (OSError,ValueError,AttributeError,TypeError):pass
+  self.tts.policy=self.settings['tts']['mode'];self.stt.policy=self.settings['stt']['mode']
+  self.thread=threading.Thread(target=self._loop,daemon=True);self.thread.start()
+ def status(self):
+  with self.lock:return dict(settings={k:dict(v) for k,v in self.settings.items()},errors=dict(self.errors))
+ def configure(self,data):
+  if not isinstance(data,dict) or set(data)!={'kind','mode','profile_id'}:raise ValueError('Art, Modus und Profil-ID erforderlich.')
+  kind=data['kind'];mode=data['mode'];ident=data['profile_id']
+  if kind not in self.settings or mode not in MODES:raise ValueError('Ungültiger Audio-Modus.')
+  if ident is not None and (not isinstance(ident,str) or len(ident)>128):raise ValueError('Ungültige Profil-ID.')
+  if mode=='warm':
+   profile_kind='audio' if kind=='tts' else 'stt'
+   if not ident or not any(p['id']==ident and p['kind']==profile_kind and p['runnable'] for p in self.profiles.status()['profiles']):raise ValueError('Für „Immer bereit“ ein ausführbares Profil auswählen.')
+  with self.lock:
+   updated={k:dict(v) for k,v in self.settings.items()};updated[kind]=dict(mode=mode,profile_id=ident)
+   self.path.parent.mkdir(parents=True,exist_ok=True);temp=self.path.with_suffix('.tmp');temp.write_text(json.dumps(updated));temp.replace(self.path)
+   self.settings=updated;self.errors.pop(kind,None);self.retry_after.pop(kind,None)
+   worker=self.tts if kind=='tts' else self.stt;worker.policy=mode
+  if mode=='per_request':
+   with worker.lock:active=bool(worker.job and worker.job.get('state')=='running')
+   if not active:worker.unload_idle()
+  return self.status()
+ def _loop(self):
+  while True:
+   time.sleep(8)
+   with self.lock:settings={k:dict(v) for k,v in self.settings.items()}
+   if self.endpoint.state!='running':continue
+   for kind,worker in (('stt',self.stt),('tts',self.tts)):
+    item=settings[kind]
+    if item['mode']!='warm':continue
+    with self.lock:
+     if time.monotonic()<self.retry_after.get(kind,0):continue
+    if kind=='tts':
+     with self.scheduler.cv:
+      if self.scheduler.gpu_mode!='llm' or self.scheduler.active or self.scheduler.transition:continue
+    with worker.lock:
+     if worker.warming or worker.process and worker.process.poll() is None or worker.job and worker.job.get('state')=='running':continue
+    try:
+     worker.warm(item['profile_id'])
+     with self.lock:self.errors.pop(kind,None)
+    except Exception as exc:
+     with self.lock:
+      self.errors[kind]=str(exc) if isinstance(exc,ValueError) else 'Vorladen fehlgeschlagen. Ressourcen und Laufzeit prüfen.'
+      self.retry_after[kind]=time.monotonic()+60