Keep TTS and STT workers resident between requests

This commit is contained in:
Mikei386
2026-09-29 20:37:29 +02:00
parent 3dcf2d3f76
commit 90a97567f3
11 changed files with 197 additions and 80 deletions
+7 -1
View File
@@ -42,7 +42,7 @@ class Endpoint:
rows=self.rows();worker=self.worker.status();job=self.images.status()['job'];counts={}
for key,kind in [('llm','chat'),('image','image'),('tts','audio'),('stt','stt')]:
subset=[p for p in rows if p['kind']==kind]
counts[key]=dict(enabled=sum(p['enabled'] for p in subset),available=sum(p['enabled'] and p['runnable'] for p in subset),loaded=bool(worker['state']=='ready' and kind=='chat') if kind=='chat' else bool(self.tts and self.tts.status()['job'] and self.tts.status()['job']['state']=='running') if kind=='audio' else bool(self.stt and self.stt.status()['job'] and self.stt.status()['job']['state']=='running') if kind=='stt' else bool(kind=='image' and job and job['state']=='running'),supported=kind in ('chat','image') or (kind=='audio' and self.tts is not None) or (kind=='stt' and self.stt is not None))
counts[key]=dict(enabled=sum(p['enabled'] for p in subset),available=sum(p['enabled'] and p['runnable'] for p in subset),loaded=bool(worker['state']=='ready' and kind=='chat') if kind=='chat' else bool(self.tts and self.tts.status().get('loaded')) if kind=='audio' else bool(self.stt and self.stt.status().get('loaded')) if kind=='stt' else bool(kind=='image' and job and job['state']=='running'),supported=kind in ('chat','image') or (kind=='audio' and self.tts is not None) or (kind=='stt' and self.stt is not None))
scheduler_status=self.scheduler.status()
with self.lock:return dict(video=self.video.status() if self.video else None,state=self.state,reachable=bool(self.thread and self.thread.is_alive() and self.state=='running'),port=self.config['port'],bind=os.environ.get('DECK_API_BIND','127.0.0.1'),base_url=f"http://127.0.0.1:{self.config['port']}/v1",allowed_ports=self.allowed_ports,error=self.error,counts=counts,worker=worker,scheduler=scheduler_status,profiles=[dict(id=p['id'],name=p['name'],kind=p['kind'],enabled=p['enabled'],runnable=p['runnable'],blockers=p['blockers']) for p in rows],active_requests=self.inflight)
def configure(self,data):
@@ -70,6 +70,12 @@ class Endpoint:
enabled.add(data['id'])
else:enabled.discard(data['id'])
self.config['enabled_profiles']=sorted(enabled);self.persist()
if not data['enabled'] and row['kind']=='audio' and self.tts:
job=self.tts.status().get('job') or {}
if job.get('profile_id')==row['id'] and job.get('state')!='running':self.tts.stop()
if not data['enabled'] and row['kind']=='stt' and self.stt:
job=self.stt.status().get('job') or {}
if job.get('profile_id')==row['id'] and job.get('state')!='running':self.stt.stop()
return self.status()
def start(self):
with self.lock: