Add configurable TTS and STT residency policies

stt.py

@@ -50,7 +50,7 @@ def read_upload(handler):
 
 class STT:
  def __init__(self,profiles,catalog,runtime):
-  self.profiles=profiles;self.catalog=catalog;self.runtime=runtime;self.lock=threading.RLock();self.job=None;self.process=None;self.port=None;self.key=None;self.binary_path=None;self.cancel=threading.Event()
+  self.profiles=profiles;self.catalog=catalog;self.runtime=runtime;self.lock=threading.RLock();self.job=None;self.process=None;self.port=None;self.key=None;self.binary_path=None;self.cancel=threading.Event();self.policy='auto';self.warming=False
  def build(self):
   state=self.runtime.status();ident=state.get('active')
   if not isinstance(ident,str) or len(ident)!=32 or any(c not in '0123456789abcdef' for c in ident):raise ValueError('Zuerst unter Laufzeiten → llama.cpp einen Build erstellen und aktivieren.')
@@ -75,7 +75,7 @@ class STT:
   except ValueError:projector=False
   try:self.model_entry();model=True
   except ValueError:model=False
-  with self.lock:return dict(model=model,job=dict(self.job) if self.job else None,installed=installed,projector=projector,repo=REPO,revision=REVISION,loaded=bool(self.process and self.process.poll() is None))
+  with self.lock:return dict(model=model,job=dict(self.job) if self.job else None,installed=installed,projector=projector,repo=REPO,revision=REVISION,loaded=bool(self.process and self.process.poll() is None),policy=self.policy,warming=self.warming)
  def model_entry(self):
   for x in self.catalog.status()['entries']:
    if supported(x):return self.catalog.entry(x['id'])
@@ -99,6 +99,7 @@ class STT:
   validate_wav(audio)
   if language not in ('de','en','auto'):raise ValueError('Sprache muss de, en oder auto sein.')
   with self.lock:
+   if self.warming:raise ValueError('STT wird gerade vorgeladen. Gleich erneut versuchen.')
    if self.job and self.job['state']=='running':raise ValueError('Ein STT-Auftrag läuft bereits.')
    p=next((p for p in self.profiles.status()['profiles'] if p['id']==profile_id and p['kind']=='stt'),None)
    if not p or not p['runnable']:raise ValueError('STT-Profil nicht ausführbar. Zuerst Einrichten öffnen.')
@@ -110,6 +111,29 @@ class STT:
    threading.Thread(target=self._run,args=(binary,model,projector,audio,language),daemon=True).start();return dict(self.job)
  def _run(self,binary,model,projector,audio,language):
   process=None;result=dict(state='failed',phase='Spracherkennung fehlgeschlagen. Laufzeit und Speicher prüfen.')
+  try:
+   port,key=self._ensure(binary,model,projector)
+   base=f'http://127.0.0.1:{port}';headers={'Authorization':'Bearer '+key};deadline=time.monotonic()+90
+   with self.lock:self.job['phase']='Audio wird auf der CPU transkribiert'
+   boundary='deck-'+uuid.uuid4().hex
+   body=f'--{boundary}\r\nContent-Disposition: form-data; name="file"; filename="audio.wav"\r\nContent-Type: audio/wav\r\n\r\n'.encode()+audio+b'\r\n'
+   for name,value in [('model','deck-stt')]+([] if language=='auto' else [('language',language)]):body+=f'--{boundary}\r\nContent-Disposition: form-data; name="{name}"\r\n\r\n{value}\r\n'.encode()
+   body+=f'--{boundary}--\r\n'.encode();headers['Content-Type']='multipart/form-data; boundary='+boundary
+   with urllib.request.urlopen(urllib.request.Request(base+'/v1/audio/transcriptions',data=body,headers=headers),timeout=180) as response:
+    payload=json.loads(response.read(1024*1024))
+   if not isinstance(payload.get('text'),str):raise ValueError('Die Laufzeit hat kein Transkript geliefert.')
+   text=payload['text'].split('<asr_text>')[-1].replace('<|endoftext|>','').strip()
+   result=dict(state='complete',phase='Transkription fertig · '+('Modell entladen' if self.policy=='per_request' else 'Modell bleibt geladen'),text=text)
+  except InterruptedError:result=dict(state='cancelled',phase='Transkription abgebrochen')
+  except ValueError as exc:result=dict(state='failed',phase=str(exc))
+  except Exception:pass
+  finally:
+   if result['state']!='complete' or self.policy=='per_request':self.unload_idle()
+   with self.lock:
+    if self.cancel.is_set():result=dict(state='cancelled',phase='Transkription abgebrochen')
+    self.job.update(result)
+ def _ensure(self,binary,model,projector):
+  process=None
   try:
    with self.lock:
     if self.cancel.is_set():raise InterruptedError()
@@ -134,25 +158,23 @@ class STT:
      except OSError:continue
    else:
     with self.lock:port=self.port;key=self.key
-   base=f'http://127.0.0.1:{port}';headers={'Authorization':'Bearer '+key};deadline=time.monotonic()+90
-   with self.lock:self.job['phase']='Audio wird auf der CPU transkribiert'
-   boundary='deck-'+uuid.uuid4().hex
-   body=f'--{boundary}\r\nContent-Disposition: form-data; name="file"; filename="audio.wav"\r\nContent-Type: audio/wav\r\n\r\n'.encode()+audio+b'\r\n'
-   for name,value in [('model','deck-stt')]+([] if language=='auto' else [('language',language)]):body+=f'--{boundary}\r\nContent-Disposition: form-data; name="{name}"\r\n\r\n{value}\r\n'.encode()
-   body+=f'--{boundary}--\r\n'.encode();headers['Content-Type']='multipart/form-data; boundary='+boundary
-   with urllib.request.urlopen(urllib.request.Request(base+'/v1/audio/transcriptions',data=body,headers=headers),timeout=180) as response:
-    payload=json.loads(response.read(1024*1024))
-   if not isinstance(payload.get('text'),str):raise ValueError('Die Laufzeit hat kein Transkript geliefert.')
-   text=payload['text'].split('<asr_text>')[-1].replace('<|endoftext|>','').strip()
-   result=dict(state='complete',phase='Transkription fertig · Modell bleibt geladen',text=text)
-  except InterruptedError:result=dict(state='cancelled',phase='Transkription abgebrochen')
-  except ValueError as exc:result=dict(state='failed',phase=str(exc))
-  except Exception:pass
+   return port,key
+  except Exception:
+   self.unload_idle();raise
+ def warm(self,profile_id):
+  with self.lock:
+   if self.job and self.job.get('state')=='running':return False
+   if self.process and self.process.poll() is None:return True
+   if self.warming:return False
+   p=next((p for p in self.profiles.status()['profiles'] if p['id']==profile_id and p['kind']=='stt' and p['runnable']),None)
+   if not p:raise ValueError('Das ausgewählte STT-Profil ist nicht ausführbar.')
+   self.warming=True;self.cancel.clear()
+  try:
+   headroom=cgroup_headroom()
+   if headroom is not None and headroom<4*1024**3:raise ValueError('Mindestens 4 GiB freier Deck-RAM werden benötigt.')
+   self._ensure(self.build(),self.catalog.root/p['model_id']/'model.gguf',self.catalog.root/self.projector()['id']/'model.gguf');return True
   finally:
-   if result['state']!='complete':self.unload_idle()
-   with self.lock:
-    if self.cancel.is_set():result=dict(state='cancelled',phase='Transkription abgebrochen')
-    self.job.update(result)
+   with self.lock:self.warming=False
  def unload_idle(self):
   with self.lock:
    process=self.process;self.process=None;self.port=None;self.key=None;self.binary_path=None