Add profile presets and min-p controls, embeddings and VoxCPM workers

voxcpm_test.py

@@ -0,0 +1,91 @@
+"""VoxCPM2 speech and reference-voice generation, serialized by Deck GPU scheduler."""
+import io,json,os,shutil,signal,subprocess,threading,time,uuid,wave
+from pathlib import Path
+from tts_test import TTSTests
+from image_test import probe,cgroup_headroom
+from profiles import voice_recipe
+WORKER=Path(__file__).with_name('voxcpm_worker.py')
+
+def validate_reference(audio):
+ if not isinstance(audio,bytes) or not 0<len(audio)<=8*1024**2:raise ValueError('Referenz als WAV bis 8 MiB hochladen.')
+ try:
+  with wave.open(io.BytesIO(audio)) as w:
+   if w.getnchannels() not in (1,2) or w.getsampwidth()!=2 or not 8000<=w.getframerate()<=96000 or not 1<=w.getnframes()/w.getframerate()<=30:raise ValueError('Referenz: PCM-16-WAV, mono/stereo, 1–30 Sekunden erforderlich.')
+ except (wave.Error,EOFError):raise ValueError('Ungültige WAV-Datei. PCM-16-WAV verwenden.') from None
+
+class VoxCPMTests(TTSTests):
+ def blockers(self,p):
+  if not voice_recipe(p['model']):return ['Diese Stimm-Modellfamilie wird noch nicht unterstützt.']
+  errors=[]
+  if not self.runtime.status()['installed']:errors.append('VoxCPM-Laufzeit unter Laufzeiten → VoxCPM installieren.')
+  for role,info in voice_recipe(p['model']).items():
+   try:self.profiles._component(p['model'],role,p.get('components',{}).get(role))
+   except ValueError:errors.append(info['label']+' fehlt oder ist noch nicht zugeordnet. Unter Einrichtung & Komponenten ergänzen.')
+  return errors
+ def stage(self,p,root):
+  model=root/'model';model.mkdir();entries=[p['model']]+[self.profiles._component(p['model'],r,p.get('components',{}).get(r)) for r in voice_recipe(p['model'])]
+  for entry in entries:
+   src=self.profiles.catalog.root/entry['id']/('model'+Path(entry['file']).suffix)
+   (model/Path(entry['file']).name).symlink_to(src.resolve())
+ def start(self,profile_id,text,device='5080',cfg=2.0,steps=10,seed=42,reference=None,wait=False):
+  if not isinstance(text,str) or not 1<=len(text.strip())<=1000:raise ValueError('Text mit 1–1000 Zeichen erforderlich.')
+  if device not in ('5080','3060'):raise ValueError('RTX 5080 oder RTX 3060 wählen.')
+  if type(cfg) not in (int,float) or not .5<=cfg<=5 or type(steps)!=int or not 1<=steps<=50 or type(seed)!=int or not 0<=seed<=2147483647:raise ValueError('Ungültige Guidance, Schritte oder Seed.')
+  if reference is not None:validate_reference(reference)
+  p=next((p for p in self.profiles.status()['profiles'] if p['id']==profile_id),None)
+  if not p or not p['model'] or not voice_recipe(p['model']):raise ValueError('VoxCPM2-Profil nicht gefunden.')
+  errors=self.blockers(p)
+  if errors:raise ValueError(' '.join(errors))
+  # Claim local job before scheduler admission to prevent races between two callers.
+  with self.lock:
+   if self.job and self.job['state']=='running':raise ValueError('Stimmgenerierung läuft bereits.')
+   self.cancel.clear();ident=uuid.uuid4().hex;self.job=dict(id=ident,state='running',phase='GPU wird reserviert',profile_id=profile_id,started_at=time.time())
+  release=None
+  try:
+   release=self.acquire(wait)
+   gpu=next((g for g in probe() if 'RTX '+device in g['name'] and not g['processes'] and g['free_mib']>=6500),None)
+   if not gpu:raise ValueError('RTX '+device+' benötigt mindestens 6,5 GiB freien VRAM und darf nicht fremd belegt sein.')
+   headroom=cgroup_headroom()
+   if headroom is not None and headroom<8*1024**3:raise ValueError('Mindestens 8 GiB freier Deck-Systemspeicher erforderlich.')
+   self.root.mkdir(parents=True,exist_ok=True,mode=0o700)
+   self.job.update(phase='VoxCPM2 wird geladen',gpu=gpu['name']);self.gpu_uuid=gpu['uuid']
+   threading.Thread(target=self._run_voice,args=(p,ident,dict(text=text,cfg=cfg,steps=steps,seed=seed),reference,gpu,release),daemon=True).start();return dict(self.job)
+  except Exception as exc:
+   if release:release()
+   with self.lock:self.job.update(state='failed',phase=str(exc),finished_at=time.time())
+   raise
+ def _run_voice(self,p,ident,request,reference,gpu,release):
+  root=self.root/ident;state='failed';phase='VoxCPM2-Ausführung fehlgeschlagen'
+  try:
+   root.mkdir(mode=0o700);self.stage(p,root)
+   if reference is not None:(root/'reference.wav').write_bytes(reference)
+   env={k:v for k,v in os.environ.items() if not k.startswith(('HF_','LLAMA_','DECK_'))};env.update(CUDA_VISIBLE_DEVICES=gpu['uuid'],HF_HUB_OFFLINE='1',TRANSFORMERS_OFFLINE='1',OMP_NUM_THREADS='4',HOME=str(root))
+   with self.lock:
+    if self.cancel.is_set():raise InterruptedError()
+    self.process=subprocess.Popen([str(self.runtime.paths()[0]),str(WORKER),str(root)],stdin=subprocess.PIPE,stdout=subprocess.DEVNULL,stderr=subprocess.DEVNULL,env=env,start_new_session=True,text=True)
+    process=self.process
+   process.stdin.write(json.dumps(request));process.stdin.close();process.stdin=None;deadline=time.monotonic()+600
+   while process.poll() is None:
+    if self.cancel.wait(.25):raise InterruptedError()
+    if time.monotonic()>deadline:raise ValueError('Zeitlimit für VoxCPM2 erreicht.')
+    if (root/'phase').exists():
+     with self.lock:self.job['phase']=(root/'phase').read_text()
+   if self.cancel.is_set():raise InterruptedError()
+   if not (root/'result.json').exists():raise ValueError('VoxCPM2-Worker beendet ohne Ergebnis. Speicher oder Laufzeit prüfen.')
+   result=json.loads((root/'result.json').read_text())
+   if not result['ok']:raise ValueError(result['error'])
+   output=root/'result.wav'
+   with wave.open(str(output)) as w:
+    if w.getnframes()==0 or output.stat().st_size>32*1024**2:raise ValueError('Ungültige Audioausgabe.')
+   state='complete';phase='Stimme fertig · Modell entladen'
+   with self.lock:self.job.update(duration=result['duration'],sample_rate=result['sample_rate'])
+  except InterruptedError:state='cancelled';phase='Stimmgenerierung abgebrochen · Modell entladen'
+  except Exception as exc:
+   if self.cancel.is_set():state='cancelled';phase='Stimmgenerierung abgebrochen · Modell entladen'
+   else:phase=str(exc) if isinstance(exc,ValueError) else 'VoxCPM2-Ausführung fehlgeschlagen ('+type(exc).__name__+').'
+  finally:
+   self.unload_idle();shutil.rmtree(root/'model',ignore_errors=True)
+   for name in ('reference.wav','phase'):(root/name).unlink(missing_ok=True)
+   if state=='complete':(root/'complete').touch()
+   with self.lock:self.job.update(state=state,phase=phase,finished_at=time.time())
+   release()