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

endpoint.py

@@ -40,7 +40,7 @@ def chat_upstream_error(status, raw):
 
 class Endpoint:
     def __init__(self,root,profiles,worker,scheduler,images,credentials,management_port):
-        self.root=Path(root);self.profiles=profiles;self.worker=worker;self.scheduler=scheduler;self.images=images;self.tts=None;self.stt=None;self.music=None;self.video=None;self.credentials=credentials
+        self.root=Path(root);self.profiles=profiles;self.worker=worker;self.scheduler=scheduler;self.images=images;self.tts=None;self.voxcpm=None;self.embeddings=None;self.stt=None;self.music=None;self.video=None;self.credentials=credentials
         self.management_port=management_port;self.lock=threading.RLock();self.http=None;self.thread=None;self.state='stopped';self.error=None;self.inflight=0
         self.allowed_ports=[int(p) for p in os.environ.get('DECK_API_PORTS','').split(',') if p]
         self.config=dict(port=self.allowed_ports[0] if self.allowed_ports else 8120,enabled_profiles=[],autostart=False)
@@ -126,6 +126,7 @@ class Endpoint:
             with self.lock:pending=self.inflight
             if not pending and self.scheduler.unload_idle():break
             time.sleep(.1)
+        if self.embeddings:self.embeddings.stop()
         with self.lock:self.http=None;self.thread=None;self.state='stopped'
     def close(self):
         # Process shutdown does not change the user's autostart preference.
@@ -133,6 +134,8 @@ class Endpoint:
         if http:http.shutdown();http.server_close()
         if self.music:self.music.stop()
         if self.tts:self.tts.stop()
+        if self.embeddings:self.embeddings.stop()
+        if self.voxcpm:self.voxcpm.stop()
         if self.stt:self.stt.stop()
         self.images.stop();self.worker.stop()
     def allowed(self):
@@ -203,8 +206,8 @@ class APIHandler(BaseHTTPRequestHandler):
             with ep.lock:
                 if not ep.allowed():raise APIError('Endpunkt wird gestoppt.',503,'endpoint_stopping')
                 ep.inflight+=1;admitted=True
-            if ep.scheduler.gpu_mode=='separator':raise APIError('Audio-Trennung aktiv. Für Chat, Bilder oder Musik die Betriebsart wechseln.',503,'separator_mode_active')
-            if ep.video and ep.scheduler.gpu_mode not in ('llm','music'):
+            if ep.scheduler.gpu_mode=='separator' and self.path not in ('/v1/embeddings','/v1/embeddings/models'):raise APIError('Audio-Trennung aktiv. Für Chat, Bilder oder Musik die Betriebsart wechseln.',503,'separator_mode_active')
+            if ep.video and ep.scheduler.gpu_mode not in ('llm','music') and self.path not in ('/v1/embeddings','/v1/embeddings/models'):
                 if ep.scheduler.gpu_mode=='switching':raise APIError('Moduswechsel läuft.',503,'mode_switching')
                 video=ep.video.status()
                 if not video.get('service',{}).get('ready'):raise APIError('Video-API noch nicht bereit.',503,'video_not_ready')
@@ -214,14 +217,17 @@ class APIHandler(BaseHTTPRequestHandler):
             if self.command=='GET' and self.path=='/v1/capabilities':
                 from api_compat import capabilities
                 return self.send(capabilities())
-            model_routes={'/v1/models':'chat','/v1/images/models':'image','/v1/audio/speech/models':'audio','/v1/audio/transcriptions/models':'stt','/v1/audio/music/models':'music'}
-            if self.command=='GET' and self.path in model_routes:return self.send(ep.model_list(model_routes[self.path]))
+            model_routes={'/v1/embeddings/models':'embeddings','/v1/models':'chat','/v1/images/models':'image','/v1/audio/speech/models':'audio','/v1/audio/transcriptions/models':'stt','/v1/audio/music/models':'music'}
+            if self.command=='GET' and self.path in model_routes:
+                listing=ep.model_list(model_routes[self.path])
+                if self.path=='/v1/audio/speech/models':listing['data']+=ep.model_list('voice')['data']
+                return self.send(listing)
             if self.command=='GET' and self.path=='/health':return self.send({'status':'ok','service':'athena-deck-api'})
             if self.video_route(ep):return
             if self.command!='POST':raise APIError('Route nicht gefunden.',404)
-            if ep.scheduler.gpu_mode=='separator':raise APIError('Audio-Trennung aktiv. Für Chat, Bilder oder Musik die Betriebsart wechseln.',503,'separator_mode_active')
-            if ep.scheduler.gpu_mode=='music' and self.path!='/v1/audio/music':raise APIError('Musik-Modus aktiv. Für andere KI-Aufträge auf LLM wechseln.',503,'music_mode_active')
-            if ep.video and ep.scheduler.gpu_mode not in ('llm','music') and self.path!='/v1/audio/transcriptions':raise APIError('Video-Modus aktiv; Chat-, Bild- und TTS-Aufträge sind gesperrt. Auf LLM zurückschalten.',503,'video_mode_active')
+            if ep.scheduler.gpu_mode=='separator' and self.path not in ('/v1/embeddings','/v1/embeddings/models'):raise APIError('Audio-Trennung aktiv. Für Chat, Bilder oder Musik die Betriebsart wechseln.',503,'separator_mode_active')
+            if ep.scheduler.gpu_mode=='music' and self.path not in ('/v1/audio/music','/v1/embeddings'):raise APIError('Musik-Modus aktiv. Für andere KI-Aufträge auf LLM wechseln.',503,'music_mode_active')
+            if ep.video and ep.scheduler.gpu_mode not in ('llm','music') and self.path not in ('/v1/audio/transcriptions','/v1/embeddings'):raise APIError('Video-Modus aktiv; Chat-, Bild- und TTS-Aufträge sind gesperrt. Auf LLM zurückschalten.',503,'video_mode_active')
             if self.path=='/v1/audio/transcriptions':return self.transcription(ep)
             if self.path=='/v1/images/edits':
                 from image_upload import read_image_upload
@@ -232,7 +238,7 @@ class APIHandler(BaseHTTPRequestHandler):
                     fields['n']=int(fields.get('n','1'))
                     return self.image(ep,fields,images)
                 except ValueError as exc:raise APIError(str(exc)) from None
-            if self.path not in ('/v1/chat/completions','/v1/images/generations','/v1/audio/speech','/v1/audio/music'):raise APIError('Route nicht implementiert.',404)
+            if self.path not in ('/v1/embeddings','/v1/chat/completions','/v1/images/generations','/v1/audio/speech','/v1/audio/music'):raise APIError('Route nicht implementiert.',404)
             if self.headers.get('Transfer-Encoding'):raise APIError('Chunked Upload wird nicht unterstützt.')
             try:length=int(self.headers.get('Content-Length','0'))
             except ValueError:raise APIError('Ungültige Content-Length.') from None
@@ -240,6 +246,7 @@ class APIHandler(BaseHTTPRequestHandler):
             try:data=json.loads(self.rfile.read(length),parse_constant=lambda _:(_ for _ in ()).throw(ValueError()))
             except (ValueError,UnicodeError):raise APIError('Ungültiges JSON.') from None
             if not isinstance(data,dict):raise APIError('JSON-Objekt erforderlich.')
+            if self.path=='/v1/embeddings':return self.embedding(ep,data)
             if self.path=='/v1/chat/completions':return self.chat(ep,data)
             if self.path=='/v1/audio/music':return self.music_generation(ep,data)
             if self.path=='/v1/audio/speech':return self.speech(ep,data)
@@ -275,17 +282,35 @@ class APIHandler(BaseHTTPRequestHandler):
         if not ep.tts:raise APIError('TTS-Laufzeit nicht eingerichtet.',501)
         if set(data)-{'model','input','voice','response_format','speed','language'}:raise APIError('Nicht unterstützte TTS-Felder.')
         if data.get('response_format','wav')!='wav':raise APIError('Derzeit wird nur WAV unterstützt; response_format=wav setzen.')
-        profile=ep.find_profile(data.get('model'),'audio')
-        try:job=ep.tts.start(profile['id'],data.get('input'),speaker=data.get('voice','Ryan'),language=data.get('language','Auto'),speed=data.get('speed'),wait=True)
-        except ValueError as exc:raise APIError(str(exc)) from None
+        from profiles import voice_recipe
+        voice=next((p for p in ep.rows() if p.get('model') and voice_recipe(p['model']) and p['name']==data.get('model') and p['enabled']),None)
+        if voice:
+            profile=ep.find_profile(data.get('model'),voice['kind']);backend=ep.voxcpm
+            if data.get('voice','default')!='default' or data.get('speed',1)!=1:raise APIError('VoxCPM2: voice=default und speed=1 verwenden; Referenzstimme im Deck-Testbereich.')
+            try:job=backend.start(profile['id'],data.get('input'),wait=True)
+            except ValueError as exc:raise APIError(str(exc)) from None
+        else:
+            profile=ep.find_profile(data.get('model'),'audio');backend=ep.tts
+            try:job=backend.start(profile['id'],data.get('input'),speaker=data.get('voice','Ryan'),language=data.get('language','Auto'),speed=data.get('speed'),wait=True)
+            except ValueError as exc:raise APIError(str(exc)) from None
         end=time.monotonic()+620
         while time.monotonic()<end:
-            current=ep.tts.status()['job']
+            current=backend.status()['job']
             if current and current['id']==job['id'] and current['state'] in ('failed','cancelled'):raise APIError(current['phase'],503)
-            try:body=ep.tts.audio(job['id'])
+            try:body=backend.audio(job['id'])
             except (ValueError,OSError):time.sleep(.25);continue
             self.sent=True;self.send_response(200);self.send_header('Content-Type','audio/wav');self.send_header('Content-Length',str(len(body)));self.send_header('Cache-Control','no-store');self.send_header('Connection','close');self.end_headers();self.wfile.write(body);return
         raise APIError('TTS-Zeitlimit überschritten.',504)
+    def embedding(self,ep,data):
+        if not ep.embeddings:raise APIError('Embedding-Worker nicht eingerichtet.',501)
+        if set(data)-{'model','input','encoding_format','user'}:raise APIError('Nicht unterstützte Embedding-Felder.')
+        if data.get('encoding_format','float')!='float':raise APIError('Nur encoding_format=float unterstützt.')
+        value=data.get('input');texts=[value] if isinstance(value,str) else value
+        if not isinstance(texts,list) or not 1<=len(texts)<=32 or any(not isinstance(t,str) or not t.strip() or len(t)>16000 for t in texts):raise APIError('input: 1–32 nichtleere Texte, jeweils maximal 16000 Zeichen. Kontextgrenze wird zusätzlich von llama.cpp geprüft.')
+        profile=ep.find_profile(data.get('model'),'embeddings')
+        try:return self.send(ep.embeddings.generate(profile,{k:v for k,v in data.items() if k!='user'}))
+        except ValueError as exc:raise APIError(str(exc)) from None
+
     def music_generation(self,ep,data):
         if ep.scheduler.gpu_mode!='music':raise APIError('Bitte in der Übersicht auf Musik umschalten.',503,'music_mode_required')
         if not ep.music:raise APIError('Musikworker nicht eingerichtet.',501)