Keep TTS and STT workers resident between requests

inference.py

@@ -19,9 +19,9 @@ class InferenceError(ValueError):
 class Scheduler:
     """FIFO admission; same-profile requests share configured slots, switches drain."""
     def __init__(self, worker):
-        self.worker=worker;self.cv=threading.Condition();self.queue=[];self.key=None;self.active=0;self.transition=False;self.gpu_mode="llm"
+        self.worker=worker;self.evict_tts=lambda:None;self.cv=threading.Condition();self.queue=[];self.key=None;self.active=0;self.tts_active=0;self.transition=False;self.gpu_mode="llm"
     def status(self):
-        with self.cv:return dict(active_requests=self.active,waiting_requests=len(self.queue),switching=self.transition)
+        with self.cv:return dict(active_requests=self.active+self.tts_active,waiting_requests=len(self.queue),switching=self.transition)
     @contextlib.contextmanager
     def lease(self, key, slots=1, prepare=None, timeout=600, allowed=lambda:True):
         ticket=object();deadline=time.monotonic()+timeout;claimed=False
@@ -32,7 +32,7 @@ class Scheduler:
                 while True:
                     if self.gpu_mode!="llm":raise InferenceError("Video-Modus aktiv oder Moduswechsel läuft; GPU-Aufträge sind gesperrt.")
                     if not allowed():raise InferenceError('Endpunkt wird gestoppt oder Profil ist nicht mehr aktiviert.')
-                    if self.queue[0] is ticket and not self.transition and (not self.active or (self.key==key and self.active<slots)):
+                    if self.queue[0] is ticket and not self.transition and (not self.active or (self.key==key and self.active<slots)) and (not self.tts_active or self.key==key and key!=('image',)):
                         self.queue.pop(0);self.active+=1;claimed=True
                         switch=self.key!=key;self.key=key
                         # Also verifies/restarts a crashed worker when there are no other leases.
@@ -47,6 +47,7 @@ class Scheduler:
             if self.transition:
                 try:
                     if switch:self.worker.stop()
+                    if key==('image',) or switch:self.evict_tts()
                     if prepare:prepare()
                 except Exception:
                     self.worker.stop()
@@ -58,15 +59,27 @@ class Scheduler:
         finally:
             if claimed:
                 with self.cv:self.active-=1;self.cv.notify_all()
-    def image_reservation(self, wait=False):
-        lease=self.lease(('image',),timeout=600 if wait else 0)
+    def image_reservation(self, wait=False,kind='image'):
+        lease=self.lease((kind,),timeout=600 if wait else 0)
         lease.__enter__()
         return lambda:lease.__exit__(None,None,None)
+    def tts_reservation(self,wait=False):
+        deadline=time.monotonic()+(600 if wait else 0)
+        with self.cv:
+            while True:
+                if self.gpu_mode!='llm':raise InferenceError('Video-Modus aktiv; TTS ist nicht verfügbar.')
+                if not self.transition and not (self.key==('image',) and self.active) and not self.queue:
+                    self.tts_active+=1;break
+                if time.monotonic()>=deadline:raise InferenceError('GPU-Auftrag läuft; TTS später erneut versuchen.')
+                self.cv.wait(min(1,max(.01,deadline-time.monotonic())))
+        def release():
+            with self.cv:self.tts_active-=1;self.cv.notify_all()
+        return release
     def unload_idle(self):
         with self.cv:
-            if self.active:return False
+            if self.active or self.tts_active:return False
             self.transition=True
-            try:self.worker.stop();self.key=None
+            try:self.worker.stop();self.evict_tts();self.key=None
             finally:self.transition=False;self.cv.notify_all()
             return True