Unload model after 60 seconds inactivity. Attempt to clear more cuda vram.

subgen.py

@@ -1,4 +1,4 @@
-subgen_version = '2025.02.54'
+subgen_version = '2025.02.55'
 
 from language_code import LanguageCode
 from datetime import datetime
@@ -682,18 +682,96 @@ def extract_audio_segment_to_memory(input_file, start_time, duration):
         logging.error(f"Error: {str(e)}")
         return None
 
-def start_model():
-    global model
-    if model is None:
-        logging.debug("Model was purged, need to re-create")
-        model = stable_whisper.load_faster_whisper(whisper_model, download_root=model_location, device=transcribe_device, cpu_threads=whisper_threads, num_workers=concurrent_transcriptions, compute_type=compute_type)
+# --- Global Model Variables ---
+_model_loading = False
+_model_lock = threading.Lock()  # Protects access to `model` and `_model_loading`
+_unload_timer = None
+_unload_timer_lock = threading.Lock()
+_unload_delay_seconds = 60  # Adjust as needed
+
+# --- Model Loading/Unloading Functions ---
+
+async def _load_model():
+    """Asynchronously load the Whisper model."""
+    global model, _model_loading
+    logging.debug("Starting asynchronous model load...")
+    try:
+        with _model_lock:
+            _model_loading = True
+
+        model = load_faster_whisper(
+            whisper_model,
+            download_root=model_location,
+            device=transcribe_device,
+            cpu_threads=whisper_threads,
+            num_workers=concurrent_transcriptions,
+            compute_type=compute_type
+        )
+        logging.debug("Model loaded asynchronously.")
+
+    except Exception as e:
+        logging.error(f"Error loading model asynchronously: {e}")
+        # Handle the error (e.g., set model to None, log it, etc.)
+        with _model_lock:
+            model = None  # Ensure model is None on load failure
+    finally:
+        with _model_lock:
+            _model_loading = False
+
+def _unload_model_callback():
+    """Callback function to unload the model after a delay."""
+    with _unload_timer_lock:
+        delete_model()  # Call delete_model inside the lock
 
 def delete_model():
-    global model
-    if clear_vram_on_complete and task_queue.is_idle():
-        logging.debug("Queue idle; clearing model from memory.")
-        model = None
-    gc.collect()
+    """Unload the Whisper model, clear memory, and cancel any pending timers."""
+    global model, _unload_timer
+
+    with _model_lock:
+        if model is not None:
+            logging.debug("Clearing model from memory (delayed).")
+            model = None
+            # Explicitly release CUDA memory if using CUDA *and* the device is CUDA
+            try:
+                if transcribe_device.lower() == 'cuda' and torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+                    logging.debug("CUDA cache cleared.")
+                gc.collect()
+            except Exception as e:
+                logging.error(f"Error clearing CUDA cache: {e}")
+
+    # Cancel the unload timer
+    with _unload_timer_lock:
+        if _unload_timer:
+            try:  # Add a try-except block in case the timer has already been cancelled
+                _unload_timer.cancel()
+            except ValueError:
+                pass  # Timer was already cancelled
+            _unload_timer = None
+
+def start_model():
+    """Start the model loading process (asynchronously) or reset the unload timer."""
+    global model, _unload_timer
+
+    with _model_lock:
+        if model is None and not _model_loading:
+            logging.debug("Starting model loading")
+            asyncio.create_task(_load_model())  # Start loading in background
+        elif _model_loading:
+            logging.debug("Model is currently loading...")
+        else:
+            logging.debug("Model is already loaded.")
+
+    # Reset the timer if the model is loaded or loading
+    with _unload_timer_lock:
+        if _unload_timer:
+            try:
+                _unload_timer.cancel()
+            except ValueError:
+                pass
+        _unload_timer = threading.Timer(_unload_delay_seconds, _unload_model_callback)
+        _unload_timer.daemon = True
+        _unload_timer.start()
 
 def isAudioFileExtension(file_extension):
     return file_extension.casefold() in \