Fix: Optimize build to prevent OOM - download model at runtime instead of build time

Dockerfile

@@ -18,8 +18,8 @@ RUN pip install --no-cache-dir -r requirements.txt
 # Copia el código de la aplicación
 COPY . .
 
-# Pre-descargar el modelo durante la construcción
-RUN python -c "import whisper; whisper.load_model('large-v3')" || echo "Model download skipped, will download at runtime"
+# NO descargar el modelo aquí - se descarga en runtime
+# Esto evita problemas de memoria durante el build
 
 EXPOSE 7860
 

app.py

@@ -21,12 +21,37 @@ async def startup_event():
     logger.info("Starting up... initializing model.")
     loop = asyncio.get_event_loop()
     try:
+        # First, ensure model is downloaded
+        logger.info("Downloading Whisper model if not already present...")
+        await loop.run_in_executor(None, _ensure_model_downloaded)
+        logger.info("Model ready.")
+        
+        # Initialize the ASR model
         await loop.run_in_executor(None, server_wrapper.init_model)
         logger.info("Model initialized successfully.")
     except Exception as e:
         logger.error(f"Error during model initialization: {e}")
 
 
+def _ensure_model_downloaded():
+    """Ensure the Whisper model is downloaded."""
+    import os
+    model_dir = os.path.expanduser('~/.cache/whisper')
+    model_path = os.path.join(model_dir, 'large-v3.pt')
+    
+    if not os.path.exists(model_path):
+        try:
+            logger.info(f"Downloading Whisper large-v3 model to {model_path}...")
+            import whisper
+            whisper.load_model('large-v3')
+            logger.info("Model downloaded successfully.")
+        except Exception as e:
+            logger.warning(f"Could not pre-download model: {e}")
+            logger.info("Will attempt to download on first use.")
+    else:
+        logger.info(f"Model already present at {model_path}")
+
+
 @app.post("/api/reset")
 async def api_reset():
     try:

requirements.txt

@@ -5,9 +5,7 @@ soundfile>=0.12.0
 librosa>=0.10.0
 numpy>=1.24.0
 torch>=2.0.0
-transformers>=4.30.0
 torchaudio>=2.0.0
 tqdm
 tiktoken
-triton>=2.0.0,<3;platform_machine=="x86_64" and sys_platform=="linux" or sys_platform=="linux2"
 openai-whisper>=20230314