Update app.py

app.py

@@ -173,9 +173,10 @@ def load_model(model_path):
             processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
         
         context_len = getattr(model.config, "max_position_embeddings", 8048)
-        logger.info(f"Model {model_path} loaded successfully")
+        logger.info(f"Model {model_path} loaded successfully to device: {model.device}")
+        model = model.to("cuda")
         return True
-    
+
     except Exception as e:
         logger.error(f"Error loading model {model_path}: {str(e)}")
         return False