Update app.py

app.py

@@ -33,11 +33,12 @@ def generate_reply(model, tokenizer, prompt, max_length, do_sample):
     )
     return tokenizer.decode(output_tokens[0], skip_special_tokens=True)
 
-@spaces.GPU
+
 def main():
     args, _ = get_args()
     model, tokenizer = load_model(args.model)
-
+    
+    @spaces.GPU
     def respond(user_message, chat_history):
         """
         Gradio expects a function that takes the last user message and the