Update app.py

app.py

@@ -93,14 +93,14 @@ def ollama_func(command):
     else:
         return "No supported command."
         
-@spaces.GPU()      
 def launch():
     global OLLAMA_SERVICE_THREAD
     OLLAMA_SERVICE_THREAD = threading.Thread(target=ollama_service_thread)
     OLLAMA_SERVICE_THREAD.start()
     print("Giving ollama serve a moment")
     time.sleep(10)
-  
+
+@spaces.GPU()  
 async def stream_chat(message: str, history: list, model: str, temperature: float, max_new_tokens: int, top_p: float, top_k: int, penalty: float):
     print(f"message: {message}")
     conversation = []