Update app.py

app.py

@@ -6,6 +6,7 @@ API_URL = "http://0.0.0.0:8000/v1/chat/completions"
 def stream_completion(message, history, max_tokens, temperature, top_p, beta):
     """Gradio callback: stream the assistant’s reply token-by-token."""
     # -------- build OpenAI-style message list (no system prompt) -------------
+    messages = []
     for user_msg, assistant_msg in history:
         if user_msg:        # past user turn
             messages.append({"role": "user", "content": user_msg})