Update app.py

app.py

@@ -40,12 +40,12 @@ def respond(
 
     try:
         stream = client.chat_completion(
-            messages,
-            max_new_tokens=max_tokens,
-            temperature=temperature,
-            top_p=top_p,
-            stream=True,
-        )
+    messages,
+    max_tokens=max_tokens,  # 'max_new_tokens'에서 'max_tokens'로 변경
+    temperature=temperature,
+    top_p=top_p,
+    stream=True,
+)
         partial_message = ""
         for response in stream:
             if response.token.special: