Update app.py

app.py

@@ -18,7 +18,14 @@ def chat(message, history):
         messages.append({"role": "assistant", "content": assistant})
     messages.append({"role": "user", "content": message})
 
-    response = model.create_chat_completion(messages=messages, max_tokens=512)
-    return response["choices"][0]["message"]["content"]
+    output = ""
+    for chunk in model.create_chat_completion(
+        messages=messages,
+        max_tokens=512,
+        stream=True
+    ):
+        delta = chunk["choices"][0]["delta"].get("content", "")
+        output += delta
+        yield output
 
 gr.ChatInterface(chat).launch(server_name="0.0.0.0", server_port=7860)