Update app.py

app.py

@@ -57,7 +57,6 @@ def qwen_api(user_message, top_p=0.9,temperature=0.7, system_message='', max_tok
     ):
         token = message.choices[0].delta.content
         response += token
-        yield response
     return response
 
 os.environ["OPENAI_API_BASE"] = "https://api-inference.huggingface.co/v1/"
@@ -93,6 +92,8 @@ def chat(question, history):
         response = qa.invoke(question)['result']
     else:
         response = qwen_api(question, gradio_history=history)
+    print(response)
+    yield response
     return response
 
 
@@ -133,7 +134,7 @@ def respond(
 chatbot = gr.Chatbot(height=600)
 
 demo = gr.ChatInterface(
-    chat,
+    fn=chat,
     fill_height=True,
     chatbot=chatbot,
     additional_inputs=[