Update app.py

app.py

@@ -67,18 +67,19 @@ def respond(
     print(messages)
 
     response = ""
-    yield response
+    # yield response
     chunks = []
-    yield response
+    # yield response
 
     for chunk in pipeline.chat(messages, **generation_kwargs):
         response += chunk.content
         chunks.append(chunk)
-        yield response
-    yield response
+        # yield response
+    # yield response
     messages.append(chatglm_cpp.ChatMessage(role="assistant", content=response))
 
     print(messages)
+    return response
 
 
 """