Update app.py

app.py

@@ -48,7 +48,11 @@ def respond(
     settings.temperature = temperature
     settings.top_p = top_p
 
-    return agent.get_chat_response(message, llm_sampling_settings=settings, returns_streaming_generator=True)
+    for new_token in agent.get_chat_response(message, llm_sampling_settings=settings, returns_streaming_generator=True):
+        partial_message += new_token
+        if '<|im_end|>' in partial_message:
+            break
+        yield partial_message
     # stop_tokens = ["</s>", "[INST]", "[INST] ", "<s>", "[/INST]", "[/INST] "]
     # chat_template = '<s>[INST] ' + system_message
     # # for human, assistant in history: