Update app.py

app.py

@@ -61,7 +61,7 @@ def predict(message, history):
     generate_kwargs = dict(
         model_inputs,
         streamer=streamer,
-        max_new_tokens=9999,
+        max_new_tokens=1024,
         do_sample=True,
         top_p=0.95,
         top_k=1000,