modify output length

app.py

@@ -26,7 +26,9 @@ def chat(user_input, max_new_tokens=20480):
         temperature=0.7,
         top_p=1,
     )
-    yield response.choices[0].message.content
+    resp_text = response.choices[0].message.content
+    print(resp_text)
+    yield resp_text