Update app.py

app.py

@@ -280,6 +280,7 @@ def generate_response():
 
     response_tokens = ''
     def generate_and_log_tokens(model, generator):
+        global response_tokens
         for token in generate_tokens(model, generator):
             if token == b'': # or (max_new_tokens is not None and i >= max_new_tokens):
                 log(user_request, response_tokens.decode("utf-8", errors="ignore"))