Update app.py

app.py

@@ -281,8 +281,8 @@ def generate_response():
     response_tokens = []
     def generate_and_log_tokens(model, generator):
         for token in generate_tokens(model, generator):
-            if token == model.token_eos(): # or (max_new_tokens is not None and i >= max_new_tokens):
-                log(user_request, model.detokenize(response_tokens).decode("utf-8", errors="ignore"))
+            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"))
                 break
             response_tokens.append(token)
             yield token