same here, try to not serialize too much if unneeded

src/transformers/tokenization_utils.py

@@ -503,7 +503,8 @@ class PreTrainedTokenizer(object):
         tokenizer_config_file = os.path.join(save_directory, TOKENIZER_CONFIG_FILE)
 
         tokenizer_config = copy.deepcopy(self.init_kwargs)
-        tokenizer_config["init_inputs"] = copy.deepcopy(self.init_inputs)
+        if len(self.init_inputs) > 0:
+            tokenizer_config["init_inputs"] = copy.deepcopy(self.init_inputs)
         for file_id in self.vocab_files_names.keys():
             tokenizer_config.pop(file_id, None)