Many Chinese models in hugging face like T5 use different tokenizer and model. For example, Chinese T5 uses T5ForConditionalGeneration and BertTokenizer. Is there some way to handle this kind of problems? Thank you!