How to run the tokenizer independently from the model?