BERT models and tokenizers are already stored and managed in the bert_models module and should not be stored here. In addition, since there may be situations where the user would like to use cpu instead of mps for inference when using it as a library, the automatic switching process to mps was removed.