* Fixing OPT fast tokenizer option. * Remove dependency on `pt`. * Move it to GPT2 tokenization tests. * Added a few tests.
train_new_from_iterator