How to avoid PreTrainedTokenizerFast.decode to add space between tokens

Hmm… clean_up_tokenization_spaces?