Bert: Slow vs fast decoding inconsistency

[pcuenca]

Fast:

>>> from transformers import AutoTokenizer
>>> t = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased")
>>> t.decode(t.encode("l'eure"), skip_special_tokens=True)
"l ' eure"

Slow:

>>> t = AutoTokenizer.from_pretrained("google-bert/bert-base-uncased", use_fast=False)
>>> t.decode(t.encode("l'eure"), skip_special_tokens=True)
"l'eure"

I believe the "slow" interpretation to be the correct one. I tried to check the reference implementation but saw no decoding code.

Maybe related (thanks @xenova for the pointer): huggingface/transformers#31938 (defaults Bert to clean_up_tokenization_spaces=True)