Nicolas Patry
Nicolas Patry
Yes I'm not sure we should do something about it.
My comment from last year still stands. `from torch._subclasses import fake_tensor` is private, therefore I don't think we should do anything about it until it's officially supported. That being said...
I tried with every latest version of the packages, and the reproducer runs for me. Happy to help whenever there's a reproducer !
I've seen that happen when using network mounted disks. If the network is flaky then the read might fail even though the rest went fine. Error should be transient though....
Hi, not sure if I can enlighten. In general, I'm not sure why `forced_decoder_ids` is useful for, since if you know what ids you should get, there's no need to...
Yes, I know. I don't *think* it's necessary but I just usually give the benefit of the doubt when something was coded intentionally.
It turns out the file in https://github.com/google/sentencepiece is actually not even valid protobuf 4.x ...
Nope. it passes with protobuf 3.20 in the tests, but not with 4.x..... I think we should wait for them to upgrade before doing it ourselves (the .proto files are...
Can we refactor in order to do : ```python output = fill_mask_pipeline("Text to predict ", tokenizer_kwargs=tokenizer_kwargs) ``` Instead ? Accepting kwargs directly is very hard to maintain down the line...