FIX [Generation] Fix some issues when running the MaxLength criteria on CPU

[younesbelkada]

What does this PR do?

Hey @zucchini-nlp @gante

In TRL we do have a class that call generate with MaxLengthCriteria and running the CI on main seems to cause some issues after #29116 😱
Not sure how to repro apart from running the failing CI and I did not flagged anything alarming outside my CI but to be on the safe zone I propose to init the torch.full tensor in bool directly, I assume it's fine according to the type hint

Here is the traceback:

self = [<transformers.generation.stopping_criteria.MaxLengthCriteria object at 0x7f652ccf8810>]
input_ids = tensor([[31373,   995, 19277],
        [31373,   995, 27455],
        [31373,   995,  4562],
        [31373,   995, 26964]])
scores = None, kwargs = {}, is_done = tensor([False, False, False, False])
criteria = <transformers.generation.stopping_criteria.MaxLengthCriteria object at 0x7f652ccf8810>

    @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> torch.BoolTensor:
        is_done = torch.full((input_ids.shape[0],), False, device=input_ids.device)
        for criteria in self:
>           is_done = is_done | criteria(input_ids, scores, **kwargs)
E           RuntimeError: "bitwise_or_cpu" not implemented for 'Float'

https://github.com/huggingface/trl/actions/runs/8058401333/job/22011132917

[younesbelkada]

cc @zucchini-nlp @gante what do you think about these changes ? 🙏 I can also spend some time on trying to repro with a minial snippet but if these changes look good to you perhaps we can move forward with it

PS: and for some reason I can't request review @zucchini-nlp 🤯

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[zucchini-nlp]

Okay for me, but it's weird that max_length criteria was returning a float tensor

[younesbelkada]

@zucchini-nlp torch.full(xxx) returns a float tensor by default unless you don't explicitly cast it or init the tensor with the desired dtype. The thing I am not sure about is that why it didn't get flagged on other tests

[gante]

Perfect, thank you for having a look at the issue 🤗