[BART] test_generate_fp16 fails after PR#3140

[sshleifer]

passes after git checkout d6de6423 (commit preceding #3140)
Traceback:

                 unfinished_sents.mul_((~eos_in_sents).long())

            # stop when there is a </s> in each sentence, or if we exceed the maximul length
>           if unfinished_sents.max() == 0:
E           RuntimeError: cuda runtime error (716) : misaligned address at /pytorch/aten/src/THC/THCReduceAll.cuh:327

src/transformers/modeling_utils.py:992: RuntimeError

@thomwolf @patrickvonplaten

[patrickvonplaten]

Interesting - will investigate. Probably something with the device of unfinished_sents then!

@julien-c

[sshleifer]

Caused by previously unexposed kwargs changing behavior:
passing

do_sample=False, early_stopping=True

in the unit test fixes them.

[Laksh1997]

I'm getting this error @sshleifer

[Laksh1997]

#5221

[Laksh1997]

I have early stopping = False, but do_sample = True