torch.unique leads to "RuntimeError: isDifferentiableType" when used in loss function

[sambaPython24]

Hello, when I use the .unique function in PyTorch
as suggested in this post

combined = torch.cat((t1, t2))
uniques, counts = combined.unique(return_counts=True)
difference = uniques[counts == 1]
intersection = uniques[counts > 1]

and try to use it in a custom loss function, I get the following error

RuntimeError: isDifferentiableType(variable.scalar_type()) INTERNAL ASSERT FAILED at "..\\torch/csrc/autograd/functions/utils.h":64, please report a bug to PyTorch. 

cc @ezyang @gchanan @zou3519 @bdhirsh @jbschlosser

[ailzhang]

Marking as high pri since it hits internal assert but the use case looks valid.

[ezyang]

What version of PyTorch? Pretty sure @albanD fixed this one on master

[sambaPython24]

Hey, it is

print(torch.__version__)
1.7.1

and it is run on a CPU. Do you need any other information?

[albanD]

This has been fixed in master in #47930
You can use a nightly build to get the fix right now. Or wait for the upcoming 1.8 release.