CI fails on Canberra metric tests on XLA

[vfdev-5]

🐛 Bug description

Canberra metric tests on XLA are failing:

>       assert canberra.pairwise([np_y_pred, np_y])[0][1] == pytest.approx(res)
E       assert 4.911688311688311 == nan ± ???
E         +4.911688311688311
E         -nan ± ???

https://github.com/pytorch/ignite/pull/1690/checks?check_run_id=1966203329

cc @sdesrozis

[vfdev-5]

Probably, the issue is related to the pull request where it happens.

[sdesrozis]

It seems weird because the mentioned PR not related to Canberra...

[vfdev-5]

Seems like the issue is here:

errors = torch.abs(y - y_pred) / (torch.abs(y_pred) + torch.abs(y))

where the random data gives x / tensor([ 4., 5., 11., 7., 0., 10., 10., 8., 7., 7.], device='xla:0') -> nan

[KickItLikeShika]

@sdesrozis, Yes it's definitely weird! i have not touched the Canberra! And now i'm working on other metrics #1695, i'm excited what's going to happen when i finish the tests and send the PR tomorrow, probably this we more things more clear, it would show us where is the problem exactly

[vfdev-5]

Fixed by #1699