✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: ac4dcc1. Link to the linter CI: here}

@ogrisel - might be you will be interested in this

Thanks for this! One question about the pragma nocover - is this because we don't have Intel GPUs in any of our CI or because the default PyTorch we use doesn't contain support for it? Or a third reason? Basically, it would be nice to make use of it so that we can catch bugs/it doesn't bitrot.

Thanks for this! One question about the pragma nocover - is this because we don't have Intel CPUs in any of our CI or because the default PyTorch we use doesn't contain support for it? Or a third reason? Basically, it would be nice to make use of it so that we can catch bugs/it doesn't bitrot.

Great question, it truly stems from me not having done my homework right. I had included a branch for the case that a PyTorch < 2.4 was used for testing, where the later call to torch.xpu would definitely fail. Since the CI systems use 2.4.1+ (from the lock files as far as I see) Codecov complained to me that line 1349 in _testing.py from the PR was uncovered. I had followed the convention in the same function for cupy underneath, but a better solution would be to # nocover the if hasattr(xp, "xpu"): or remove that branch entirely, rather than block coverage for all of the xpu skip checks. I'm willing to go with either, let me know what you'd recommend @betatim

For the impact on sklearn: it will only introduce a lot of skipped tests as none of the runners (as far as I see) have any GPU hardware that would interface with integrated GPUs starting with Intel's Meteor Lake or any Intel discrete GPUs. This seems to also unfortunately be the case with the Apple PyTorch Metal branch (as shown on codecov). I'll add some screenshots of some functional array_api testing in our environments using an Intel Max GPU.

Thanks for the explanations! Let's keep it the way the cupy nocover is. The main reason I asked is that I looked at the MPS clause above which had no "nocover" and that made me wonder why it was needed here.

Fine for me to skip testing this in CI. It would have been nice to have it but not required. Getting "exotic" hardware for CI is hard work :D