Needs rebase. I almost prefer the originals, because now I have to remember the builtin type hierarchy :)

Needs rebase. I almost prefer the originals, because now I have to remember the builtin type hierarchy :)

Yea I agree that those "type promotions" are certainly annoying, and actually plain wrong (e.g. int doesn't subclass float at runtime). So it's not strange to want to be explicit about it, and write int | float instead of float. I've actually tried doing this for a while, but quickly realized that it wasn't a good idea.

One issue is that the typing spec explicitly says that we should float instead of float | int (https://typing.python.org/en/latest/reference/best_practices.html#types). The same goes for ruff's PYI041 rule, which is enabled by default.

But besides formalities, there are actual practical problems that it can lead to. The main issue is that both mypy and pright don't always treat them as equivalent, which can lead to incorrect inference. It also can negatively affect the runtime of type-checkers in a very bad way. The incorrect type inference can be seen in scipy/scipy-stubs#492, and in data-apis/array-api-compat#291 (comment) I also explain my thoughts on the performance stuff in a bit more detail.

Diff from mypy_primer, showing the effect of this PR on type check results on a corpus of open source code:

vision (https://github.com/pytorch/vision)
- torchvision/utils.py:270: error: Unused "type: ignore" comment  [unused-ignore]

the circleci devdocs build failure seems unrelated

Thanks Joren.