Support NumPy 2.0

[jakirkham]

Hi All 👋

NumPy is planning to release 2.0 on June 16th ( numpy/numpy#24300 (comment) ). NumPy 2.0.0 RC packages for conda & wheels came out at the beginning of April ( numpy/numpy#24300 (comment) )

Also it looks like PyTorch has added NumPy 2.0 support ( pytorch/pytorch#107302 ) as part of 2.3.0

NumPy has put out a migration guide. More details are in the release notes

If torchvision makes use of NumPy's C API (and produces wheels that use it), having a release of torchvision with wheels built against NumPy 2.0.0rc1 would be helpful to ensure NumPy 1 & 2 compatible wheels (as wheels built against NumPy 1 won't be compatible with NumPy 2). More details in this NumPy 2 ABI doc

Also as NumPy is tracking ecosystem support for NumPy 2.0, it would be helpful to share torchvision current support status in issue (with any plans): numpy/numpy#26191

[rgommers]

It looks like some testing and fixes are in order indeed (in progress in gh-8461)

If torchvision makes use of NumPy's C API (and produces wheels that use it)

There should not be such a dependency. numpy isn't a build requirement in pyproject.toml, and there are no numpy headers being used directly.

[NicolasHug]

Thanks a lot for the update @jakirkham and @rgommers .

From #8461 and #8465 it looks like there's no more errors with numpy 2.0 (CI is mostly red but the relevant unittests-linux jobs are passing). The only required fix was #8463. So torchvision 0.19 should be fully compatible.

I also commented on numpy/numpy#26191 (comment).

[rgommers]

Great, thanks @NicolasHug!

[adamjstewart]

It looks like numpy 2 is still not supported on Windows? The wheel has an upper bound on numpy.

[adamjstewart]

Ah, I found #8530 and #8531. Maybe we should reopen this issue?

[NicolasHug]

No need to re-open this, we're aware of the problem and we're working on a fix. This is tracked in #8531 for torchvision and more generally in pytorch/pytorch#131668 for pytorch, as this isn't a torchvision-specific problem.