On Sun, Sep 14, 2025 at 11:29 PM Hameer Abbasi <hameerabb...@yahoo.com> wrote:
> Hello! > > Is a universal2 wheel an option? > No. universal2 wheels only make the build & testing part harder, because now you have two separate architectures you should test on (typically people who ship them don't even realize this, don't test the non-native part and just YOLO release them). And universal2 wheels are kinda a bad idea to begin with, which is why we don't ship them in the first place. They're strictly worse in all respects both for maintenance burden and for the average end user. Cheers, Ralf > Best, > Hameer > > Le 14.09.2025 à 22:49, Ralf Gommers via NumPy-Discussion < > numpy-discussion@python.org> a écrit : > > > > > On Wed, Sep 10, 2025 at 3:28 PM Nathan via NumPy-Discussion < > numpy-discussion@python.org> wrote: > >> >> >> On Wed, Sep 10, 2025 at 7:02 AM Ralf Gommers via NumPy-Discussion < >> numpy-discussion@python.org> wrote: >> >>> >>> >>> On Wed, Sep 10, 2025 at 2:06 PM Hugo van Kemenade <h...@python.org> >>> wrote: >>> >>>> I believe Azure Pipelines and GitHub Actions essentially use the same >>>> infra, and Azure Pipelines is also dropping macOS Intel x86-64: >>>> >>> >>> I don't think that is correct. They use the same *runner images* and the >>> macos-13 image is indeed deprecated for Azure as well, but there's a large >>> difference in the hardware offered. Azure only just started giving preview >>> access to macOS arm64 runners, and macos-14/15 images still run on x86-64 >>> hardware by default. >>> >>> That said, I'm really not a fan of trying to move to Azure, for multiple >>> reasons: >>> 1. It's a lot of work for a temporary extension of the support window >>> only >>> 2. Azure is pretty awful to work with, and we have never done any >>> release builds on Azure. Starting now is very unappealing. >>> 3. Release process wise there are both new opportunities for mistakes, >>> and also we'd not be able to do synchronized uploads so Chuck would have to >>> upload the macOS x86-64 wheels manually before starting the trusted >>> publishing process (to avoid sdist becoming visible to users first). >>> 4. It indeed would prevent from fully moving to trusted publishing and >>> no longer doing manual uploads (however, the security aspect is not the >>> most important to me here, we only delay an improvement, it's not really a >>> regression). >>> >>> The case for doing the extra work isn't all that compelling, since NumPy >>> is quite easy to build on macOS - it's the easiest platform by some >>> distance, even easier than Linux because it's the only platform where a >>> BLAS library is always available (unless a user is still on macOS <=12, >>> which is a really small fraction of users). So it's very easy to build >>> compared to SciPy, Matplotlib, etc. - and I very much doubt that those >>> projects will be building on Azure. >>> >>> *** >>> >>> What Andrew suggests would be the more reasonable route: if testing with >>> cibuildwheel for x86-64 on arm64 through Rosetta works, we can decide to >>> accept that some features are not as well tested as they should be, but at >>> least most functionality is tested, and for the rest we rely on it not >>> changing. >>> >>> I'm not as convinced as Andrew that this will work, but I also don't >>> know that it won't. So let's see if we can move to macos-15 on arm64 and >>> get this to work? If it works well enough, then keep support for one more >>> year (and if not, then not)? >>> >> >> This makes sense to me. Let’s make a best-effort attempt using whatever >> approach cibuildwheel promulgates based on rosetta. If it requires a bunch >> of extra work we can re-assess. >> > > It can be made to work, but it isn't super easy. Here is a working build > with OpenBLAS https://github.com/numpy/numpy/pull/29756. That took quite > a while to get right. I briefly tried the Accelerate build, but that didn't > seem to be detected. I'll wait with doing more until cibuildwheel decides > whether this stays a cross build or will be changed to a native build under > Rosetta2 (https://github.com/pypa/cibuildwheel/issues/2592). > > Cheers, > Ralf > > > >> >> Apple’s last Intel-based hardware came out in 2019. That’s not too long >> ago and it would be a shame to encourage people to create more e-waste if >> we can reasonably avoid it. >> >> >>> Cheers, >>> Ralf >>> >>> >>> >>>> > mac-OS 13 Ventura >>>> > In accordance with our policy to support the n-1 version of OS >>>> images, we will initiate the deprecation of macOS 13 Ventura starting 1st >>>> September 2025, with plans to retire it by 14th November 2025. Customers >>>> currently utilizing macOS-13 in their pipelines are encouraged to >>>> transition to macOS-14 or macOS-15 images. >>>> > >>>> > Key Dates: >>>> > • Deprecation Start Date: 1st September 2025 >>>> > • Full Retiral Date: 14th November 2025 >>>> >>>> >>>> https://devblogs.microsoft.com/devops/upcoming-updates-for-azure-pipelines-agents-images/#mac-os >>>> >>>> (Although I see [1] above has delayed the schedule by one month: now >>>> starting 1st Oct, complete by 4th Dec, so I expect AP will follow this >>>> too.) >>>> >>>> Hugo >>>> _______________________________________________ >>>> NumPy-Discussion mailing list -- numpy-discussion@python.org >>>> To unsubscribe send an email to numpy-discussion-le...@python.org >>>> https://mail.python.org/mailman3//lists/numpy-discussion.python.org >>>> Member address: ralf.gomm...@gmail.com >>>> >>> _______________________________________________ >>> NumPy-Discussion mailing list -- numpy-discussion@python.org >>> To unsubscribe send an email to numpy-discussion-le...@python.org >>> https://mail.python.org/mailman3//lists/numpy-discussion.python.org >>> Member address: nathan12...@gmail.com >>> >> _______________________________________________ >> NumPy-Discussion mailing list -- numpy-discussion@python.org >> To unsubscribe send an email to numpy-discussion-le...@python.org >> https://mail.python.org/mailman3//lists/numpy-discussion.python.org >> Member address: ralf.gomm...@googlemail.com >> > _______________________________________________ > NumPy-Discussion mailing list -- numpy-discussion@python.org > To unsubscribe send an email to numpy-discussion-le...@python.org > https://mail.python.org/mailman3//lists/numpy-discussion.python.org > Member address: einstein.edi...@gmail.com > >
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3//lists/numpy-discussion.python.org Member address: arch...@mail-archive.com
