Conversation
neutrinoceros
left a comment
neutrinoceros
left a comment
There was a problem hiding this comment.
as is, you're effectively skipping all free-threaded builds. I would recommend instead to omit CIBW_BUILD entirely so the whole array of supported versions is run.
Indeed, no version of numpy has ever supported Python 3.10 on windows arm, so it's no use doing it downstream. |
e5bf007 to
b42effd
Compare
I have 0 experience with the free-threaded builds. Should we do them here?
My goal was to do to only oldest (forward compitable with ABO3) and latest (for future proofing) so I kept ``"cp311-* cp315-*"` there. As Python is getting more and more stable I don't see a reason to do them all. I'm fine removing that if you think it is worth to test them all. (We would need to re-do the PR skip from before.) |
b42effd to
5ee9f92
Compare
|
@neutrinoceros Let me know what you think. My idea is to reduce the maintenance burden and false positives, like the windows-arm one. I would like to remove more 😬 , like the x32 stuff. However, I think this is enough for now. |
|
Numpy stopped publishing wheels for win x32 a couple months back and as far as I could measure on astropy (using the same method as numpy folks) the downloads are really anecdotal. |
| CIBW_SKIP: "*-musllinux*" | ||
| CIBW_BUILD: "cp311-* cp315-*" |
There was a problem hiding this comment.
I would actually suggest keeping as much as cibw's configuration as possible in pyproject.toml:
- a single source of truth is preferable to two
- static configuration is easier to maintain and reason with
neutrinoceros
left a comment
neutrinoceros
left a comment
There was a problem hiding this comment.
a couple comments. I'm only looking at cibw configuration
Co-authored-by: Clément Robert <cr52@protonmail.com>
2 participants
This PR started as an investigation to solve #412 (comment), but I don't believe that windows-11-arm would ever pass on lower Python version. With ABI3 and in the light on the upcoming EoL of Python 3.10 we can clean the setup a little bit:
PS: There is no reason to use conda in the testing here. Different from netcdf4-python, where the netcdf-c, hdf5 from conda helps a lot, we can do an all pip setup for this repository. I'll change the CIs to reflect that here in a tomorrow.