job fails on python 3.12

[louisabraham]

Description

This job failed

The reason is this pip bug so we should require pip >= 23.2

I'm trying a fix where I call pip install -U pip from CIBW_BEFORE_BUILD.

Build log

https://github.com/louisabraham/pydivsufsort/actions/runs/5851153678/job/15861602978

CI config

louisabraham/pydivsufsort@baf932e

[louisabraham]

My fix didn't work: https://github.com/louisabraham/pydivsufsort/actions/runs/5851270990/job/15861950058

[joerick]

I think the reason is numpy, not pip, see numpy/numpy#23808

[henryiii]

You can set PIP_PRE=1, that will get numpy 1.26.0b1.

[henryiii]

Actually, even better, set PIP_ONLY_BINARY=:all:, then you'll never try to build numpy or other dependent wheels, and it will get the only binary it can find, even though it's a pre-release. :) (PIP_ONLY_BINARY=numpy also works)

Edit: though if you have a lower bound, then you still need the PRE.

[yunjunz]

I got the same issue as well. Setting PIP_ONLY_BINARY=:all: does not seem to solve the issue.

[henryiii]

Do you have a lower bound? For some reason, that changes the resolution to more strongly avoid pre-releases. You need PIP_PRE if so.

[henryiii]

Oh, by the way, your example is incorrect. You can’t set environment variables directly, it must be via CIBW_ENVIRONMENT or use CIBW_ENVIRONMENT_PASS with it. It’s less error prone if you use pyproject.toml config.

[brenthuisman]

Actually, even better, set PIP_ONLY_BINARY=:all:, then you'll never try to build numpy or other dependent wheels, and it will get the only binary it can find, even though it's a pre-release. :) (PIP_ONLY_BINARY=numpy also works)

Unfortunately, both break. PIP_ONLY_BINARY doesn't work for PyPy builds (they don't publish PyPy binaries), and PIP_PRE gets you that 1.26 beta, but it doesn't (yet) build for PyPy (1.25.x does).

[joerick]

If you want to, TOML overrides should allow you to solve that. You could set the PIP_ONLY_BINARY environment only on CPython, for example.

[brenthuisman]

Yep, I'm awaiting CI on that :)

Just thought it would be good to point out there are no binary numpy wheels for PyPy.

[update] Just passed 🎉

[henryiii]

Yes there are. But only for PyPy 3.9. 3.8 isn’t supported and 3.10 doesn’t seem to be added yet.

[joerick]

NumPy 1.26.0 is now released, this issue should be solved.