I think we build them already (https://github.com/matplotlib/matplotlib/actions/workflows/cibuildwheel.yml) on every merge to main and 3.5.x, but do not upload them anyplace yet.

oooooh - Thats almost perfect!

You can get a list of them with:

curl   -H "Accept: application/vnd.github.v3+json"   https://api.github.com/repos/matplotlib/matplotlib/actions/artifacts

( following https://docs.github.com/en/rest/reference/actions) and from there it should be possible to install the correct version. The only issue is then that the download is pretty large (284 MB!).

So an amended version of this request is to upload two artifacts from that workflow:

1. The current wheels as is
2. A subset of the wheels (e.g. Ubuntu, x86_64, CPython) named something like CI-test-wheels

Coming here from #21637 (comment), I have a branch on my fork that, inspired by h5py (c.f. #21637 (comment)), adds a GitHub action workflow that is able to:

• Run on a scheduled CRON job and also on demand through workflow dispatch.
• Find the latest completed build of the wheels that originated from main.
• Download that builds "wheels" artifact that contains all of the wheels built.
• Install anaconda-client (from GitHub and not PyPI given Update anaconda-client release on PyPI anaconda/anaconda-client#540).
• Use the anaconda-client CLI API to upload the subset of all the wheels that are x86_64.manylinux wheels, given

A subset of the wheels (e.g. Ubuntu, x86_64, CPython) named something like CI-test-wheels

, to the scipy-wheels-nightly Anaconda Cloud organization as nightlies. This gets all the way up to the actual upload and then fails as the value of ANACONDA_ORG_UPLOAD_TOKEN that I currently have in my fork's repo secrets is just "FAKETOKEN" as I'm not a maintainer.

Example:

So to my mind this is working and as it doesn't need to build anything it is pretty fast too. 👍

@tacaswell, @QuLogic, @jklymak, @greglucas, @ianhi if this approach sounds good to you I can PR (though I should also add docs given #21637 (comment)). It looks like the other projects in the scipy-wheels-nightly Anaconda Cloud organization are uploading all of their wheels though, so if you'd like me to have this also upload all of the wheels instead of just the x86_64.manylinux I can certainly do that.