Use tool.scikit-build.cmake.version

[KyleFromNVIDIA]

Description

cmake.minimum-version has been deprecated since scikit-build-core 0.8, and is now causing conflicts in 0.10 due to its attempts to auto-detect cmake.version from CMakeLists.txt. Bump the minimum scikit-build-core to 0.10 and use the suggested cmake.version.

Contributes to rapidsai/build-planning#58

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

Just one suggestion, otherwise LGTM. Let's apply these changes across all of RAPIDS.

Should we be concerned about this affecting the 24.08 release?

[henryiii]

causing conflicts in 0.10 due to its attempts to auto-detect cmake.version from CMakeLists.txt

It's not supposed to do this unless minimum-version is unset or set to 0.10+.

[henryiii]

Oh, is minimum-version unset and you are using cmake.minimum-version? I might not have worked out that possibility (Edit: it seems fine, actually, in our tests). What's breaking, though, if it auto-detects the version?

[bdice]

#1638 (comment)

That's right, @henryiii. scikit-build-core 0.10.0 sees that the minimum-version (for scikit-build-core) is unset, and thus automatically sets cmake.version, which conflicts with our manually set cmake.minimum-version.

    WARNING: Use cmake.version instead of cmake.minimum-version with scikit-build-core >= 0.8
    ERROR: Cannot set both cmake.minimum_version and cmake.version; use version only for scikit-build-core >= 0.8.

[henryiii]

But it's not supposed to do this. I'm trying to see if I can reproduce this in our tests. It's supposed to take minimum-version if it's set and version isn't.

[henryiii]

I see the bug. I don't see why I can't trigger the bug in our tests yet. Could you try a branch? scikit-build/scikit-build-core#853

[jameslamb]

Could you try a branch?

Sure! I just pushed to #1638.

The wheel-build-* jobs will be the ones to watch: https://github.com/rapidsai/rmm/actions/runs/10275261685/job/28433610205?pr=1638

Using that PR which contains 0 other changes instead of this one because it'd tell us if new behavior from your scikit-build-core branch would allow RAPIDS to build successfully with a new version of scikit-build-core that contains that patch.

[jameslamb]

@henryiii it does look to me like the wheel builds are succeeding with your patch!

https://github.com/rapidsai/rmm/actions/runs/10275391215/job/28434079051?pr=1638

(ignore the failing conda-* jobs there... those are still pulling in scikit-build-core==0.10.0)

[jakirkham]

We are seeing this error on CI:

    ERROR: Cannot set cmake.verbose if minimum-version is set to 0.10 or higher

This should be fixed by upstream PR: rapidsai/devcontainers#377

[henryiii]

0.10.1 is out, with this regression fixed. In general, not setting minimum-version means you might be affected by updates, but this was a mistake. I'd recommend going 0.10+, of course, but the regression should be fixed. :)

[jameslamb]

Thank you SO MUCH for the quick fix @henryiii !!!

You're right of course, we should be setting minimum-version.

[henryiii]

Pushed to conda-forge too, that should be live in a few minutes.

[jameslamb]

Pushed to conda-forge too, that should be live in a few minutes.

Yep I see 0.10.1 there: https://anaconda.org/conda-forge/scikit-build-core

Thanks again @henryiii !!

[jameslamb]

The devcontainers build should be fixed once rapidsai/devcontainers#377 is merged.

[jakirkham]

/merge

[jakirkham]

Thanks all! 🙏