Specifying --compiler=intelem is only possible using numpy.disutils (https://github.com/numpy/numpy/blob/main/numpy/distutils/intelccompiler.py). Since numpy.distutils is deprecated I guess it's natural that we stop supporting it as well.

To the question of supporting custom compilers, I'd be fine with making it simple at first and then see if we can add support later once our new build system is stable (and if we still find that there's a need for that)

Edit: I think it will still be possible to select the compiler through environment variables CC, CXX, ...

I have an build refactor removing distutils and numpy.disutils and only uses setuptools that successfully builds our wheels and passes tests

That's great !

Been worried about the whole distutils issue. I think this makes sense. Thanks for the work @thomasjpfan

Adding to the meson-related discussion here, I was recently playing with the editable installs (mesonbuild/meson-python#47) with meson and meson-python and it is nice.

Some thoughts: I've also found using meson to be quite useful and fast in developing a scikit-learn compatible package (https://github.com/neurodata/scikit-tree). Some cons are that using meson needs some getting used to, and there are some weird things you can't easily do (like run Python in the same root directory). But one can use spin (https://github.com/scientific-python/spin) to make the developer exp easier. I will say the meson.build files are also way easier to read and parse and understand compared to mega setup.py files.

With the recent inclusion of executable examples in the browser using JupyterLite in #25887, I prefer for meson to work with Pyodide before switching over. Currently there is work to get SciPy to build with Pyodide here: pyodide/pyodide#3649

Yes, @lesteve was looking into building scikit-image with meson for Pyodide in pyodide/pyodide#3874 which would be more comparable to scikit-learn build-wise. Scipy is another story.

I am going to close this one. There does not seem there has any demand for this, if this changes we can have a closer look.

I guess if someone has any interest with support Intel compilers with the Meson build, my advise would be to create a separate issue. Setting CC et al would be the first option. There may also be improvements to our meson.build files. There is a scipy issue about supporting Intel compilers that may be have some useful info, see scipy/scipy#17075.

Note that Spack has CI pipelines that exercise this. It should work, and otherwise there'll be a report at some point.

Note that Spack has CI pipelines that exercise this. It should work, and otherwise there'll be a report at some point.

OK great to hear, it seems indeed there is some kind of scikit-learn Spack recipe: https://github.com/spack/spack/blob/develop/var/spack/repos/builtin/packages/py-scikit-learn/package.py