Guillaume Fraux

I think this is worth it: anything that include torch takes a lot of memory to compile, and sometime the front node has limits on how much memory a single...

> I think min(cpu_count, 8) is sufficient I'm not sure. We barely have 8 files to compile, so if the build fails with `make -j` it will also fail with...

We should check with the oldest torch version we support if this is still required. I think this was only an issue when using the conda version of torch though.

We had a discussion about this today, here is a quick summary: - use separate classes/objects for spherical harmonics and solid harmonics instead of a "normalize" parameter - use functional-looking...

The idea for jax was that we can use the calculator object to cache the state we need (right now we are using a hidden global cache). We are pretty...

This should reduce the number of wheels we need to build in #106

Not really: https://github.com/lab-cosmo/sphericart/actions/runs/7931040669/job/21654454335?pr=107 I think we need to pass a flag to pybind11 to use ABI3. The torch wheels I'm pretty confident are not directly linking to libpython, so they...

So this is not required for sphericart-torch (since we don't link to libpython directly) and is not supported by pybind11 for sphericart-jax. We'll have to find another way.

Regarding the distribution of the wheels, I had an idea last week: we could have some code in Python to dynamically pick a version of the shared library (depending on...

I guess the CI configuration now needs to be expanded to include all OS + all torch versions + more architectures?