I have not tried Python / TorchScript / Jax. I could make it work for the examples. I have not ported the backward method, which would be instrumental for training workloads. I am working on that.
I am using sphericart now only for inference of MACE models in LAMMPS in kokkos. I am very interested in learning more about metatensor, sphericart is the only piece that I see with hardcoded cuda, the rest can be managed switching device=xpu for Intel gpus. Thanks for the comments. I will make changes.

We have 20% of speed up when we use the SYCL version of sphericart in symmetrix with LAMMPS (42k atmos, 100 steps, double precision)

Let me know if there is anything else to change.

@alvarovm Thank you very much for the contribution! We're looking through the code, hopefully we'll be able to make some finishing touches and merge it relatively soon

I'm going to revert all the DTYPE -> T changes for the moment to make things easier to review. We'll open a different PR for them later

cscs-ci run

cscs-ci run

I have not tested the backward function, and the pytorch integration, I will complete this during April.

Inference with C++ and Kokkos is OK, we are using this for production.