Check GPU availability for cupy import

[ymelul]

cupy could be installed in the environment but:

• CUDA version could be not compatible
• The computer could not have a GPU

[ichristen]

Hi @ymelul ,

Thanks for this suggestion. I agree that a change similar to this would be very useful, and we should also handle the case where multiple GPUs are present.

However, there are files/subpackages other than algorithms that use cupy. I think we should instead do a package-level test/import for GPU-acceleration.

Torch is additionally a dependency for gradient-based optimization, and we've been meaning to add jax as well #89 , so this package-level import could also handle backend switching for gradients. For non-gradient stuff, I think we will still only support cupy because the drop-in replacement for numpy is more maintainable and the GPU memory is transferrable (with no cost) to torch and jax.

Best,
Ian

[ichristen]

I'm going to close this pull request for now, because we would want to implement the full package-level import before merging into main. I've created an issue #188 .