Should Device::cuda_if_available return an error if the cuda feature is enabled, but the cuda device is not found?

[dsgallups]

Hello! I was wondering about the implementation detail behind Device::cuda_if_available. It seems like, since this is a runtime query for a device, then I had expected it to return the CPU if there is CUDA device at the ordinal index. I'd love to know more about the purpose of this function and where it's used (as well as a potential metal_if_available or host-agnostic device targeting).

Thank you!