Resolve a validated device keyword to a concrete, available torch device

check_device_arg() has already rejected anything but "cpu", "cuda", "mps", "gpu" or "auto"; this turns that keyword into "cpu", "cuda" or "mps", matching what Python sbi's process_device() does. Requires torch (call require_torch() first; every caller here is about to build a network, so it already has).

Usage

resolve_device(device)

Arguments

device

One of "cpu", "cuda", "mps", "gpu", "auto".

Value

"cpu", "cuda" or "mps".

Details

"cuda"/"mps" name a specific device, so asking for one that is not there errors rather than downgrading silently – a silent fallback would hide a real problem (a missing CUDA build, a non-Apple-silicon Mac). "gpu"/"auto" never named a specific device, so it resolves CUDA -> MPS -> CPU and can fall back silently, mirroring sbi's "gpu".