Support for torch on AMD / ROCm and NVIDIA / CUDA #439
Description
This is not an urgent item but in the future if / when we want to use AMD GPUs we need libtorch built for ROCm. This is mutually exclusive with libtorch built for CUDA but both builds use the same at::cuda interface so WCT code need not change to support one or the other.
If we wish to give our users maximum flexibility we can, in principle, move this from being a compile-time to a run-time choice by hacking wcb to build multiple instances of libtorch-dependent-plugins, each one on one of the builds of libtorch. This could include a third CPU-only variant. This would let a single binary distribution of WCT work on a variety of hardware.
Specifically the pytorch and two spng plugins would need this treatment. Though, perhaps by then we can consolidate them into one plugin, or dispense with the pytorch one if/when spng is ready to replace it. See also #343 for thoughts on naming of this package.