This would allow for device targeting found in other popular ML frameworks. Not everything is ideal to run on CUDA.