Outputting Data to Multiple GPUs #56
Description
Description of feature
It seems anathema to a project whose goal is making small models that are io-bottlenecked more effecient that that project should be compatible with multiple GPUs, but especially with preload_to_gpu, it's possible that we want to control this behavior. Outside of that, I am not sure I see much value in this proposition because a user can just manage that themselves since we yield CPU-based samples otherwise.