it could be nice to avoid the overhead of moving the gpu box tensors to cpu, potentially.