How to use PyTorch-Direct, "Enabling direct memory access" or similar method?

Memory Format
1

I am reading PyTorch-Direct: Introducing Deep Learning Framework with GPU-Centric Data Access for Faster Large GNN Training | NVIDIA On-Demand
and

PyTorch-Direct: Introducing Deep Learning Framework with GPU-Centric Data Access for Faster Large GNN Training | NVIDIA On-Demand

It makes me wonder how to use these and it it’s supported as a first party feature nowadays since some time has passed.