> But I don't see a scenario where pytorch needs network access.

Training models across multiple compute nodes? That’s a big one.