Datamodule and DDP

[astrocyted]

I'm trying to wrap my head around why the Datamodule's setup hook is supposed to be called by all processes and GPUs independently:
https://github.com/Lightning-AI/lightning/blob/cb17fe90ec3b04b680e1b425431e3201529501c5/src/lightning/pytorch/core/hooks.py#LL357C1-L360C24

I mean since in the docs examples very commonly dataset splitting/partitioning is done in setup() e.g.:
https://github.com/Lightning-AI/lightning/blob/cb17fe90ec3b04b680e1b425431e3201529501c5/src/lightning/pytorch/demos/mnist_datamodule.py#L199

I just dont understand why anyone would want dataset splitting be carried independenly in different nodes, eventhough if they all share the same random seeds, its still redundant. why not just do it once on rank zero and then copy the datamodule similar to prepare_data? Could you explain the necessity of any operation within setup having to be done independently on all processes?