📚 Documentation

As mentioned in the Changelog, version 2.5.0 introduces a robust scheme for seeding NumPy dependent dataloader workers.

As described in the related PR #20369, until this version Lightning was not handling correctly the seeding for NumPy, meaning that all projects using NumPy for controlling randomness (e.g. random transformations for augmentation) are affected (see here and also here).

Since NumPy is commonly used for augmentations in deep learning projects (see also the example from PyTorch docs) and proper control of seeding/randomness is essential for any project, I think it would be beneficial to add a note/warning in the docs to use version 2.5.0 or higher if randomness depends on NumPy.

cc @Borda