Non‑deterministic batch ordering in PyTorch DataLoader with multiple workers

Goal: Achieve deterministic batch ordering when using torch.utils.data.DataLoader with num_workers>0 while preserving the performance benefits of multiprocess data loading.

Constraint: Setting a global seed via torch.manual_seed does not propagate uniquely to each worker, so workers may generate identical augmentation sequences unless a worker_init_fn is supplied. The documentation leaves seed assignment to the user, creating uncertainty about the best practice for reproducibility.

Questions:

• Should PyTorch introduce a built‑in mechanism that automatically adjusts seeds for each DataLoader worker?
• Is there an officially endorsed pattern that guarantees deterministic ordering without requiring a custom worker_init_fn?
• What are the practical trade‑offs of implementing per‑worker seeding versus disabling multiprocessing for reproducibility?