This field contains the worker-specific seed:
torch.utils.data.get_worker_info().seed
So I guess something like the below (untested!) could work too: worker_init_fn=lambda id: np.random.seed(torch.utils.data.get_worker_info().seed)