Ray Train: Can it continue training with fewer workers when one goes down?

I am curious whether Ray Train supports dynamic scaling like for example torchrun --nnodes=1:4 torchrun (Elastic Launch) — PyTorch 2.5 documentation

I see the ScalingConfig can take a Domain or a dict instead of just an int value, but the documentation does not say how to use those: ray.train.ScalingConfig — Ray 2.41.0