When using Ray’s Data and Train modules together, if the dataset hasn’t been materialized, it provides a batch execution (using read_* and map method) where the batches are sent to training as soon as they are ready. There’s also an option (here just use read_* method) to use a collate_fn within iter_torch_batches. I’m curious whether the collate_fn used inside that is processed in parallel across the Ray cluster, and how this approach differs from the first one I mentioned. (edited)