When `accelerator.prepare` is called on a
`datasets.IterableDataset`, use the `shard` method to
split the dataset across the available processes. This
allows for more efficient data loading and processing.
Without load and slice overhead of `IterableDatasetShard`