IterableDataset with batch samples of different size

Hi,
I want to train a large dataset with accelerate.
However, my dataset has different shape (same batch, different sequence length) of samples.
This [line](https://github.com/huggingface/accelerate/blob/b2956acbe925845bf202224797334cfe4da18396/src/accelerate/data_loader.py#L356) requires the big batch for different processes to maintain a same shape so that it can concat.

I was wondering if we can just broadcast the tensors to other processors without forcing the data to be of same shape?