I would like to use map_groups() on ray.data.grouped_data.GroupedData with GPU support.
I see that map_groups is mapped to map_batches internally. However, providing num_gpus=1 as parameter fails because it requires a fixed batch_size and map_group uses as batch_size=-1 (which depends on the group size).
Is it possible at all to enable GPU support for a map_groups() call?