distributed training dataloader setting

[wjlyclll]

Thanks for publishing this useful work! May I ask a question will you keep different gpus training data with same aspect ratio (BucketManagers in different process share same seed) in one iteration during distributed training？Hope for your reply ~~