ffcv
ffcv copied to clipboard
Distributed support for QUASI_RANDOM
Supports distributed training with QUASI_RANDOM traversal order.
Relies on the same code as non-distributed to generate the order. Then, divides this into num_ranks contiguous chunks. The last two ranks can have some overlap to ensure that all ranks have the same amount of data.
Thanks so much for this PR, sorry for the late reply! Did you try training anything with this sampler/were the results okay? If so, I'll merge it in to v1.1.0