load_gbs in build_dataloader, how it works?

[shensimeteor]

Hi everyone,

Anyone knows this load_gbs parameter (https://github.com/NVIDIA/NeMo-Aligner/blob/main/nemo_aligner/data/nlp/builders.py#L471), how it works?

I noticed if I set load_gbs = True, the dataloader, in each iteration, will respond data of whole global_batch_size (divided by data_parallel_size), while if it's False, it reponds data of single micro_batch_size.

My issue is if using load_gbs=True, the data will exceed memory so dataloader will crash. So I'm wondering if I can switch it to False. Or will it have other side effects of changing it to False?

Thanks!