When doing dataset.map, does it play nicely with multi-GPU accelerate? I guess dataset.map creates a lock file. So do I need to surround it with main_process_first?
with PartialState().main_process_first:
dataset = dataset.map(...)
Is this documented anywhere?