# FairScale should only depends on torch, not things higher level than torch.
torch >= 1.6.0
numpy >= 1.21