Why use local_loss=True with gather_with_grad=True?

[ThisisBillhe]

Hi,

Thanks for your great work! I notice that both local_loss and gather_with_grad are set to True in LIFT/configs/lift_vit_b16_128m.yaml. However, since local_loss is set to True, we do not need grad for features from other ranks, right?

[Jingfeng0705]

Hi! Yes, I believe you are right. We inherited these config files from some other repos, so some attributes may conflict with each other. During training, we didn't gather the loss from all ranks since our single gpu batch size was large enough.