Oscar icon indicating copy to clipboard operation
Oscar copied to clipboard
verified publisher icon microsoft

fix the deadlock problem when using distributed training in VQA fintune

Open Light-V opened this issue 3 years ago • 0 comments

When using distributed training, the process with local_rank!=0 will not call torch.distributed.barrier() and cause a deadlock.

Light-V avatar May 19 '22 06:05 Light-V