When using -m torch.distributed.launch --nproc_per_node=2 as an example it uses --local-rank and not the expected --local_rank which generate a fatal error.
|
parser.add_argument("--local_rank", type=int, default=-1, |
Setting the arg to local-rank in train.py resolved the issue and it then works as expected
When using -m torch.distributed.launch --nproc_per_node=2 as an example it uses --local-rank and not the expected --local_rank which generate a fatal error.
ViT-pytorch/train.py
Line 281 in 460a162
Setting the arg to local-rank in train.py resolved the issue and it then works as expected