SoWhereAreYou

It has been solved.But it is another question could you help me? def compute_loss_with_global_feature(feature, loss_func, enable_grad=True): world_size = dist.get_world_size() rank = dist.get_rank() shape = (world_size, ) + feature.shape all_feature =...

That is really sad。。And your work is really excellent sincerely and im amdire for It.At the end could you tell me if the thought of the loss is moco to...

Thank you, i would try it

 I had try to debug like this and find it hangs in gelu function.. That is very scary.I found with the time change the version of torch in dockerfile...