Automatic rank assignment in init_process_group

I am trying to setup distributed training and encountered some problems with initialization of process group.
Since I have a shared file-system between nodes, I chose initialization with file://. But I got this error:

ValueError: Error initializing torch.distributed using file:// rendezvous: rank parameter missing

Then I found in documentation, that “automatic rank assignment is not supported anymore”, although documentation for init_process_group imply otherwise.

Is there a way to not tell init_process_group rank explicitly. And what is the point of init_process_group if I have to pass rank explicitly.

Good point.

This used to be possible and was not reinstated when we moved to c10d for PyTorch 1.0. I created an issue on GitHub to bring this functionality back: