Change the command to initialize DDP subprocess with Hydra #16694
Labels
3rd party
Related to a 3rd-party
feature
Is an improvement or enhancement
strategy: ddp
DistributedDataParallel
won't fix
This will not be worked on
Description & Motivation
Hi,
the current behavior to initialize DDP subprocesses with Hydra is here, which overrides the hydra.run.dir with os.getcwd().
This is assuming the os.getcwd() will return the output directory that Hydra assigned. However, since Hydra 1.2, Hydra will keep the cwd as the project directory by default. This command will result in logging the subprocess in the project directory.
Pitch
Logging the DDP subprocess correctly in the output directory created by Hydra
Alternatives
use:
instead
Additional context
No response
cc @Borda @justusschock @awaelchli
The text was updated successfully, but these errors were encountered: