DDP support for training loop #110
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Easier for the trainers to communicate with each other
This is for DDP, for some reason DDP prefer to use ipv6 to communicate, disable ipv6 is simpler than extending the deployment to support ipv6.
Use torchrun to setup the environment and launch training script
It is does not necessarily run the rank 0 processes, but should still be fine
The task can be idle while waiting for other workers to start
It does not make sense to auto scale the cluster since the size is fixed
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add a new parameter
NUM_TRAINERSto specify the number of gpu instances. Only works with DDP branch of DeepEM.gpu_idsis important for specifying the number of processes in each instance, but does not actually restrict the gpus used. If its length is larger than the actual number of gpus the training will fail.batch_sizeandnum_workersare now batch_size and num_worker per process/gpu. So you need to scale down the numbers from previous training script to match the change.Not sure the current version is very compatible with spot instance training. Elastic training cannot be all spot instances and we are not using it now, any instance loss will cause the entire training to fail, to resume seems to require a new
rdvz_idwhich only happens with aupdate parametercommand for now.