Skip to content

Conversation

LorrinWWW
Copy link
Collaborator

@DanFu09 Allreduce and CocktailSGD should work with amp now. Could you have a look and see if H3 is happy with that?

@DanFu09
Copy link
Collaborator

DanFu09 commented Feb 24, 2023

This seems to be failing when I use it with gradient compression:

Screenshot 2023-02-24 at 12 55 18 PM

Purple run: AMP + gradient compression. It spikes while still in the warmup stage. https://wandb.ai/hazy-research/cocktail-sgd/runs/2qor1jh4?workspace=user-danfu

Lime green run: AMP + all reduce. https://wandb.ai/hazy-research/cocktail-sgd/runs/3lfu92oo?workspace=user-danfu

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants