Skip to content

Conversation

@siddharth9820
Copy link
Collaborator

Why?

  1. Reduce scatters - happen on weight gradients, and researchers increasingly want to do these in fp32.
  2. All gathers - with torch.autocast, these were happening in fp32 by default. We can afford to do these in bf16/fp16.

@siddharth9820 siddharth9820 marked this pull request as draft February 26, 2024 15:50
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.

2 participants