Skip to content

Conversation

@erhoo82
Copy link
Contributor

@erhoo82 erhoo82 commented Feb 14, 2023

This is to enable the overlap of AMAX reduction and the other DP-communications e.g. gradient reduction and parameter gathering.

@erhoo82 erhoo82 force-pushed the slym/amax_proc_group branch 2 times, most recently from 1f83729 to 36017e8 Compare February 14, 2023 19:25
@erhoo82 erhoo82 changed the title Draft: Use a separate communicator for AMAX reduction Use a separate communicator for AMAX reduction Feb 14, 2023
@erhoo82 erhoo82 force-pushed the slym/amax_proc_group branch from 36017e8 to 7505295 Compare February 14, 2023 22:08
@erhoo82 erhoo82 force-pushed the slym/amax_proc_group branch from 7505295 to 55a3d26 Compare February 15, 2023 05:22
@erhoo82
Copy link
Contributor Author

erhoo82 commented Feb 15, 2023

Thanks for the comments. I have applied the suggested changes.

@crcrpar crcrpar self-requested a review February 15, 2023 17:49
@crcrpar crcrpar merged commit 0c8400a into NVIDIA:master Feb 16, 2023
yuanzhedong pushed a commit to yuanzhedong/apex that referenced this pull request Jul 14, 2023
* add amax reduction group for fp8 training

* use a separate communicator for amax reduction across DP-ranks

reflect suggestion
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