Skip to content

Implement own DDP used without CPU offloading (#100)#100

Merged
mseeger merged 1 commit into
mainfrom
own_ddp2
May 1, 2026
Merged

Implement own DDP used without CPU offloading (#100)#100
mseeger merged 1 commit into
mainfrom
own_ddp2

Conversation

@mseeger
Copy link
Copy Markdown
Contributor

@mseeger mseeger commented May 1, 2026

Don't use DDP provided by Lightning Fabric. Instead, just do the all_reduce at the end of each iteration.
This is a major simplification, and results seem to get better.

Closes #88, #92, #98.


By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

@mseeger mseeger self-assigned this May 1, 2026
@mseeger mseeger merged commit 1e2ec65 into main May 1, 2026
7 checks passed
@mseeger mseeger deleted the own_ddp2 branch May 1, 2026 21:33
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.

UserWarning: The AccumulateGrad node's stream does not match the stream of the node that produced the incoming gradient

1 participant