ZeRO2-Offload: Load balance gradient copying to CPU#1067
Merged
Conversation
tjruwase
requested review from
RezaYazdaniAminabadi,
ShadenSmith,
arashashari,
awan-10,
cli99,
conglongli,
minjiaz and
niumanar
as code owners
…/zero2_offload_balance_backward
eltonzheng
approved these changes
May 12, 2021
samyam
reviewed
May 13, 2021
Contributor
samyam
left a comment
samyam
left a comment
There was a problem hiding this comment.
This is quite clever. :)
Just documenting what we discussed on the potential simplification by just re-ordering the parameters in a round robin fashion directly in the self.fp16_groups to avoid all the book-keeping.
…/zero2_offload_balance_backward
Contributor
Author
My initial attempt at doing this hurt performance. So I am making it a TODO for when I have more time to investigate. |
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
During backward in ZeRO-2 Offload, reduced gradients are accumulated into CPU memory for later optimizer step. Due to the previous model partitioning scheme, gradient copying to CPU occurs one rank at a time, which slows down backward and under-utilizes PCIe. This PR introduces a new model partitioning scheme that spreads gradient copying evenly among all (most) ranks at any point in time. This improves backward time and PCIe utilization.