Optimizing distributed Adam when running with one work queue #1551
Conversation
This file contains 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
The buckets perform communication together, so they are effectively a large bucket.
This was referenced Dec 7, 2022
crcrpar
reviewed
Jan 13, 2023
crcrpar
approved these changes
Jan 19, 2023
timmoon10
added a commit
to timmoon10/apex
that referenced
this pull request
Mar 2, 2023
Handles checkpoints generated before NVIDIA#1551.
yuanzhedong
pushed a commit
to yuanzhedong/apex
that referenced
this pull request
Jul 14, 2023
…1551) * Coalesce reduce-scatters in distributed Adam * Support variable-size param buckets in dist Adam optimizer * Support contiguous grad buffer with variable-size param buckets * Add dist Adam unit test with contiguous grad buffers * Optimize compute/communication overlap in dist Adam optim step * Restore Dist Adam default of splitting params across default-sized buckets * Support initializing multiple dist Adam param buckets together The buckets perform communication together, so they are effectively a large bucket. * Handle recent change in PyTorch API for coalescing NCCL calls
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
When @erhoo82 ran NeMo-Megatron with
CUDA_DEVICE_MAX_CONNECTIONS=1, he observed poor overlapping between the model's backward compute and the distributed Adam optimizer's gradient reduce-scatters. In particular, if multiple reductions are launched in a row, then only the last one is possible to overlap. This PR makes several changes to optimize performance: