add basic support for the optimi adamw optimizer #1727
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Outdated suggestions cannot be applied.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.
https://optimi.benjaminwarner.dev/kahan_summation/
Kahan Summation¶
Kahan summation2 is a technique to reduce the numerical error of adding multiple low precision numbers by accumulating errors in a separate compensation buffer. The addition of the compensation buffer increases the effective summation precision by the precision of the compensation buffer.
Using Kahan summation to improve low precision model training was first introduced by Zamirai et al in Revisiting BFloat16 Training. Zamirai et al discovered the primary source of numerical error from low precision training is during the optimizer’s model weight update step. They add Kahan summation to the SGD & AdamW weight update steps to reduce the update’s numerical inaccuracy, increasing low precision training to the equivalent of full precision training across tested models.