Remove Fused LayerNorm deprecation #1931
Merged
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Outdated suggestions cannot be applied.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.
A deprecation warning for Fused LayerNorm was added in PR #1789, informing users that FLN would be removed in Composer 0.13. The thought behind this was that Low Precision LayerNorm was strictly better, since it does not depend on Apex, and has equivalent performance.
However, our NLP testing plan found that Low Precision LayerNorm is not scriptable (i.e. we cannot use torchscript to export a model with LowPrecision LayerNorm for inference), making it inferior to Fused LayerNorm for users who want to export their models.
We should hold off deprecating Fused LayerNorm until this issue has been resolved, since it is an important component of many NLP models. Both algorithms can live side-by-side for the time being---the documentation for LPLN clearly states it is an alternative to FLN.