Implement LayerNormalization fusion #280
Merged
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
robertknight
force-pushed
the
fuse-layer-norm
branch
2 times, most recently
from
5e90d30
to
779d6e3
Compare
This is useful for extracting constant values from nodes when creating graph optimizations.
- Add the ability to create symbols which match only constant values - Add the ability to give operators in patterns an identifier which can be used to look up the operator node after the pattern has matched. This is useful to perform additional tests on the operator such as matching particular attributes.
robertknight
force-pushed
the
fuse-layer-norm
branch
from
779d6e3
to
bd8a086
Compare
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Outdated suggestions cannot be applied.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.
This fuses subgraphs like below, excluding the "Add" in the top left corner, into a
LayerNormalizationoperator:So far this has been tested with BERT, GPT-2 and Whisper models.
This is a rather complex fusion and as a result, is liable to be brittle. That is, other equivalent ways of expressing the operator will not match. Also it doesn't help with operators which are variations of this such as RMSNorm (used in Llama etc.). In future it would probably make sense to break this up so that the individual sub-steps (centering, variance normalization, shift + scale) are fused separately.
TODO: