Fuse MatMul + Mul/Div by constant #487
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
A subgraph of the form `Mul(MatMul(Mul(A, c), Mul(B, d)), e)` where all of the `Mul`s are optional and c, d and e are constants can be rewritten as `MatMul(A, B, alpha = c * d * e)` where `alpha` is the scaling already handled by the `C = alpha * AB + beta * C` result that GEMM already computes. Such scaling is common in transformers as part of attention operations (SDPA). The initial implementation only handles two specific cases of this form which have been seen in real models. This should be generalized to all possible cases.
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Outdated suggestions cannot be applied.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.
A subgraph of the form
Mul(MatMul(Mul(A, c), Mul(B, d)), e)where all of theMuls are optional and c, d and e are constants can be rewritten asMatMul(A, B, alpha = c * d * e)wherealphais the scaling already handled by theC = alpha * AB + beta * Cresult that GEMM already computes. Such scaling is common in transformers as part of attention operations (SDPA).The initial implementation only handles two specific cases of this form which have been seen in real models. This should be generalized to all possible cases.