Fix LayerNorm fusion selecting the Div output as scale or bias - #31148
Open
Sammy-Dabbas wants to merge 1 commit into
Open
Fix LayerNorm fusion selecting the Div output as scale or bias#31148Sammy-Dabbas wants to merge 1 commit into
Sammy-Dabbas wants to merge 1 commit into
Conversation
The fusion could pick an operand produced inside the matched subgraph, then remove its producer, leaving a dangling input that failed graph resolve (microsoft#30513). Exclude in-subgraph operands from scale/bias candidacy.
|
Azure Pipelines: There may be pipelines that require an authorized user to comment /azp run to run. |
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #31147.
The scale selection in LayerNormFusion and SimplifiedLayerNormFusion picks a Mul input by rank alone, so a rank-1 Div output can be chosen as scale. The fusion removes the Div and the fused node references a dangling input, failing session creation with "Node input 'nm' is not a graph input, initializer, or output of a previous node". The bias loop can pick the Mul output the same way. This reproduces with keepdims=1, so it is separate from #31144.
The fix excludes operands produced inside the matched subgraph from scale and bias candidacy; when none remain, the fusion is skipped. Valid fusions are unaffected because a real scale always comes from outside the chain. Two negative tests added, one per fusion class; both fail before the change. No local C++ build available, so CI is the executable check.