Repository navigation
fix: align causal mask bias with attention mask value - #108
Open
dajiaohuang wants to merge 1 commit into
Open
dajiaohuang wants to merge 1 commit into
dajiaohuang wants to merge 1 commit into
Conversation
Use -1e4 instead of -1e9 for masked_bias to match the -10000.0 value used in attention_mask processing, ensuring consistent pre-softmax attention scores for causally masked and padded tokens. Fixes salesforce#49
|
Thanks for the contribution! Unfortunately we can't verify the commit author(s): RepoStew Worker <w***@.repo>. One possible solution is to add that email to your GitHub account. Alternatively you can change your commits to another email and force push the change. After getting your commits associated with your GitHub account, sign the Salesforce Inc. Contributor License Agreement and this Pull Request will be revalidated. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Fix inconsistent mask values in
CodeGenAttentionthat cause pre-softmax attention score discrepancies between causally masked tokens and padded tokens.Problem
In
modeling_codegen.py:masked_bias = -1e9(used for causal mask)attention_mask = (1.0 - attention_mask) * -10000.0(used for padding mask)These use different large negative values (-1e9 vs -10000), causing inconsistent pre-softmax attention scores as noted in issue #49.
Fix
Change
masked_biasfrom-1e9to-1e4to match the-10000.0value used in attention mask processing.Testing
The change aligns the two mask values, ensuring consistent behavior during inference when both causal masking and padding are present.
Fixes #49