Fix attention bias calculation and dbias handling #199
Conversation
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
Corrects parenthesis to apply the matrix scaling before transpose when building the attention bias, aligning with the intended formula and improving numerical stability/broadcasting. Passes window size into the attention kernel to enable proper windowed masking and behavior.
Fixes incorrect dbias dimension handling in backward by deriving batch and query length from the bias tensor, not reused vars, ensuring correct allocation. Updates expansion/reduction logic for MQA/GQA and broadcasted dims (batch or seqlen_q == 1) to sum over the right axes, preventing mis-shaped outputs. Removes unused variables for clarity.
Contributor
There was a problem hiding this comment.
Pull Request Overview
This PR fixes the attention bias calculation to improve numerical stability by correcting the order of operations, and updates the dbias handling logic during backward operations by introducing dedicated variables for dimension tracking and adding window size parameter support.
- Corrects parenthesization in attention bias calculation to ensure multiplication happens before transpose operations
- Adds window_size parameter to the attention interface call
- Refactors dbias dimension tracking by replacing
batch_size_biasandseqlen_q_biaswithbatch_size_dbiasandseqlen_q_dbias
Reviewed Changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.
| File | Description |
|---|---|
| examples/modeling/modeling_doge.py | Fixes operator precedence in attention bias calculation and adds window_size parameter to attention call |
| csrc/flash_dmattn/flash_api.cpp | Removes unused mask/bias dimension variables and introduces dedicated dbias dimension tracking variables for proper backward pass handling |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
github-actions
bot
requested review from
Evanwu1125,
SNHuan,
Thanksyy,
ftgreat,
juliohsu,
wubingheng111 and
zacliu2023
github-actions
bot
assigned
Evanwu1125,
ftgreat,
juliohsu,
SNHuan,
Thanksyy,
wubingheng111 and
zacliu2023
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.
Correct the order of operations in the attention bias calculation for improved numerical stability and introduce window size handling. Adjust dbias shape and broadcasting logic to ensure proper dimension management during backward operations.