Bugfix: Attention computation is not equivalent if fused_attn is false - #145
Open
kandelak wants to merge 1 commit into
Open
Bugfix: Attention computation is not equivalent if fused_attn is false#145kandelak wants to merge 1 commit into
kandelak wants to merge 1 commit into
Conversation
facebook-github-bot
added
the
CLA Signed
Author
|
To remind you. This problem came up also for other users (Here for instance: #149) and this PR solves it. |
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
There is a bug in calculation of the attention if fused_attn is set to false.
To replicate this bug, set fused_attn to false and you will get very poor reconstruction whereas if it is set to true (default), it works. With this change, it works again (reimplemented non-efficient fused_attention basically)
Possible reason: The training was done using F.scaled_dot_product_attention which is internally different from the "else branch" where attention calculation happens in a non-efficient way.