Supporting memory efficient dropout in flash attention #23
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
…d to ensure execution on device other than cuda:0
iclementine
reviewed
Jun 4, 2024
iclementine
approved these changes
Jun 5, 2024
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Outdated suggestions cannot be applied.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.
The main difference from Tri Dao's cuda implementation is the way we handle philox rng state. We cannot easily control the per thread philox offset increment since Triton makes thread abstraction opaque. That means we cannot reproduce the dropout masks exactly the same way as the cuda version.
The second difference is about the nuances of where to apply scaling, masking, casting and everything. It's an experience not guided by principles but rather trial and error.