Skip to content

Conversation

@mengqin
Copy link

@mengqin mengqin commented Dec 1, 2025

This is basic Sage Attention 3 support. Because it is still unstable and differs significantly from previous versions of Sage Attention, a separate switch --use-sage-attiention3 is provided to enable or disable it. You need to install Sage Attention 3 in your environment before enabling it.

Attention 3 takes effect is reduced to 1024 (although the improvement is
not significant at this scale).
@mengqin mengqin requested a review from guill as a code owner December 2, 2025 16:21
@mengqin mengqin requested a review from rattus128 December 2, 2025 16:26
@Kosinkadink
Copy link
Collaborator

@mengqin for this PR, could you remove the startup argument for sage3? sage3 is very lossy, so after this PR gets merged with that gone, we can create a node that will only apply sage3 to certain blocks + sampling steps to make the results much better.

we would want to avoid users turning on sage3 for everything, as that will basically make everything return very poor results.

@mengqin
Copy link
Author

mengqin commented Dec 11, 2025

@mengqin for this PR, could you remove the startup argument for sage3? sage3 is very lossy, so after this PR gets merged with that gone, we can create a node that will only apply sage3 to certain blocks + sampling steps to make the results much better.

we would want to avoid users turning on sage3 for everything, as that will basically make everything return very poor results.

Thanks for the review. I removed the command-line argument switch and only kept the relevant attention functions and their registrations.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants