Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Introduce Pipelined PA in 1.19 release #54

Open
wants to merge 5 commits into
base: v1.19.0
Choose a base branch
from

Conversation

mswiniarsk
Copy link
Contributor

No description provided.

mswiniarsk and others added 5 commits December 10, 2024 12:09
Introduce Pipelined PA, that applies the concepts of FlashAttention into PagedAttention, so that we can calculate softmax in parallel to matmul_av.
Cleanup code, and leave old attn implementation as a fallback option when index_reduce operator is not available (releases older than 1.19).

---------

Co-authored-by: Michal Adamczyk <[email protected]>
…y possible layout (#46)

* Fix adjustment tensor shape in PipelinedPA so that it matches with original shape

Previous solution caused a bug in models that have specific layout of [num_blocks, 1, num_heads , ...]
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.

1 participant