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.

Sign up for GitHub

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

Jump to bottom

Revert "MHA refac: rope without complex operations + query only as input of the forward" #22

Merged

vince62s merged 1 commit into main from revert-20-mha

Jun 12, 2024

Contributor

vince62s commented Jun 12, 2024 •

edited

Loading

Reverts #20

there were two oversights:

cos/sin embeddings need to be [, head_dim//2] for flash2_with_kv_cache but [, head_dim] for the non interleaved calculation
making key/value optional (since = to query in case of self) added a test that make things slower.

will split PR for clarity.


          Revert "refac rope without complex (#20)"

d353967

This reverts commit acd7b9a.

vince62s merged commit d4b8791 into main

4 checks passed

francoishernandez deleted the revert-20-mha branch

July 1, 2024 13:46

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

Labels

None yet