GitLab

Correct fused attention output after each step to reduce intermediate memory use.
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>