openvino icon indicating copy to clipboard operation
openvino copied to clipboard
verified publisher icon openvinotoolkit

Refactor pagedAttention transpose

Open mangguo321 opened this issue 4 months ago • 1 comments

Details:

  • Move transpose functions from executor_pa.cpp to transpose.hpp to reuse in xattention and executor_pa.cpp. Modify transpose_16NxK logic to handle tails

Tickets:

mangguo321 avatar Dec 03 '25 03:12 mangguo321

Test on EMR, no regression in performance and accuracy. image

mangguo321 avatar Dec 09 '25 02:12 mangguo321