[cuda backend] optimized L_kv threshold for sdpa implementation selection. #20142

Gasoonjia wants to merge 3 commits into

mainpytorch/executorch:mainfrom

g4-opt-sliding-splitkpytorch/executorch:g4-opt-sliding-splitkCopy head branch name to clipboard