[CPU] Optimize PagedAttention performance#27718

Closed

luo-cheng2021 wants to merge 8 commits intoopenvinotoolkit:masterfrom luo-cheng2021:luocheng/pa_vs_sdpa

+4,095-107

Commits on Nov 23, 2024

add linux-perf

luo-cheng2021
authored and
chenhu-wang
committed

Commits on Nov 25, 2024

when work items count is small, use hq instead of hk to get works; 2nd token uses static scheduler if B is small
luo-cheng2021
committed
fix broken rope pattern in chatglm PagedAttention path
luo-cheng2021
committed
fuse split/reshape into PagedAttention
luo-cheng2021
committed

Commits on Nov 27, 2024

Commits on Nov 28, 2024

prefer first token to loop along H
luo-cheng2021
committed

Commits on Nov 29, 2024

fix test case for PagedAttentionWithSplitFusion
luo-cheng2021
committed