[CUDA] Enable XQA decode for GroupQueryAttention with attention sink - #29162

Merged

Tianlei Wu (tianleiwu) merged 7 commits into

mainfrom

tlwu/20260618/xqa_head_sink

Jun 20, 2026

Commits on Jun 19, 2026

Optimize FlashDecode split planning for local-window GQA
Tianlei Wu (tianleiwu)
committed
1e86cf4
Enable XQA for GQA head_sink decode
Tianlei Wu (tianleiwu)
committed
b1ee723
enable xqa by default when head sink exists
Tianlei Wu (tianleiwu)
committed
2d72641
add test and doc
Tianlei Wu (tianleiwu)
committed
00ba600
Merge remote-tracking branch 'origin/main' into tlwu/20260618/xqa_head_sink
Tianlei Wu (tianleiwu)
committed
2a919c1
scoped env var
Tianlei Wu (tianleiwu)
committed
8c8df9b
Address PR review feedback: fix head_size comment, profile_gqa import, and XQA testing doc
Tianlei Wu (tianleiwu)
committed
7306218