vulkan: sparse prefill flash attention for qwen4exp top-k masks - #10

Closed

LynxPDA wants to merge 15 commits into

Nathanw1014:strix-halo-vulkanfrom

LynxPDA:pr/sparse-fa-pp-fix

Commits on Sep 11, 2026

vulkan: sparse flash attention for qwen4exp top-k masks
LynxPDA
committed
20edd90
vulkan: size the sparse-FA compaction per-subgroup tally to the workgroup
LynxPDA
committed
7e56d92
docs: recipe - shared per-subgroup tally sizing bug in the sparse FA compaction
LynxPDA
committed
c46ecbc
docs: recipe - the sparse-FA shared-tile tiling contract
LynxPDA
committed
949cc2c
vulkan: fix grouped union prefill host addressing for non-square head/batch shapes
LynxPDA
committed
ec22659
docs: recipe - push constants carry the destination shape, not the dispatch tile's
LynxPDA
committed
161ab59
feat: enable the grouped union prefill for GQA caches by default
LynxPDA
committed
0b08831

Commits on Sep 12, 2026

vulkan: overlap the union scan of group g+1 with flash attention of group g
LynxPDA
committed
0209047

Commits on Sep 13, 2026

Commits on Sep 14, 2026

vulkan: widen the top-k radix pass and batch the emit scan
LynxPDA
committed
188988f
vulkan: fold the QSA score transpose into the top-k gather
LynxPDA
committed
380b00a
vulkan: clear the fusion label when a fusion is declined
LynxPDA
committed
929aa8d