Implement SparseK Attention mechanism — new GGML operator with CPU backend (GPU planned next)#16817

Open

yael-works wants to merge 16 commits intoggml-org:masterfrom

yael-works:feature/sparsek-attn-sycl

Commits on Oct 28, 2025

Add skeleton for GGML_OP_SPARSEK_ATTN (SparseK Attention): new operator definition and tensor creation, backend implementation pending to ggml.c/h
yael-works
committed
Add CPU support for SparseK Attention (without performance checks)
yael-works
committed
Merge branch 'master' into feature/sparsek-attn-sycl
yael-works
committed

Commits on Oct 29, 2025

fix: add missing prototypes for ggml_sparsek_attn_set/get_params in ggml.h Co-authored-by: Yael Shuker <[email protected]>
yael-works
committed

Commits on Oct 30, 2025

fix SparseK CPU operator implementation

committed
fix SparseK CPU operator implementation

committed
trigger refresh
yael-works
committed
test commit from Gitty
Gitty Burstein
committed
remove test file
Gitty Burstein
committed
feat: implement SparseK attention core logic

committed
Implement final optimized SparseK Attention (CPU) Co-authored-by: Yael <[email protected]>
yael-works
committed
style: remove trailing whitespace and fix indentation in test-backend-ops.cpp

committed
delete Trailing whitespace
Gitty Burstein
committed
Update tests/test-backend-ops.cpp

GittyBurstein
and
CISC
authored
Update tests/test-backend-ops.cpp

GittyBurstein
and
CISC
authored
Update tests/test-backend-ops.cpp

GittyBurstein
and
CISC
authored