From-scratch SM80 (A100/A800) CUDA kernels for GDN (Gated DeltaNet) and QSA (query-key sparse attention). Apache-2.0.
gdn ampere qsa a100 linear-attention cuda-kernel sparse-attention llm-inference sm80 gated-delta-rule
-
Updated
Sep 8, 2026 - Cuda