SparseGDN: Chunk Sparse in Linear Attention
SparseGDN introduces a hardware-aware framework for Gated Delta Networks that reduces computation and memory traffic by combining structured intra-chunk approximation with input-dependent recurrent-event selection. Evalu...