Tag
PulseCol introduces a periodically refreshed column-sparse attention method for diffusion language models, achieving higher sparsity and up to 1.95x end-to-end speedup over FlashAttention while maintaining model quality.