{"type":"link","version":"1.0","title":"An efficient-attention mechanism's asymptotic memory win only appears if its block structure is never materialized, which requires hand-written kernels","author_name":"AI Archs","author_url":"https://ai-arch.pages.dev","provider_name":"AI Archs","provider_url":"https://ai-arch.pages.dev","url":"https://ai-arch.pages.dev/n/risk-efficient-attention-needs-custom-kernels-to-realize-savings","thumbnail_url":"https://ai-arch.pages.dev/og/risk-efficient-attention-needs-custom-kernels-to-realize-savings.png","thumbnail_width":1200,"thumbnail_height":630}