Memory Efficiency on Jaehun's Bloghttps://jaehun.me/en/tags/memory-efficiency/Recent content in Memory Efficiency on Jaehun's BlogHugoen-USTue, 08 Sep 2026 03:26:23 +0000Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attentionhttps://jaehun.me/en/posts/native-sparse-attention-hardware-aligned-and-natively-trainable-sparse-attention/Mon, 07 Jul 2025 00:00:00 +0900https://jaehun.me/en/posts/native-sparse-attention-hardware-aligned-and-natively-trainable-sparse-attention/<p><a href="https://arxiv.org/abs/2502.11089v2"target="_blank" class="inline-flex items-center gap-1" >Paper<svg class="h-3 w-3 flex-shrink-0" id="external-link" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M15 3h6v6m-11 5L21 3m-3 10v6a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2V8a2 2 0 0 1 2-2h6"/></svg> </a></p> <h2 id="native-sparse-attention-nsa--11-faster-even-at-64k-tokens-accuracy-intact">Native Sparse Attention (NSA) — 11× faster even at 64k tokens, accuracy intact<a href="#native-sparse-attention-nsa--11-faster-even-at-64k-tokens-accuracy-intact" class="heading-anchor" aria-label="Link to this section"><svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></g></svg></a></h2><h2 id="one-line-summary-tldr">One-line summary (TL;DR)<a href="#one-line-summary-tldr" class="heading-anchor" aria-label="Link to this section"><svg class="h-4 w-4" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"><g fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2"><path d="M10 13a5 5 0 0 0 7.54.54l3-3a5 5 0 0 0-7.07-7.07l-1.72 1.71"/><path d="M14 11a5 5 0 0 0-7.54-.54l-3 3a5 5 0 0 0 7.07 7.07l1.71-1.71"/></g></svg></a></h2><p><strong>NSA combines a three-branch sparse attention — “compress → select → slide” — with a GQA/MQA-friendly kernel to make decoding 11.6× faster and training up to 9× faster at 64k context, while improving average performance over Full Attention.</strong></p>