https://k4i.top/zh/tags/vllm/Recent content in Vllm on k4i's blogHugo -- gohugo.iozhsky_io@outlook.com (K4i)sky_io@outlook.com (K4i)All content is subject to the license of <a rel="license noopener" href="https://creativecommons.org/licenses/by-nc-sa/4.0/" target="_blank">CC BY-NC-SA 4.0</a> .Wed, 22 Apr 2026 10:00:00 +0800https://k4i.top/zh/posts/paged-attention/Wed, 22 Apr 2026 10:00:00 +0800sky_io@outlook.com (K4i)Sat, 30 May 2026 00:04:33 +0800https://k4i.top/zh/posts/paged-attention/<h2 id="memory-management-problem">显存管理问题</h2> <h3 id="fragmentation">碎片化问题</h3> <p>上一篇 <a href="https://k4i.top/zh/posts/kv-cache/">KV cache</a> 解释了为什么自回归解码可以缓存 key 和 value。KV cache 帮我们避免了大量重复计算，但也引出了一个新的系统问题：<strong>这些不断增长的缓存到底放在哪里？</strong></p>K4ifeatured imagellm推理systemsvllmmemoryAILLM Inference Internals