<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Layer-Wise Loading on Text Matrix</title><link>https://txtmix.com/tags/layer-wise-loading/</link><description>Recent content in Layer-Wise Loading on Text Matrix</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Tue, 21 Jul 2026 20:06:14 +0800</lastBuildDate><atom:link href="https://txtmix.com/tags/layer-wise-loading/index.xml" rel="self" type="application/rss+xml"/><item><title>AirLLM 技术拆解：单张 4GB GPU 如何跑 70B 大模型</title><link>https://txtmix.com/posts/tech/airllm-layer-wise-loading-70b-on-4gb-gpu/</link><pubDate>Sun, 19 Jul 2026 03:10:00 +0800</pubDate><guid>https://txtmix.com/posts/tech/airllm-layer-wise-loading-70b-on-4gb-gpu/</guid><description>&lt;h2 id="核心判断">核心判断&lt;/h2>
&lt;p>AirLLM 不是把模型&amp;quot;塞进&amp;quot;显存，而是把模型&amp;quot;流过&amp;quot;显存。它把 70B（700 亿）参数的模型按 Transformer 层切分后存放在磁盘上，推理时只把当前计算层加载进 GPU，算完就丢弃，再加载下一层。因此 GPU 显存只需容纳单层参数——这是它能在 4GB 显存上跑 70B 模型的根本原因，也是它和 vLLM、TGI、TensorRT-LLM 这类&amp;quot;整模型驻留显存&amp;quot;路线的本质分歧。&lt;/p></description></item></channel></rss>