<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Internals on 安橙的博客</title><link>https://blog.ans20xx.com/tags/internals/</link><description>Recent content in Internals on 安橙的博客</description><generator>Hugo -- 0.162.1</generator><language>zh</language><lastBuildDate>Mon, 25 May 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://blog.ans20xx.com/tags/internals/index.xml" rel="self" type="application/rss+xml"/><item><title>Day 11 · 显存管理</title><link>https://blog.ans20xx.com/posts/ai/day11/</link><pubDate>Mon, 25 May 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day11/</guid><description>拆开 PyTorch 的 caching allocator:Segment / Block 两级结构、small/large pool 分裂、stream 池隔离;掌握 memory_summary / memory_snapshot 工具链,学会 OOM 排查的标准流程,并理解 expandable_segments 如何根治碎片化。</description></item><item><title>Day 10 · 算子与后端</title><link>https://blog.ans20xx.com/posts/ai/day10/</link><pubDate>Thu, 21 May 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day10/</guid><description>拆开 PyTorch 算子实现层:c10 是底座,ATen 是算子库,native_functions.yaml 是真理之源。从代码生成管线到 TORCH_LIBRARY 注册机制,跑通一个自定义 CUDA 算子。</description></item><item><title>Day 09 · Autograd 原理</title><link>https://blog.ans20xx.com/posts/ai/day09/</link><pubDate>Wed, 20 May 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day09/</guid><description>用 ~200 行 Python 手写一个 mini-autograd:理解动态计算图的构建、链式法则的层层应用、拓扑排序的反向遍历。在最小可运行实现里看清 PyTorch backward() 的真面目。</description></item><item><title>Day 08 · PyTorch 核心抽象</title><link>https://blog.ans20xx.com/posts/ai/day08/</link><pubDate>Tue, 19 May 2026 00:00:00 +0000</pubDate><guid>https://blog.ans20xx.com/posts/ai/day08/</guid><description>打开 PyTorch 黑盒:Tensor 与 Storage 的分离、Dispatcher 的多重派发机制、Autograd Engine 的工作原理。跟踪一行 a + b 从 Python 一路调到 CUDA kernel 的完整路径。</description></item></channel></rss>