<?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>CUDA IPC on Ryan's Blog</title><link>/tags/cuda-ipc/</link><description>Recent content in CUDA IPC on Ryan's Blog</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Sun, 01 Nov 2026 18:00:00 +0800</lastBuildDate><atom:link href="/tags/cuda-ipc/index.xml" rel="self" type="application/rss+xml"/><item><title>合璧之路：用 1200 行手搓 Agentic RL 双卡训推分离闭环（nano-miles + nano-openshell）</title><link>/posts/building-agentic-rl-nano-miles-openshell/</link><pubDate>Sun, 01 Nov 2026 18:00:00 +0800</pubDate><guid>/posts/building-agentic-rl-nano-miles-openshell/</guid><description>深度拆解如何将轻量强化学习框架 nano-miles 与零信任进程沙箱 nano-openshell 组装为一个完整的智能体 RL 闭环。涵盖多轮 Interleaved Loss Mask 的 Shift-by-One 严格对齐、无特权 Docker 容器下命名空间降级与看门狗进程树回收、GPU 0 训练与 GPU 1 推理物理分离架构，以及绕过硬件无 P2P 限制的单次连续 PCIe DMA Staging CUDA IPC（~122ms 零拷贝热更新）。附 RunPod 2x RTX 4090 真实 8 步收敛日志（准确率由 12.5% 跃升至 50%）。</description></item><item><title>大道至简：用 720 行纯 Python 手搓 miles 最小闭环（SGLang + GRPO + 50ms CUDA IPC 热更新）</title><link>/posts/building-nano-miles-from-scratch/</link><pubDate>Sun, 01 Nov 2026 00:30:00 +0800</pubDate><guid>/posts/building-nano-miles-from-scratch/</guid><description>深度拆解企业级强化学习框架 miles 源码，剥离数万行分布式通信噪音，仅用 720 行纯 PyTorch 代码完整复现 TITO 零重分词采样、纯向量化 GRPO 组内优势白化、SGLang 50ms 显存级 CUDA IPC 零拷贝热更新，以及全异步解耦生产者驱动器。包含 13 项硬核单测与 RunPod L40S 真实收敛遥测日志。</description></item></channel></rss>