<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
<title>gogongxt</title>
<link>https://gogongxt.com/</link>
<description>知识分享</description>
<language>zh-CN</language>
<lastBuildDate>Sat, 21 Nov 2026 16:30:51 GMT</lastBuildDate>
<atom:link href="https://gogongxt.com/feed.xml" rel="self" type="application/rss+xml"/>
<item>
<title>cp</title>
<link>https://gogongxt.com/posts/743081c8</link>
<guid isPermaLink="true">https://gogongxt.com/posts/743081c8</guid>
<pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate>
<description>cp核心来说，就是分成几个不同的思路点：</description>
</item>
<item>
<title>SGLang 分布式 Hang / GPU 调试指南</title>
<link>https://gogongxt.com/posts/d1b06489</link>
<guid isPermaLink="true">https://gogongxt.com/posts/d1b06489</guid>
<pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate>
<description>补充判别:</description>
</item>
<item>
<title>gpu-topo</title>
<link>https://gogongxt.com/posts/87e16e3c</link>
<guid isPermaLink="true">https://gogongxt.com/posts/87e16e3c</guid>
<pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate>
<description>NIC : Network Interface Card 就是网卡</description>
</item>
<item>
<title>Retract机制</title>
<link>https://gogongxt.com/posts/36295bad</link>
<guid isPermaLink="true">https://gogongxt.com/posts/36295bad</guid>
<pubDate>Thu, 16 Jul 2026 00:00:00 GMT</pubDate>
<description>一条 input 长度逼近 KV cache 容量的请求,在 decode 阶段会触发一次看起来很费时间的 “retract” 交互。本文讲清楚 retract 到底是什么、为什么会被触发、怎么实现的,以及它在普通模式和 PD(prefill-decode)分离模式下走的是两条完全不同的路。</description>
</item>
<item>
<title>coredump调试cuda kernel</title>
<link>https://gogongxt.com/posts/da25d1fe</link>
<guid isPermaLink="true">https://gogongxt.com/posts/da25d1fe</guid>
<pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate>
<description>首先新建test_coredump.py脚本，脚本内容如下：</description>
</item>
<item>
<title>hisparse</title>
<link>https://gogongxt.com/posts/454aab87</link>
<guid isPermaLink="true">https://gogongxt.com/posts/454aab87</guid>
<pubDate>Wed, 15 Jul 2026 00:00:00 GMT</pubDate>
<description>长上下文推理里,稀疏注意力解决了”算不动”的问题,却没解决”装不下”的问题。一个 1M 上下文的请求,即便每步 decode 只看 top-k 个 token,它的完整 KV cache 仍然得常驻 GPU HBM,否则没法被快速访问。结果是 batch size 很快被显存撑爆,throughput 在并发还没拉起来时就见顶了。</description>
</item>
<item>
<title>miniconda换目录</title>
<link>https://gogongxt.com/posts/7a275973</link>
<guid isPermaLink="true">https://gogongxt.com/posts/7a275973</guid>
<pubDate>Tue, 14 Jul 2026 00:00:00 GMT</pubDate>
<description>如果把conda/miniconda换一个目录，会出现一些脚本bash bang等硬编码写死路径的问题，需要修复</description>
</item>
<item>
<title>tbo-sbo</title>
<link>https://gogongxt.com/posts/dfbc1801</link>
<guid isPermaLink="true">https://gogongxt.com/posts/dfbc1801</guid>
<pubDate>Tue, 02 Jun 2026 00:00:00 GMT</pubDate>
<description>tbo 就是 two batch overlap，最开始由deepseek提出，主要的目标是基于deepep的实现思路，优化计算和通信的overlap</description>
</item>
<item>
<title>通信操作</title>
<link>https://gogongxt.com/posts/ae3d8aa2</link>
<guid isPermaLink="true">https://gogongxt.com/posts/ae3d8aa2</guid>
<pubDate>Mon, 18 May 2026 00:00:00 GMT</pubDate>
<description>主要介绍AllGather和AllReduce，别的通信操作可以参考下面的文档</description>
</item>
<item>
<title>vllm-ascend</title>
<link>https://gogongxt.com/posts/16f9b6d1</link>
<guid isPermaLink="true">https://gogongxt.com/posts/16f9b6d1</guid>
<pubDate>Mon, 18 May 2026 00:00:00 GMT</pubDate>
<description>一句话总结：vllm-ascend 是 vllm 的硬件插件（out-of-tree platform plugin），通过 Python entry_points 机制注入，覆盖 GPU 相关的算子/调度/通信层，复用 vllm 的 API 服务、模型定义、请求管理等上层逻辑。</description>
</item>
<item>
<title>awq</title>
<link>https://gogongxt.com/posts/3cbaabed</link>
<guid isPermaLink="true">https://gogongxt.com/posts/3cbaabed</guid>
<pubDate>Wed, 06 May 2026 00:00:00 GMT</pubDate>
<description>AWQ 的研究团队发现了一个非常有意思的现象：对于 LLM 的权重量化，只有 0.1%-1% 的权重（Salient Weights）决定了最终的精度。</description>
</item>
<item>
<title>量化-04-SmoothQuant</title>
<link>https://gogongxt.com/posts/ba6f0c8a</link>
<guid isPermaLink="true">https://gogongxt.com/posts/ba6f0c8a</guid>
<pubDate>Sat, 02 May 2026 00:00:00 GMT</pubDate>
<description>之前我们讲的算法是OBQ/GPTQ是量化权重的，把权重量化到W4/8，实际计算时还要转成bf16，并没有使用上INT8 tensor core加速计算</description>
</item>
<item>
<title>投机采样</title>
<link>https://gogongxt.com/posts/99587e5b</link>
<guid isPermaLink="true">https://gogongxt.com/posts/99587e5b</guid>
<pubDate>Wed, 29 Apr 2026 00:00:00 GMT</pubDate>
<description>论文：</description>
</item>
<item>
<title>拉格朗日乘数法</title>
<link>https://gogongxt.com/posts/11bad293</link>
<guid isPermaLink="true">https://gogongxt.com/posts/11bad293</guid>
<pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate>
<description>拉格朗日乘数法（Lagrange Multiplier Method）是数学优化中处理等式约束极值问题的经典方法。它通过引入“拉格朗日乘子”，巧妙地将一个有约束的问题转化为无约束的问题。</description>
</item>
<item>
<title>量化-02-OBQ</title>
<link>https://gogongxt.com/posts/1831e68d</link>
<guid isPermaLink="true">https://gogongxt.com/posts/1831e68d</guid>
<pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate>
<description>因为GPTQ是基于OBQ改进的，所以我们这里先讲OBQ算法</description>
</item>
<item>
<title>量化-03-GPTQ</title>
<link>https://gogongxt.com/posts/c3b29e4a</link>
<guid isPermaLink="true">https://gogongxt.com/posts/c3b29e4a</guid>
<pubDate>Tue, 28 Apr 2026 00:00:00 GMT</pubDate>
<description>上文提到，OBQ 的致命痛点在于 “行依赖（Row-wise dependency）”</description>
</item>
<item>
<title>量化-01-综述</title>
<link>https://gogongxt.com/posts/845b34d2</link>
<guid isPermaLink="true">https://gogongxt.com/posts/845b34d2</guid>
<pubDate>Sat, 25 Apr 2026 00:00:00 GMT</pubDate>
<description>最新博主在系统学习量化，之前都是以工程实践为主，缺少系统性的学习，最近找了一篇量化综述，学习了一下，收获颇丰</description>
</item>
<item>
<title>n卡驱动关系</title>
<link>https://gogongxt.com/posts/a2dc8e31</link>
<guid isPermaLink="true">https://gogongxt.com/posts/a2dc8e31</guid>
<pubDate>Thu, 23 Apr 2026 00:00:00 GMT</pubDate>
<description>需要保证一条自下而上的兼容链条：</description>
</item>
<item>
<title>sglang在pd分离下的router请求</title>
<link>https://gogongxt.com/posts/a94e59bb</link>
<guid isPermaLink="true">https://gogongxt.com/posts/a94e59bb</guid>
<pubDate>Wed, 22 Apr 2026 00:00:00 GMT</pubDate>
<description>本文档详细说明 SGLang Router 在 Prefill-Decode (PD) 分离模式下，如何将客户端请求转换并分发到 prefill 和 decode 服务器。</description>
</item>
<item>
<title>deepseek_v32部署</title>
<link>https://gogongxt.com/posts/dd4775d8</link>
<guid isPermaLink="true">https://gogongxt.com/posts/dd4775d8</guid>
<pubDate>Tue, 21 Apr 2026 00:00:00 GMT</pubDate>
<description>所有实例（Prefill 和 Decode）均需设置以下环境变量：</description>
</item>
</channel>
</rss>
