<?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>DSA on Yunsheng Ni</title><link>https://niyunsheng.github.io/tags/dsa/</link><description>Recent content in DSA on Yunsheng Ni</description><generator>Hugo -- 0.154.5</generator><language>en-us</language><copyright>Content is licensed under CC BY-NC-SA 4.0.</copyright><lastBuildDate>Tue, 18 Aug 2026 10:00:00 +0800</lastBuildDate><atom:link href="https://niyunsheng.github.io/tags/dsa/index.xml" rel="self" type="application/rss+xml"/><item><title>From MLA to DSA to HCA/CSA: DeepSeek V3.2 and V4 Attention</title><link>https://niyunsheng.github.io/deepseek-mla-dsa-hca-csa/</link><pubDate>Tue, 18 Aug 2026 10:00:00 +0800</pubDate><guid>https://niyunsheng.github.io/deepseek-mla-dsa-hca-csa/</guid><description>MLA carried DeepSeek from V2 through V3.2 and was adopted across the industry — then V4 dropped it. V3.2 keeps MLA and bolts a lightning indexer on top: dense KV cache, sparse read. DeepSeek-V4 throws that layout away — every layer is a 128-token sliding window plus a compressed long-range branch, and layers alternate between CSA (m=4, indexer-selected) and HCA (m&amp;rsquo;=128, dense over blocks). Walking both through the reference implementations in huggingface/transformers.</description></item></channel></rss>