<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0">
  <channel>
    <title>Kevin Ye Blog</title>
    <link>https://www.kevinye.org/blog/</link>
    <description>Engineering, reflections, and reviews.</description>
    <language>zh-CN</language>
    <lastBuildDate>Thu, 16 Jul 2026 00:00:00 GMT</lastBuildDate>
    <atom:link xmlns:atom="http://www.w3.org/2005/Atom" href="https://www.kevinye.org/rss.xml" rel="self" type="application/rss+xml" />
    
    <item>
      <title>DeepSeek Sparse Attention：从 SWA 到稀疏 Attention 计算</title>
      <link>https://www.kevinye.org/blog/deepseek-sparse-attention/</link>
      <guid>https://www.kevinye.org/blog/deepseek-sparse-attention/</guid>
      <description>从 SWA 的固定窗口出发，解释 Lightning Indexer 如何为完整前缀打分并选出 top-k、融合式 Sparse Attention 算子如何按索引读取 KV，以及 DSA 如何降低 Attention 的计算量与显存读取量。</description>
      <pubDate>Thu, 16 Jul 2026 00:00:00 GMT</pubDate>
      <category>Engineering</category>
      <category>DeepSeek Sparse Attention</category><category>SWA</category><category>Attention</category><category>Transformer</category>
    </item>
    <item>
      <title>MLA 如何压缩 KV Cache：低秩潜变量、矩阵吸收与解耦 RoPE</title>
      <link>https://www.kevinye.org/blog/multi-head-latent-attention/</link>
      <guid>https://www.kevinye.org/blog/multi-head-latent-attention/</guid>
      <description>从 MHA、MQA、GQA 的 KV head 共享关系出发，解释 MLA 如何通过低秩 KV latent、矩阵吸收和解耦 RoPE 压缩推理缓存。</description>
      <pubDate>Sat, 11 Jul 2026 00:00:00 GMT</pubDate>
      <category>Engineering</category>
      <category>MLA</category><category>Attention</category><category>Transformer</category>
    </item>
    <item>
      <title>Transformer Tensor Parallel：列切分、行切分与 All-Reduce</title>
      <link>https://www.kevinye.org/blog/tensor-parallelism/</link>
      <guid>https://www.kevinye.org/blog/tensor-parallelism/</guid>
      <description>用 column-wise 和 row-wise 两种切分方式解释 Transformer TP：中间分片如何继续本地计算，输出贡献又如何通过 all-reduce 合并。</description>
      <pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate>
      <category>Engineering</category>
      <category>Tensor Parallelism</category><category>Transformer</category><category>Distributed Computing</category><category>AI Infra</category>
    </item>
    <item>
      <title>从 EngineCore step 切入理解 vLLM PD 分离的异步执行机制</title>
      <link>https://www.kevinye.org/blog/vllm-pd-async-execution/</link>
      <guid>https://www.kevinye.org/blog/vllm-pd-async-execution/</guid>
      <description>从 EngineCore step 的异步执行模型出发，梳理 vLLM 在 PD 分离场景下与 NIXL KV Cache 传输协同的关键路径。</description>
      <pubDate>Fri, 26 Jun 2026 00:00:00 GMT</pubDate>
      <category>Engineering</category>
      <category>vLLM</category><category>LLM Serving</category><category>PD Disaggregation</category><category>KV Cache</category><category>AI Infra</category>
    </item>
  </channel>
</rss>