
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>Yusheng Zheng</title>
      <link>https://www.yunwei37.com/blog</link>
      <description>Yusheng Zheng is a systems researcher working on GPU runtimes, distributed AI infrastructure, programmable systems, and agent observability.</description>
      <language>en-us</language>
      <managingEditor>yunwei356@gmail.com (Yusheng Zheng)</managingEditor>
      <webMaster>yunwei356@gmail.com (Yusheng Zheng)</webMaster>
      <lastBuildDate>Fri, 18 Sep 2026 15:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.yunwei37.com/tags/thunderbolt/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.yunwei37.com/blog/qwen-tp2-nextn-thunderbolt-tx-e2e</guid>
    <title>Two RTX 5090s, One Qwen Model, and a Thunderbolt Flow-Control Bug</title>
    <link>https://www.yunwei37.com/blog/qwen-tp2-nextn-thunderbolt-tx-e2e</link>
    <description>A complete experiment report on running Qwen 27B with TP2 and native speculative decoding across two RTX 5090 hosts, finding a Thunderbolt transport bottleneck, and validating an opt-in Linux TX-E2E fix.</description>
    <pubDate>Fri, 18 Sep 2026 15:00:00 GMT</pubDate>
    <author>yunwei356@gmail.com (Yusheng Zheng)</author>
    <category>llm-inference</category><category>gpu</category><category>linux</category><category>thunderbolt</category><category>systems</category>
  </item>

    </channel>
  </rss>
