https://contact.alessandrosangiorgi.net/tags/qwen/Recent content in Qwen on Alessandro Sangiorgi — GPU Performance EngineerHugoenSun, 27 Sep 2026 00:00:00 +0000https://contact.alessandrosangiorgi.net/posts/qwen3-8-vllm-responses-api-system-message-chat-template/Sun, 27 Sep 2026 00:00:00 +0000https://contact.alessandrosangiorgi.net/posts/qwen3-8-vllm-responses-api-system-message-chat-template/<p>I run <a href="https://github.com/openai/codex">Codex CLI</a> against a locally hosted <strong>unsloth/Qwen3.8-27B-NVFP4</strong>, served by vLLM on an RTX 5090. It&rsquo;s wired up as a Codex profile (<code>qwen38.config.toml</code>) pointing at <code>http://10.8.0.34:8000/v1</code>. It had been working. Then one session it started failing, reliably, with this:</p> <div class="highlight"><pre tabindex="0" style="color:#e6edf3;background-color:#0d1117;-moz-tab-size:4;-o-tab-size:4;tab-size:4;-webkit-text-size-adjust:none;"><code class="language-json" data-lang="json"><span style="display:flex;"><span>{<span style="color:#7ee787">&#34;error&#34;</span>:{<span style="color:#7ee787">&#34;message&#34;</span>:<span style="color:#a5d6ff">&#34;System message must be at the beginning.&#34;</span>,<span style="color:#7ee787">&#34;type&#34;</span>:<span style="color:#a5d6ff">&#34;BadRequestError&#34;</span>,<span style="color:#7ee787">&#34;param&#34;</span>:<span style="color:#79c0ff">null</span>,<span style="color:#7ee787">&#34;code&#34;</span>:<span style="color:#a5d6ff">400</span>}} </span></span></code></pre></div><p>The interesting part wasn&rsquo;t that it failed — it&rsquo;s <em>when</em> it failed. The very first prompt of a session always went through fine. The break only showed up on the <strong>second</strong> turn, right after the model made a tool call (a shell command) and Codex sent the result back. First turn: 200 OK. First tool round-trip: 400. Every time.</p>