<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[双3090 nvlink Xid 175 - 为了解决问题 反而有意外惊喜]]></title><description><![CDATA[<p dir="auto">这几天发生了几次 GPU 挂掉（GPU没反应 nvtop 没反应）所以叫AI 解决</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>变更</th>
<th>原因</th>
</tr>
</thead>
<tbody>
<tr>
<td>驱动从 <code>nvidia-driver-595-open</code> 切换到 <code>nvidia-driver-595</code> (proprietary)</td>
<td>open kernel 驱动强制使用 GSP-RM，无法关闭。要关 GSP-RM 必须切回 proprietary。</td>
</tr>
<tr>
<td>关闭 GSP-RM (<code>NVreg_EnableGpuFirmware=0</code>)</td>
<td>Xid 175 是 GSP 固件超时挂起。关闭 GSP-RM 完全消除这一类故障。</td>
</tr>
<tr>
<td><code>gpu-memory-utilization</code> 从 0.9388 调到 <strong>0.9</strong></td>
<td>下午的 Xid 74 是 NVLink 错误，怀疑是显存使用率太高（98.6%)</td>
</tr>
</tbody>
</table>
<h3>当前配置 测试结果</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>测试项目</th>
<th>数值</th>
</tr>
</thead>
<tbody>
<tr>
<td>首响应时间 TTFT（短 prompt，冷启动）</td>
<td><strong>163 ms</strong></td>
</tr>
<tr>
<td>总响应时间（10 tokens）</td>
<td>293 ms</td>
</tr>
<tr>
<td>Prefill 1K</td>
<td>1,991 tok/s</td>
</tr>
<tr>
<td>Prefill 4K</td>
<td><strong>2,036 tok/s</strong></td>
</tr>
<tr>
<td>Prefill 16K</td>
<td>1,985 tok/s</td>
</tr>
<tr>
<td><strong>Decode（单流）</strong></td>
<td><strong>69.3 tok/s</strong></td>
</tr>
<tr>
<td>50K prompt 冷启动 TTFT</td>
<td>25.06 秒</td>
</tr>
<tr>
<td>50K prompt 缓存命中 TTFT</td>
<td>0.69 秒</td>
</tr>
<tr>
<td>缓存加速比</td>
<td>36.5×</td>
</tr>
</tbody>
</table>
<h3>配置 对比</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>测试项目</th>
<th>nvidia-driver-595-open, GSP-on, util=0.9388</th>
<th>当前 nvidia-driver-595, GSP-off, util=0.9</th>
</tr>
</thead>
<tbody>
<tr>
<td>Decode</td>
<td>67 tok/s</td>
<td><strong>69.3 tok/s</strong></td>
</tr>
<tr>
<td>Prefill 4K</td>
<td>1,289 tok/s</td>
<td><strong>2,036 tok/s</strong></td>
</tr>
<tr>
<td>Prefill 16K</td>
<td>—</td>
<td><strong>1,985 tok/s</strong></td>
</tr>
<tr>
<td>50K cold TTFT</td>
<td>~35 秒（估算）</td>
<td><strong>25.06 秒</strong></td>
</tr>
<tr>
<td>显存峰值（双并发负载）</td>
<td>23.6+ GiB</td>
<td><strong>22.7 GiB</strong></td>
</tr>
</tbody>
</table>
<hr />
<h3>vLLM 启动参数 (docker-compose.yml)</h3>
<pre><code class="language-yaml">--model /models/heretic-gptq-int4
--served-model-name qwen3.6-27b-heretic
--quantization gptq_marlin
--dtype float16
--tensor-parallel-size 2
--max-model-len 262144
--gpu-memory-utilization 0.9        # via ~/vllm/.env: GPU_MEM_UTIL=0.9
--max-num-seqs 2
--max-num-batched-tokens 8192
--kv-cache-dtype fp8_e5m2
--trust-remote-code
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder
--enable-prefix-caching
--enable-chunked-prefill
--disable-custom-all-reduce         # 必需 — 否则双 socket NUMA setup 在 CUDA graph profiling 阶段崩溃
--host 0.0.0.0
--port 8000
</code></pre>
<h3>硬件 &amp; 操作系统</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>配置</th>
</tr>
</thead>
<tbody>
<tr>
<td>GPU</td>
<td>2× NVIDIA GeForce RTX 3090 (24 GiB / card, NVLink)</td>
</tr>
<tr>
<td>功耗上限</td>
<td>250 W / GPU</td>
</tr>
<tr>
<td>OS</td>
<td>Ubuntu 24.04 LTS (noble)</td>
</tr>
<tr>
<td>内核</td>
<td>6.8.0-124-generic</td>
</tr>
<tr>
<td>NVIDIA 驱动</td>
<td><code>nvidia-driver-595</code> proprietary, v595.71.05 (DKMS-built)</td>
</tr>
<tr>
<td>GSP-RM</td>
<td><strong>禁用</strong> (<code>NVreg_EnableGpuFirmware=0</code> in <code>/etc/modprobe.d/nvidia-no-gsp.conf</code>)</td>
</tr>
<tr>
<td>Docker</td>
<td>29.5.2 + nvidia-container-toolkit 1.19.1</td>
</tr>
</tbody>
</table>
<h3>vLLM 引擎</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>配置</th>
</tr>
</thead>
<tbody>
<tr>
<td>版本</td>
<td>vLLM v0.21.0 (官方镜像 <code>vllm/vllm-openai:v0.21.0</code>)</td>
</tr>
<tr>
<td>部署方式</td>
<td>Docker Compose，systemd 启动 (<code>vllm-heretic.service</code>)</td>
</tr>
</tbody>
</table>
]]></description><link>https://lcz.me/topic/412/双3090-nvlink-xid-175-为了解决问题-反而有意外惊喜</link><generator>RSS for Node</generator><lastBuildDate>Sun, 26 Jul 2026 23:29:53 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/412.rss" rel="self" type="application/rss+xml"/><pubDate>Wed, 03 Jun 2026 15:57:02 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 双3090 nvlink Xid 175 - 为了解决问题 反而有意外惊喜 on Wed, 03 Jun 2026 20:54:50 GMT]]></title><description><![CDATA[<p dir="auto">不错，nvlink以后个人市场很难看到了，也算有考古价值了。</p>
]]></description><link>https://lcz.me/post/4889</link><guid isPermaLink="true">https://lcz.me/post/4889</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Wed, 03 Jun 2026 20:54:50 GMT</pubDate></item><item><title><![CDATA[Reply to 双3090 nvlink Xid 175 - 为了解决问题 反而有意外惊喜 on Wed, 03 Jun 2026 16:42:43 GMT]]></title><description><![CDATA[<p dir="auto">速度不错 <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=9786174bac0" class="not-responsive emoji emoji-android emoji--+1" style="height:23px;width:auto;vertical-align:middle" title=":+1:" alt="👍" /></p>
]]></description><link>https://lcz.me/post/4874</link><guid isPermaLink="true">https://lcz.me/post/4874</guid><dc:creator><![CDATA[Tony Wang]]></dc:creator><pubDate>Wed, 03 Jun 2026 16:42:43 GMT</pubDate></item></channel></rss>