<?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[Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens]]></title><description><![CDATA[<h1>DeepSeek-V4.1-Flash 性能测试报告</h1>
<blockquote>
<p dir="auto">测试时间：2026-10-23 ｜ 测试方式：API 流式/非流式实测（127.0.0.1:8080 + 公网隧道）<br />
引擎：Project Maya v1.1.25（<code>build/maya-ds4</code>）· 模型：<strong>DeepSeek-V4.1-Flash-1017（Maya-S 量化）</strong></p>
</blockquote>
<hr />
<h2>一、软硬件环境</h2>
<h3>硬件</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>规格</th>
</tr>
</thead>
<tbody>
<tr>
<td>CPU</td>
<td>AMD EPYC 7K62 48 核 / 96 线程（L2 24 MiB · L3 192 MiB）</td>
</tr>
<tr>
<td>内存</td>
<td>252 GB DDR4（单 NUMA 节点）</td>
</tr>
<tr>
<td>GPU 0</td>
<td>NVIDIA RTX PRO 5000 Blackwell · 48 GB · PCIe 4.0 x16</td>
</tr>
<tr>
<td>GPU 1</td>
<td>NVIDIA GeForce RTX 5060 Ti · 16 GB · PCIe 4.0 x16</td>
</tr>
<tr>
<td>GPU 2</td>
<td>NVIDIA RTX PRO 4500 Blackwell · 32 GB · PCIe 4.0 x16</td>
</tr>
<tr>
<td>磁盘</td>
<td>Predator GM7000HS 2TB NVMe（模型 339.5 GB 存放于此）</td>
</tr>
</tbody>
</table>
<h3>软件</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>版本</th>
</tr>
</thead>
<tbody>
<tr>
<td>操作系统</td>
<td>Ubuntu 24.04.5 LTS（内核 7.0.0-38-generic）</td>
</tr>
<tr>
<td>NVIDIA 驱动</td>
<td>595.99.02（CUDA 13.2 runtime）</td>
</tr>
<tr>
<td>CUDA 编译工具</td>
<td>nvcc CUDA 12.0 构建链 / sm_120</td>
</tr>
<tr>
<td>Maya 引擎</td>
<td><strong>v1.1.25</strong>（2026-10-23 发布）</td>
</tr>
<tr>
<td>模型量化</td>
<td>Maya-S：IQ2_XXS/IQ2_S/IQ3_XXS/IQ3_S 专家 + Q6_K Engram（339.5 GB）</td>
</tr>
<tr>
<td>草稿器</td>
<td>DSpark（<code>--dspark on</code>），保留率 ~75%</td>
</tr>
<tr>
<td>隧道</td>
<td>cloudflared 2026.10.1（quick tunnel）</td>
</tr>
<tr>
<td>Python</td>
<td>3.12.3</td>
</tr>
</tbody>
</table>
<h3>引擎配置</h3>
<p dir="auto"><code>--max-context 393216</code>（384K）· <code>--tp 2</code> · 3 GPU · 无视觉 · 专家分层：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>GPU</th>
<th>分得层</th>
<th>专家驻留</th>
</tr>
</thead>
<tbody>
<tr>
<td>CUDA0（48G）</td>
<td>[0, 42)</td>
<td>3534 / 6528（54%）</td>
</tr>
<tr>
<td>CUDA1（16G）</td>
<td>[16, 25)</td>
<td>961 / 3840（25%）</td>
</tr>
<tr>
<td>CUDA2（32G）</td>
<td>[26, 39) + head</td>
<td>2321 / 5376（43%）</td>
</tr>
</tbody>
</table>
<hr />
<h2>二、性能实测数据</h2>
<h3>1. 生成速度（流式，免思考，1200 tok/轮）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>轮次</th>
<th>首 token</th>
<th>速度</th>
</tr>
</thead>
<tbody>
<tr>
<td>冷启动</td>
<td>1.5 s</td>
<td>51.8 tok/s</td>
</tr>
<tr>
<td>热身第 2 轮</td>
<td>0.0 s</td>
<td><strong>56.3 tok/s</strong></td>
</tr>
<tr>
<td>热身第 3 轮</td>
<td>0.0 s</td>
<td><strong>55.4 tok/s</strong></td>
</tr>
</tbody>
</table>
<h3>2. 思考模式（thinking budget 修复后）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>指标</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>数学题端到端</td>
<td>4.2 s（思考 84 tok + 答案 48 tok）</td>
</tr>
<tr>
<td>全程速度</td>
<td>31.6 tok/s</td>
</tr>
<tr>
<td>结论</td>
<td>v1.1.22 的"思考失控吃光预算"问题已修复</td>
</tr>
</tbody>
</table>
<h3>3. Prefill（prompt 读取）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>场景</th>
<th>tokens</th>
<th>速度</th>
<th>耗时</th>
</tr>
</thead>
<tbody>
<tr>
<td>短文档</td>
<td>2,925</td>
<td>319 tok/s</td>
<td>9.2 s</td>
</tr>
<tr>
<td>长 prompt</td>
<td>14,541</td>
<td><strong>601 tok/s</strong></td>
<td>24.2 s</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">长 prompt 更快的原因：v1.1.19+ 的长 prompt 跳层（DeepInfer 方法，20-39 层只算最后 ~2.7K token）+ NVMe 页缓存变热。<strong>文档越长读取效率越高</strong>。</p>
</blockquote>
<h3>4. 对话缓存</h3>
<p dir="auto">重复同题首 token <strong>0.44 s</strong>（conversation checkpoint 命中），同类新题 0.0-1.5 s。</p>
<h3>5. 公网链路（cloudflared quick tunnel）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>指标</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>TLS 握手</td>
<td>0.14-0.31 s</td>
</tr>
<tr>
<td>首字节（TTFB）</td>
<td><strong>1.0-1.7 s</strong></td>
</tr>
<tr>
<td>可用性</td>
<td>HTTP 200，<code>/v1/models</code> 返回完整 384K 配置</td>
</tr>
</tbody>
</table>
<h3>6. 资源占用（384K KV cache 下）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>资源</th>
<th>状态</th>
</tr>
</thead>
<tbody>
<tr>
<td>GPU 显存</td>
<td>三卡满配计划内（专家池占 ~95 GB），空闲 450-590 MiB/卡</td>
</tr>
<tr>
<td>RAM</td>
<td>已用 113 G + 页缓存 125 G，可用 137 G</td>
</tr>
</tbody>
</table>
<hr />
<h2>三、纵向对比（同机 DeepSeek Maya-S）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>版本/配置</th>
<th>热身生成</th>
<th>16K prefill</th>
<th>备注</th>
</tr>
</thead>
<tbody>
<tr>
<td>v1.1.22 @ 256K</td>
<td>46-48 tok/s</td>
<td>479 tok/s</td>
<td>2026-10-22</td>
</tr>
<tr>
<td><strong>v1.1.25 @ 384K</strong></td>
<td><strong>55-56 tok/s</strong></td>
<td><strong>601 tok/s</strong></td>
<td>2026-10-23，<strong>+16% / +26%</strong></td>
</tr>
</tbody>
</table>
<p dir="auto">与 GLM Maya-L（同机 v1.1.25，384K，parallel 2）对比：GLM 44-46 tok/s，DeepSeek <strong>55-56 tok/s 快约 20%</strong>。</p>
<h2>四、校准与调优记录</h2>
<ul>
<li>DS4 校准（<code>./maya.sh --calibrate</code>）：<code>--tp 2</code>(40.0) &gt; <code>--tp 1</code>(39.4) &gt; <code>--tp 3</code>(37.2)——<strong>tp 2 为最优</strong>；3 卡显存互不相等（48/16/32 GB），TP 只能两两配对，其余各跑各的层</li>
<li>PCIe lane 自适应：引擎实测 CPU lane（0.24-1.1 ms/专家）vs GPU 流式（1.1-1.5 ms）vs 磁盘（5.3-6.3 ms），自动关闭不划算的流式</li>
<li>页缓存：随使用增长（当前 125 GB），磁盘 miss 持续减少</li>
</ul>
<h2>五、结论</h2>
<ol>
<li><strong>本机 DeepSeek Maya-S 已接近该硬件的理论上限</strong>：三卡利用率均衡，RAM/SSD 分层工作正常</li>
<li>v1.1.25 相比 v1.1.22 全面提升：生成 +16%、长 prompt 读取 +26%</li>
<li>384K 上下文的 KV 开销被引擎优化完全吸收，无回退</li>
</ol>
]]></description><link>https://lcz.me/topic/2534</link><generator>RSS for Node</generator><lastBuildDate>Sun, 25 Oct 2026 00:57:19 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2534.rss" rel="self" type="application/rss+xml"/><pubDate>Fri, 23 Oct 2026 12:52:39 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Sun, 25 Oct 2026 00:20:56 GMT]]></title><description><![CDATA[<p dir="auto">更新 更高精度版本的性能<br />
<img src="https://upload.lcz.me/uploads/d2db6699-c6d1-46b1-a9d5-a2473f8fdada.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/27502</link><guid isPermaLink="true">https://lcz.me/post/27502</guid><dc:creator><![CDATA[Peng Xiao]]></dc:creator><pubDate>Sun, 25 Oct 2026 00:20:56 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Sat, 24 Oct 2026 23:57:43 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/52c6a9cc-8628-4332-a404-27104dbe4d73.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /><br />
使用48G+32G两张卡性能更好，接近60 tokens/s了</p>
]]></description><link>https://lcz.me/post/27495</link><guid isPermaLink="true">https://lcz.me/post/27495</guid><dc:creator><![CDATA[Peng Xiao]]></dc:creator><pubDate>Sat, 24 Oct 2026 23:57:43 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Sat, 24 Oct 2026 15:39:03 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/peng-xiao" aria-label="Profile: Peng-Xiao">@<bdi>Peng-Xiao</bdi></a> 能看出速度是个什么趋势么？ 是受限于最慢那个卡， 还是三个卡的速度中和了一下。</p>
]]></description><link>https://lcz.me/post/27424</link><guid isPermaLink="true">https://lcz.me/post/27424</guid><dc:creator><![CDATA[johnnybegood]]></dc:creator><pubDate>Sat, 24 Oct 2026 15:39:03 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Fri, 23 Oct 2026 14:53:12 GMT]]></title><description><![CDATA[<p dir="auto">跑应该能跑起来，内存放不下就放nvme磁盘，性能就没测过了</p>
]]></description><link>https://lcz.me/post/27104</link><guid isPermaLink="true">https://lcz.me/post/27104</guid><dc:creator><![CDATA[Peng Xiao]]></dc:creator><pubDate>Fri, 23 Oct 2026 14:53:12 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Fri, 23 Oct 2026 14:48:10 GMT]]></title><description><![CDATA[<p dir="auto">128G内存可以不？可以的话，我再买一张4090 48G <a class="plugin-mentions-user plugin-mentions-a" href="/user/peng-xiao" aria-label="Profile: peng-xiao">@<bdi>peng-xiao</bdi></a></p>
]]></description><link>https://lcz.me/post/27103</link><guid isPermaLink="true">https://lcz.me/post/27103</guid><dc:creator><![CDATA[NY Zhao]]></dc:creator><pubDate>Fri, 23 Oct 2026 14:48:10 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Fri, 23 Oct 2026 14:34:33 GMT]]></title><description><![CDATA[<p dir="auto">力大飞砖！力大飞砖！</p>
]]></description><link>https://lcz.me/post/27099</link><guid isPermaLink="true">https://lcz.me/post/27099</guid><dc:creator><![CDATA[johnnybegood]]></dc:creator><pubDate>Fri, 23 Oct 2026 14:34:33 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Fri, 23 Oct 2026 14:09:03 GMT]]></title><description><![CDATA[<p dir="auto">内存总是越大越好，实在用不上就做磁盘缓存</p>
]]></description><link>https://lcz.me/post/27092</link><guid isPermaLink="true">https://lcz.me/post/27092</guid><dc:creator><![CDATA[Peng Xiao]]></dc:creator><pubDate>Fri, 23 Oct 2026 14:09:03 GMT</pubDate></item><item><title><![CDATA[Reply to Strata跑deepseek-V4.1-flash, Maya 实测 decode 50 tokens on Fri, 23 Oct 2026 13:31:13 GMT]]></title><description><![CDATA[<p dir="auto">我也一直在关注这个项目，最近A 卡的也都支持了， 都在考虑是否有必要把 ram 括到256g 了</p>
]]></description><link>https://lcz.me/post/27080</link><guid isPermaLink="true">https://lcz.me/post/27080</guid><dc:creator><![CDATA[farmer node]]></dc:creator><pubDate>Fri, 23 Oct 2026 13:31:13 GMT</pubDate></item></channel></rss>