<?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[R9700的定制版Qwen3.8 27b又来了]]></title><description><![CDATA[<h1>Halogen / R9700 定制版性能测试</h1>
<p dir="auto">Halogen 让 <strong>AI Max 395</strong> 一下子牛逼了，现在 <strong>R9700</strong> 也有了定制版，号称 <strong>Prefill 干到 3000+ tok/s</strong>。</p>
<p dir="auto">有兴趣的可以试试：</p>
<p dir="auto"><a href="https://github.com/GGZ14/vllm-mxfp4" rel="nofollow ugc">https://github.com/GGZ14/vllm-mxfp4</a></p>
<hr />
<h2>Decode 性能</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Category</th>
<th style="text-align:right">Update P50</th>
<th style="text-align:right">Decode 速度</th>
</tr>
</thead>
<tbody>
<tr>
<td>Chat</td>
<td style="text-align:right">42.3 ms</td>
<td style="text-align:right"><strong>67.1 tok/s</strong></td>
</tr>
<tr>
<td>Code</td>
<td style="text-align:right">42.5 ms</td>
<td style="text-align:right"><strong>120.5 tok/s</strong></td>
</tr>
<tr>
<td>File Edit</td>
<td style="text-align:right">42.5 ms</td>
<td style="text-align:right"><strong>138.0 tok/s</strong></td>
</tr>
<tr>
<td>JSON</td>
<td style="text-align:right">42.4 ms</td>
<td style="text-align:right"><strong>153.1 tok/s</strong></td>
</tr>
<tr>
<td>Math</td>
<td style="text-align:right">42.5 ms</td>
<td style="text-align:right"><strong>140.0 tok/s</strong></td>
</tr>
<tr>
<td>Prose</td>
<td style="text-align:right">42.3 ms</td>
<td style="text-align:right"><strong>69.2 tok/s</strong></td>
</tr>
<tr>
<td>Reasoning</td>
<td style="text-align:right">34.3 ms</td>
<td style="text-align:right"><strong>123.9 tok/s</strong></td>
</tr>
<tr>
<td>Summarization</td>
<td style="text-align:right">34.2 ms</td>
<td style="text-align:right"><strong>141.7 tok/s</strong></td>
</tr>
</tbody>
</table>
<hr />
<h2>Prefill 性能</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th style="text-align:right">Prefill Depth</th>
<th style="text-align:right">Prefill 速度</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align:right">2K</td>
<td style="text-align:right"><strong>3552 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">8K</td>
<td style="text-align:right"><strong>3536 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">16K</td>
<td style="text-align:right"><strong>3619 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">32K</td>
<td style="text-align:right"><strong>3437 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">64K</td>
<td style="text-align:right"><strong>3192 tok/s</strong></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>64K Context 下依然能维持 3192 tok/s，16K 更是达到 3619 tok/s。</strong></p>
<hr />
<h2>并发性能</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th style="text-align:right">并发数</th>
<th style="text-align:right">总吞吐</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align:right">1</td>
<td style="text-align:right"><strong>120 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">2</td>
<td style="text-align:right"><strong>215 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">4</td>
<td style="text-align:right"><strong>322 tok/s</strong></td>
</tr>
<tr>
<td style="text-align:right">8</td>
<td style="text-align:right"><strong>471 tok/s</strong></td>
</tr>
</tbody>
</table>
<hr />
<h2>简单看一下数据</h2>
<ul>
<li><strong>Prefill：</strong> 基本稳定在 <strong>3000–3600 tok/s</strong></li>
<li><strong>Decode：</strong> 根据任务类型，大约 <strong>67–153 tok/s</strong></li>
<li><strong>单并发：</strong> 约 <strong>120 tok/s</strong></li>
<li><strong>4 并发：</strong> 总吞吐约 <strong>322 tok/s</strong></li>
<li><strong>8 并发：</strong> 总吞吐已经达到 <strong>471 tok/s</strong></li>
<li><strong>长上下文 Prefill：</strong> 到 64K 仍有 <strong>3192 tok/s</strong></li>
</ul>
<p dir="auto">如果这些数据在实际 Agent 工作负载下也能稳定复现，那 R9700 这套方案就相当有意思了。</p>
]]></description><link>https://lcz.me/topic/1813</link><generator>RSS for Node</generator><lastBuildDate>Fri, 23 Oct 2026 16:55:32 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1813.rss" rel="self" type="application/rss+xml"/><pubDate>Sat, 19 Sep 2026 00:45:56 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sun, 20 Sep 2026 04:18:52 GMT]]></title><description><![CDATA[<p dir="auto">R9700 32G 疯涨的原因找到了。</p>
]]></description><link>https://lcz.me/post/19460</link><guid isPermaLink="true">https://lcz.me/post/19460</guid><dc:creator><![CDATA[williamlouis]]></dc:creator><pubDate>Sun, 20 Sep 2026 04:18:52 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 12:33:37 GMT]]></title><description><![CDATA[<p dir="auto">帮你整理了下格式，下不为例。</p>
]]></description><link>https://lcz.me/post/19349</link><guid isPermaLink="true">https://lcz.me/post/19349</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sat, 19 Sep 2026 12:33:37 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 10:58:52 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/paul-hou" aria-label="Profile: paul-hou">@<bdi>paul-hou</bdi></a> <a href="/post/19331">说</a>:</p>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a><br />
這是模型胡言亂語了？</p>
</blockquote>
<p dir="auto">智力不够，才27b我不能怪他。</p>
]]></description><link>https://lcz.me/post/19335</link><guid isPermaLink="true">https://lcz.me/post/19335</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Sat, 19 Sep 2026 10:58:52 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 10:12:46 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E5%85%89%E7%92%9E" aria-label="Profile: 张光璞">@<bdi>张光璞</bdi></a><br />
這是模型胡言亂語了？</p>
]]></description><link>https://lcz.me/post/19331</link><guid isPermaLink="true">https://lcz.me/post/19331</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Sat, 19 Sep 2026 10:12:46 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 08:11:57 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/paul-hou" aria-label="Profile: paul-hou">@<bdi>paul-hou</bdi></a> <a href="/post/19257">说</a>:</p>
<p dir="auto">正名一下，GGZ14是原始版本，不是FORK<br />
已知FORK是<br />
<a href="https://github.com/magiccodingman/vllm-radiance" rel="nofollow ugc">https://github.com/magiccodingman/vllm-radiance</a><br />
<a href="https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main" rel="nofollow ugc">https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main</a></p>
</blockquote>
<p dir="auto"><img src="https://upload.lcz.me/uploads/7b69c8fa-8ddd-4faf-9bf2-12d5db14498d.jpeg" alt="010bd060-ef72-4b62-8e87-ec0f3c89ad9f-image.jpeg" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/19321</link><guid isPermaLink="true">https://lcz.me/post/19321</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Sat, 19 Sep 2026 08:11:57 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 01:18:34 GMT]]></title><description><![CDATA[<p dir="auto">正名一下，GGZ14是原始版本，不是FORK<br />
已知FORK是<br />
<a href="https://github.com/magiccodingman/vllm-radiance" rel="nofollow ugc">https://github.com/magiccodingman/vllm-radiance</a><br />
<a href="https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main" rel="nofollow ugc">https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main</a></p>
]]></description><link>https://lcz.me/post/19257</link><guid isPermaLink="true">https://lcz.me/post/19257</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Sat, 19 Sep 2026 01:18:34 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 01:02:23 GMT]]></title><description><![CDATA[<p dir="auto">数字能看，但两个 fork 的口径先拆开，不然收益会算重。</p>
<ol>
<li>表里的 decode t/s (med) 是各 category 的中位数，输出长度差很多：json/math/summarization 这类短输出天然拉高中位数，chat/prose（67–69 t/s）才接近长文稳态。<a class="plugin-mentions-user plugin-mentions-a" href="/user/johnnybegood" aria-label="Profile: johnnybegood">@<bdi>johnnybegood</bdi></a> 说的 3090 code 190+ 是另一种口径（模型/量化/上下文都不同），跨卡比之前先把这三个变量对齐。</li>
<li>Prefill 3000 t/s 在 27B 上不是不可能，前提是权重真按 FP4/MXFP4 参与 GEMM、KV 用 FP8。RDNA4 上 FP8 是原生矩阵路径；MXFP4 是走原生低精度还是 unpack 后再算，要看 vLLM 日志里的实际 dtype —— 若权重被 dequant 回 FP16，prefill 会明显掉档。</li>
<li>两个 fork 别混比：GGZ14 的 mxfp4 和 paiton 的 0.29.0/ROCm 10，模型（Unsloth NVFP4）、vLLM 版本、ROCm 版本三个变量同时在动。要对比就固定模型和 prompt 集，pp/tg 分开报，关掉 prefix cache。</li>
<li>验证：同一起点新起进程，固定 1k/4k/16k/32k 输入各跑一遍；rocm-smi 看显存带宽利用率和有没有 host offload；decode 取长输出（≥512 token）的稳态值。</li>
</ol>
<p dir="auto">NVFP4 权重在 ROCm 上能不能真跑起来，是这套方案成不成立的关键，跑通的话把 dtype 和 pp/tg 曲线贴出来一起看。</p>
]]></description><link>https://lcz.me/post/19252</link><guid isPermaLink="true">https://lcz.me/post/19252</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Sat, 19 Sep 2026 01:02:23 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 01:01:20 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E6%88%91%E7%88%B1%E5%B0%8F%E7%8C%AA" aria-label="Profile: 我爱小猪">@<bdi>我爱小猪</bdi></a> 3090的code是 190+</p>
]]></description><link>https://lcz.me/post/19251</link><guid isPermaLink="true">https://lcz.me/post/19251</guid><dc:creator><![CDATA[johnnybegood]]></dc:creator><pubDate>Sat, 19 Sep 2026 01:01:20 GMT</pubDate></item><item><title><![CDATA[Reply to R9700的定制版Qwen3.8 27b又来了 on Sat, 19 Sep 2026 00:54:45 GMT]]></title><description><![CDATA[<p dir="auto">這個R9700優化的FORK的第二版也推出了<br />
<a href="https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main" rel="nofollow ugc">https://github.com/Eliovp-BV/paiton-vllm-plugin/tree/main</a></p>
<p dir="auto">VLLM 升級到0.29.0<br />
ROCM 也升級到10<br />
預設模型是用Unsloth的NVFP4，沒錯…就是NV專用的NVFP4。</p>
]]></description><link>https://lcz.me/post/19247</link><guid isPermaLink="true">https://lcz.me/post/19247</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Sat, 19 Sep 2026 00:54:45 GMT</pubDate></item></channel></rss>