<?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[單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4]]></title><description><![CDATA[<p dir="auto">之前本地佈署都是靠著網站上的大神們的分享<br />
這個假日測了整整二天VLLM，最終使用了這個方案，結果不錯分享一下。</p>
<p dir="auto"><a href="https://github.com/magiccodingman/vllm-radiance" rel="nofollow ugc">https://github.com/magiccodingman/vllm-radiance</a></p>
<p dir="auto">模型</p>
<ul>
<li>主模型：/models/amd/Qwen3.8-27B-Quark-AWQ-MXFP4（served name: qwen3.8-27b-mxfp4）</li>
<li>Draft 模型：/models/tcclaviger/Qwen3.8-27B-DFlash2-FP8</li>
<li>量化：MXFP4（W4A8），RADIANCE_MXFP4=1 + 全套 radiance 優化旗標（WPERM、HOIST_QUANT、EPIFAST、SKINNY_GEMM、R4D attention 等）</li>
</ul>
<p dir="auto">引擎參數</p>
<ul>
<li>tensor-parallel-size: 1（單卡 R9700 / gfx1201）</li>
<li>gpu-memory-utilization: 0.98</li>
<li>kv-cache-dtype: fp8，--kv-cache-memory 7,783,339,733 bytes（~7.3GB 固定分配）</li>
<li>max-num-seqs: 8</li>
<li>max-model-len: 163,840</li>
<li>max-num-batched-tokens: 8,192</li>
<li>attention-backend: R4D（+ RADIANCE_USE_R4D_GDN / R4D_AR / R4D_AR_QUANT）</li>
<li>speculative decoding: dflash，num_speculative_tokens=7，TRITON_ATTN，disable_padded_drafter_batch</li>
<li>mamba-cache-mode: align</li>
<li>prefix caching: 啟用</li>
<li>no-async-scheduling</li>
<li>tool call: qwen3_xml parser；reasoning parser: qwen3</li>
<li>override generation config: temperature 0.7 / top_p 0.95 / top_k 20</li>
<li>chat template: qwen-fixed-v22.3.jinja（掛載為 /chat-template.jinja）</li>
<li>language-model-only、trust-remote-code、HF_HUB_OFFLINE</li>
</ul>
<p dir="auto">實際運行狀態（/metrics）</p>
<ul>
<li>KV cache：288 blocks，共 180,098 tokens 容量，fp8</li>
<li>prefix cache 命中率：1.85M / 2.44M ≈ 75.9%</li>
<li>dflash spec decode：48,588 draft tokens → 17,180 accepted，整體接受率約 35%；position 0 接受率 69%（4920/7108），逐位遞減到 position 6 約 15%</li>
<li>目前 0 running / 0 waiting，KV 使用率 0%</li>
<li>AITER：只啟用 UNIFIED_ATTENTION，其餘（MHA/MOE/MLA/RMSNORM/FP4BMM/FP8BMM）全關</li>
</ul>
<p dir="auto">Prefill（長 context 預填）速度：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>prompt 長度</th>
<th>實際 tokens</th>
<th>TTFT</th>
<th>prefill 速度</th>
</tr>
</thead>
<tbody>
<tr>
<td>8K</td>
<td>7,470</td>
<td>3.22s</td>
<td>2,319 tok/s</td>
</tr>
<tr>
<td>32K</td>
<td>29,742</td>
<td>11.49s</td>
<td>2,589 tok/s</td>
</tr>
<tr>
<td>64K</td>
<td>59,409</td>
<td>17.95s</td>
<td>3,311 tok/s</td>
</tr>
<tr>
<td>128K</td>
<td>118,772</td>
<td>45.52s</td>
<td>2,609 tok/s</td>
</tr>
</tbody>
</table>
<p dir="auto">Prefill 穩定在 2,300-3,300 tok/s，128K context 約 45 秒完成預填。</p>
<p dir="auto">併發測試（每路 max 256 tok）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>併發路數</th>
<th>wall time</th>
<th>總輸出</th>
<th>聚合速度</th>
<th>單路速度</th>
</tr>
</thead>
<tbody>
<tr>
<td>1（先前）</td>
<td>-</td>
<td>298 tok</td>
<td>41.5 tok/s</td>
<td>41.5</td>
</tr>
<tr>
<td>2</td>
<td>7.5s</td>
<td>459 tok</td>
<td>61.4 tok/s</td>
<td>29-34</td>
</tr>
<tr>
<td>4</td>
<td>8.7s</td>
<td>982 tok</td>
<td>113.5 tok/s</td>
<td>25-42</td>
</tr>
<tr>
<td>8</td>
<td>12.7s</td>
<td>1,941 tok</td>
<td>152.4 tok/s</td>
<td>27-42</td>
</tr>
</tbody>
</table>
<p dir="auto">實際在DSH的使用，單路思考時是很穩定在40 t/s左右，編程時會到70~80 t/s<br />
，比起llama.cpp + UD-Q4_K_XL 思考時25~40t/s，編程時50~75 t/s 效能好上不少。</p>
]]></description><link>https://lcz.me/topic/1529</link><generator>RSS for Node</generator><lastBuildDate>Mon, 07 Sep 2026 13:42:50 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1529.rss" rel="self" type="application/rss+xml"/><pubDate>Sun, 06 Sep 2026 15:03:32 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 11:13:08 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/kos-or" aria-label="Profile: kos-or">@<bdi>kos-or</bdi></a> 有測試內建的MTP，結果就是9xx的prefill，30左右的decode，看來真的跟dflash2有關了。</p>
]]></description><link>https://lcz.me/post/16421</link><guid isPermaLink="true">https://lcz.me/post/16421</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Mon, 07 Sep 2026 11:13:08 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 10:53:40 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/16247">said</a>:</p>
<p dir="auto">2,319 tok/s</p>
</blockquote>
<p dir="auto">Prefill 這速度讓人訝異, 難道是 MXFP4 + DFlas2-FP8 有神奇效果 <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f60d.png?v=2fb7360d8c6" class="not-responsive emoji emoji-android emoji--heart_eyes" style="height:23px;width:auto;vertical-align:middle" title=":heart_eyes:" alt="😍" /></p>
]]></description><link>https://lcz.me/post/16419</link><guid isPermaLink="true">https://lcz.me/post/16419</guid><dc:creator><![CDATA[kos or]]></dc:creator><pubDate>Mon, 07 Sep 2026 10:53:40 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 10:39:33 GMT]]></title><description><![CDATA[<p dir="auto">不確定是dflash2有加快，還是整個vllm設定和amd自己量化的模型有特別的優化，過程中裝了三種vllm，確實另外二個就是1200左右的prefill，decode 還有慢到只有15左右的。</p>
<p dir="auto">今天整天用下來比之前用vulkan + Q6或Q4快很多，不過2併發的編程就不用想了，run2只能做很簡單的小事，不然會整體拖慢速度。</p>
]]></description><link>https://lcz.me/post/16414</link><guid isPermaLink="true">https://lcz.me/post/16414</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Mon, 07 Sep 2026 10:39:33 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 09:32:50 GMT]]></title><description><![CDATA[<p dir="auto">还是前缀缓存命中了？</p>
]]></description><link>https://lcz.me/post/16398</link><guid isPermaLink="true">https://lcz.me/post/16398</guid><dc:creator><![CDATA[zhenyu huang]]></dc:creator><pubDate>Mon, 07 Sep 2026 09:32:50 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 09:32:10 GMT]]></title><description><![CDATA[<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> profill速度怎么那么快 是因为量化等级小吗</p>
]]></description><link>https://lcz.me/post/16397</link><guid isPermaLink="true">https://lcz.me/post/16397</guid><dc:creator><![CDATA[zhenyu huang]]></dc:creator><pubDate>Mon, 07 Sep 2026 09:32:10 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 08:02:31 GMT]]></title><description><![CDATA[<p dir="auto">電費，設備，預算，使用情境，老婆的臉色，目前只支持一張R9700。</p>
]]></description><link>https://lcz.me/post/16380</link><guid isPermaLink="true">https://lcz.me/post/16380</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Mon, 07 Sep 2026 08:02:31 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 06:38:45 GMT]]></title><description><![CDATA[<p dir="auto">用这个模型建议两张R9700，单流速度150以上</p>
]]></description><link>https://lcz.me/post/16359</link><guid isPermaLink="true">https://lcz.me/post/16359</guid><dc:creator><![CDATA[andyfay]]></dc:creator><pubDate>Mon, 07 Sep 2026 06:38:45 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 02:38:32 GMT]]></title><description><![CDATA[<p dir="auto">我是純LLM SERVER的方式在用，所以可以壓榨所有的VRAM來使用，另外是放棄了多模態沒上載入。二併發的情況下還是滿好用的。<br />
這套99.9%都是HERMES+DSV4F安裝好的，那2G應該可以想辦法搞出來。<br />
DSH新會話大約是3秒就開始動作，所以PREFILL是破3000的，這是我用這個的重點。</p>
<p dir="auto"><img src="https://upload.lcz.me/uploads/127ddb57-9700-405a-a0e8-b9ea40271a7e.jpeg" alt="1225f3f2-cf51-4b23-9959-f10088fdb83b-image.jpeg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/16326</link><guid isPermaLink="true">https://lcz.me/post/16326</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Mon, 07 Sep 2026 02:38:32 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Mon, 07 Sep 2026 01:41:02 GMT]]></title><description><![CDATA[<p dir="auto">多谢分享，但我试了，同样参数我显存不够（大概多了2g），我开机100M显存占用。</p>
]]></description><link>https://lcz.me/post/16315</link><guid isPermaLink="true">https://lcz.me/post/16315</guid><dc:creator><![CDATA[linghu007]]></dc:creator><pubDate>Mon, 07 Sep 2026 01:41:02 GMT</pubDate></item><item><title><![CDATA[Reply to 單張R9700 AI PRO 32G + VLLM + amd/Qwen3.8-27B-Quark-AWQ-MXFP4 on Sun, 06 Sep 2026 15:13:23 GMT]]></title><description><![CDATA[<p dir="auto">補充一下，OS是CACHYOS，關閉所有圖形介面後，開機VRAM只佔65M。</p>
]]></description><link>https://lcz.me/post/16249</link><guid isPermaLink="true">https://lcz.me/post/16249</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Sun, 06 Sep 2026 15:13:23 GMT</pubDate></item></channel></rss>