<?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[X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响]]></title><description><![CDATA[<p dir="auto">测试了不MXFP4，速度120t/s。<br />
FP8原生还在尝试<br />
NVFP4，速度150t/s。</p>
]]></description><link>https://lcz.me/topic/1882</link><generator>RSS for Node</generator><lastBuildDate>Fri, 23 Oct 2026 22:15:52 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1882.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 22 Sep 2026 05:18:58 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Wed, 23 Sep 2026 09:54:42 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/9572c0fb-8ff4-4ca3-afb0-6e22e397067a.png" alt="Screenshot 2026-09-23 at 17.52.36.png" class="img-markdown  img-fluid" /><br />
<img src="https://upload.lcz.me/uploads/e99ad187-e39f-41a0-b07e-7897da11745d.png" alt="Screenshot 2026-09-23 at 17.52.55.png" class="img-markdown  img-fluid" /><br />
以上问了两轮，速度比较稳定。思考比较长，聪明的代价</p>
]]></description><link>https://lcz.me/post/20268</link><guid isPermaLink="true">https://lcz.me/post/20268</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Wed, 23 Sep 2026 09:54:42 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Wed, 23 Sep 2026 09:50:34 GMT]]></title><description><![CDATA[<p dir="auto">测试速度感觉有问题，自己使用单会话60t/s，3个会话基本上也没降多少。单任务再DSH中长上下文，保持50t/s，目前是双卡FP8，使用Dflash，应该是最佳</p>
]]></description><link>https://lcz.me/post/20267</link><guid isPermaLink="true">https://lcz.me/post/20267</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Wed, 23 Sep 2026 09:50:34 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Wed, 23 Sep 2026 09:11:38 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://upload.lcz.me/uploads/4913b2d0-a050-49cf-b8f3-3fc0b090835d.png" alt="Screenshot 2026-09-23 at 17.10.12.png" class="img-markdown  img-fluid" /><br />
<img src="https://upload.lcz.me/uploads/340435a9-0a9e-42a3-a2c1-6473be5a20dd.png" alt="Screenshot 2026-09-23 at 17.10.23.png" class="img-markdown  img-fluid" /><br />
<img src="https://upload.lcz.me/uploads/a390bbc0-536b-42ab-81b3-695133f22b42.png" alt="Screenshot 2026-09-23 at 17.10.31.png" class="img-markdown  img-fluid" /><br />
<img src="https://upload.lcz.me/uploads/906315ca-fc5c-40f1-873a-bbd4475e8dd4.png" alt="Screenshot 2026-09-23 at 17.10.39.png" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/20256</link><guid isPermaLink="true">https://lcz.me/post/20256</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Wed, 23 Sep 2026 09:11:38 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Wed, 23 Sep 2026 02:16:27 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/brian" aria-label="Profile: Brian">@<bdi>Brian</bdi></a><br />
原文件有提到Explicit text-only NVFP4→MXFP4 requantization... conversion is lossy and occurs only in device memory at model load.</p>
<p dir="auto">所以加快是有代價的，平衡點的做法是用Launch80/Qwen3.8-27B-PARO-MXFP4這個模型，Launch80他們自己量化的模型，多了什麼旋轉校正。</p>
<p dir="auto">VLLM的鏡像最好是重抓新的，就兩個方案都留著，看心情使用了。</p>
<p dir="auto">我今天就用Launch80/Qwen3.8-27B-PARO-MXFP4這個模型工作，看看有什麼差異。<br />
目前在LOW檔，智力沒有很明顯的差別。<br />
NVFP4 -&gt; 100~110 T/S<br />
PARO-MXFP4 -&gt; 55~60 T/S<br />
<img src="https://upload.lcz.me/uploads/01d8bb7f-8ae3-416d-8f67-668670d5d60c.jpeg" alt="dc9a54ca-11a0-42d4-b0a8-db7ecf7a615e-image.jpeg" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/20155</link><guid isPermaLink="true">https://lcz.me/post/20155</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Wed, 23 Sep 2026 02:16:27 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 08:57:22 GMT]]></title><description><![CDATA[<p dir="auto">我現在在用，froggeric/Qwen-Fixed-Chat-Templates還是必須要加上去。<br />
不加的話，XHIGH的思考會沒完沒了，不思考會極笨。</p>
<p dir="auto">就算加上不思考還是極笨，以前用Q6_K_XL的不思考多少還是能用。<br />
加上去後LOW和MEDIUM，讓DSH測思考的TOKEN數，差異極小，所以設哪個都好。<br />
XHIGH也不會沒完沒了，比LOW再長個10倍左右。</p>
<p dir="auto">大家的temperature設成多少？官方建議是1.0 思考模式，VLLM-RADIANCE的預設都是0.7 非思考。但不思考還真的不能用...</p>
]]></description><link>https://lcz.me/post/20034</link><guid isPermaLink="true">https://lcz.me/post/20034</guid><dc:creator><![CDATA[paul hou]]></dc:creator><pubDate>Tue, 22 Sep 2026 08:57:22 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 07:50:19 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/brian" aria-label="Profile: Brian">@<bdi>Brian</bdi></a> 有图有真相<br />
上图</p>
]]></description><link>https://lcz.me/post/20022</link><guid isPermaLink="true">https://lcz.me/post/20022</guid><dc:creator><![CDATA[williamlouis]]></dc:creator><pubDate>Tue, 22 Sep 2026 07:50:19 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 07:15:19 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/geekyang" aria-label="Profile: Geekyang">@<bdi>Geekyang</bdi></a> 应该用的还是radiance，MTP速度慢，用的Dflash快一些</p>
]]></description><link>https://lcz.me/post/20004</link><guid isPermaLink="true">https://lcz.me/post/20004</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Tue, 22 Sep 2026 07:15:19 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 07:02:20 GMT]]></title><description><![CDATA[<p dir="auto">几点拆开说：</p>
<p dir="auto">1）「显卡非直连」通常指第二张卡不走 CPU 直连 lanes，而是经 PCH 或拆分后的槽位；但只要两张卡能被同一进程枚举、能建 RCCL 通信组，vLLM TP=2 就能跑（没有 P2P 时退回 host-staged/SHM）。batch=1 的 decode 由权重带宽主导，PCIe gen4 x8 和 x16 差别很小，所以单并发下「非直连不影响」基本成立。</p>
<p dir="auto">2）NVFP4 150 / MXFP4 120 / FP8 54 的量级方向合理：NVFP4 每权重约 4.5 bit，MXFP4 约 4.25 bit（但 scale 是 2 的幂、粒度更粗），FP8 约 8 bit 还要额外 scale。真要对比得固定同一 prompt、同一 KV dtype、同一上下文长度，别混变量。</p>
<p dir="auto">3）质量上 NVFP4 的块级 scale 一般比 MXFP4 更贴 FP8，但长上下文、推理、代码任务仍要自己 A/B；体感「粗糙」很多时候是 kernel 或 KV 精度造成的，不全是权重。</p>
<p dir="auto">4）确认 TP 是否真走 P2P：rocm-smi --showtopo 看链路，RCCL_DEBUG=INFO 看是否 host-staged，再跑 rccl-tests 的 bus bandwidth；pp 和 tg 分开测。</p>
]]></description><link>https://lcz.me/post/20001</link><guid isPermaLink="true">https://lcz.me/post/20001</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Tue, 22 Sep 2026 07:02:20 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 06:29:08 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/brian" aria-label="Profile: Brian">@<bdi>Brian</bdi></a> 双R9700 vllm FP8 54t/s 用的哪个github 方案？加投机解码了嘛？</p>
]]></description><link>https://lcz.me/post/19997</link><guid isPermaLink="true">https://lcz.me/post/19997</guid><dc:creator><![CDATA[Geekyang]]></dc:creator><pubDate>Tue, 22 Sep 2026 06:29:08 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 05:51:43 GMT]]></title><description><![CDATA[<p dir="auto">用了FP8，确实质量好一些</p>
]]></description><link>https://lcz.me/post/19992</link><guid isPermaLink="true">https://lcz.me/post/19992</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Tue, 22 Sep 2026 05:51:43 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 05:49:35 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> 速度快感觉有点粗糙，也不知道是不是心理作用，NVFP4据说跟原版差别不太大，比MXFP4好一些，但是也感觉不出来</p>
]]></description><link>https://lcz.me/post/19991</link><guid isPermaLink="true">https://lcz.me/post/19991</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Tue, 22 Sep 2026 05:49:35 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 05:48:29 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/geekyang" aria-label="Profile: Geekyang">@<bdi>Geekyang</bdi></a> 是相比Qwen3.8 27b FP8模型，之前TP2速度极慢，现在感觉vllm行了，FP8也有54t/s了（DSH中显示）</p>
]]></description><link>https://lcz.me/post/19990</link><guid isPermaLink="true">https://lcz.me/post/19990</guid><dc:creator><![CDATA[Brian]]></dc:creator><pubDate>Tue, 22 Sep 2026 05:48:29 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 05:38:57 GMT]]></title><description><![CDATA[<p dir="auto">如果是双卡的话，尽量还是跑高精度的，低精度我总是智力不够</p>
]]></description><link>https://lcz.me/post/19989</link><guid isPermaLink="true">https://lcz.me/post/19989</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Tue, 22 Sep 2026 05:38:57 GMT</pubDate></item><item><title><![CDATA[Reply to X99 AI Pro R9700 双卡 vLLM TP2 NVFP4 速度起飞，显卡非直连不影响 on Tue, 22 Sep 2026 05:24:31 GMT]]></title><description><![CDATA[<p dir="auto">显卡非直连 是什么意思？官方不就是给的MXFP4方案？</p>
]]></description><link>https://lcz.me/post/19986</link><guid isPermaLink="true">https://lcz.me/post/19986</guid><dc:creator><![CDATA[Geekyang]]></dc:creator><pubDate>Tue, 22 Sep 2026 05:24:31 GMT</pubDate></item></channel></rss>