<?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[V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next]]></title><description><![CDATA[<h2>前情提要 <a href="https://lcz.me/topic/2486/40"><strong>整台洋垃圾V100 16G雙卡</strong></a></h2>
<p dir="auto">之前只測 IQ3_S。但128G DDR3 雙通道帶寬不足容量有餘，挑戰Strata 的 4bit 選項看會不會吃鱉</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th style="text-align:right">下載</th>
<th style="text-align:right">專家 RAM</th>
<th style="text-align:right">全入 RAM 需</th>
<th style="text-align:right">品質（vs Q8_0 同 top token）</th>
<th>備註</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>UD-IQ4_XS</strong>（Unsloth）</td>
<td style="text-align:right">94 GB</td>
<td style="text-align:right">55 GiB</td>
<td style="text-align:right">~72 GB</td>
<td style="text-align:right">88.9%</td>
<td>比<strong>現行 IQ3_S（87.7%）</strong> 只高一點</td>
</tr>
<tr>
<td><strong>UD-Q4_K_XL</strong>（Unsloth）</td>
<td style="text-align:right">111 GB</td>
<td style="text-align:right">72 GiB</td>
<td style="text-align:right">~88 GB</td>
<td style="text-align:right">93.1%</td>
<td>4-bit 中品質最好</td>
</tr>
<tr>
<td><strong>NVFP4</strong>（Strata GPTQ 版）</td>
<td style="text-align:right">128 GB</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">92.09%</td>
<td>V100 無 FP4 tensor core，走 8-bit 精度路徑</td>
</tr>
<tr>
<td><strong>NVFP4-NVIDIA</strong></td>
<td style="text-align:right">128 GB</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">—</td>
<td>純 round，品質低於上面</td>
</tr>
<tr>
<td><strong>NVFP4-QAD</strong>（NVIDIA 重訓）</td>
<td style="text-align:right">102 GB +110 GB 盤</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">~80 GB</td>
<td style="text-align:right">—</td>
<td>短 prompt 略快；要在本機轉 pack（20–40 分鐘）</td>
</tr>
</tbody>
</table>
<h2>玩個騷一點的 Volta裝逼成Blackwell挑戰NVFP4</h2>
<p dir="auto">為啥能玩?有鑑於ninfer推論引擎之前社群就可以在V100 32G跑起NVFP4的27B模型了<br />
<strong>NVFP4 在 V100 上跑得起來，而且不需要 FP8/FP4 硬體</strong>——NVFP4 的路徑根本不用：</p>
<h2>NVFP4 （V100 sm_70 + E5-2686 v4）的支援狀況</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>路徑</th>
<th>V100 支援</th>
<th>機制</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Decode</strong></td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /></td>
<td>引擎自己的 q8_1 kernel <strong>軟體解碼</strong> UE4M3/E2M1，無 CUDA 專屬 intrinsics，任何架構都行</td>
</tr>
<tr>
<td><strong>CPU 專家</strong></td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /></td>
<td>E5 v4 明確在清單：走 <code>nvfp4_avx2.cpp</code> 的 256-bit AVX2 lane，與 ggml-cpu 逐位元相等（無 AVX-512 只慢 ~10%/core，但 pool 是 DRAM-bound，實際無差）</td>
</tr>
<tr>
<td><strong>Prefill（預設 w4a8）</strong></td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /> 但走 dp4a</td>
<td>source 註解（prefill.cpp:1147）：「Volta (sm_70): MMQ runs on dp4a there」——int8 tensor core 的 MMQ kernel 用 ldmatrix（sm_75+），V100 改走 dp4a 路徑（V100 有 dp4a）</td>
</tr>
<tr>
<td><strong>Prefill（<code>STRATA_PF_WMMA=1</code>）</strong></td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /> opt-in</td>
<td>Volta 專屬優化：專家層改用 <strong>FP16 tensor core tiles</strong>（V100 最強的就是 FP16 TC：125 TFLOPS）；代價是 activations 變 FP16，輸出與預設路徑略有差異</td>
</tr>
<tr>
<td><code>--peer-device</code> prefill</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/274c.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--x" style="height:1.4em;width:auto;vertical-align:middle" title="❌" alt="❌" /></td>
<td>「not for NVFP4 experts yet」（我們用 layer split，不影響）</td>
</tr>
</tbody>
</table>
<h2>NVFP4（Strata GPTQ 版）雙卡實測結果</h2>
<h2>硬體</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>項目</th>
<th>規格</th>
</tr>
</thead>
<tbody>
<tr>
<td>CPU</td>
<td>Intel Xeon E5-2686 v4 @ 2.3GHz（18C/36T，單路，max 3.0GHz，AVX2）</td>
</tr>
<tr>
<td>RAM</td>
<td>121GB</td>
</tr>
<tr>
<td>GPU</td>
<td>2× Tesla V100-SXM2-16GB（PHB 橋接，<strong>無 NVLink</strong>，PCIe 實測 12.0 GB/s）</td>
</tr>
<tr>
<td>顯示驅動</td>
<td>NVIDIA 580.173.02</td>
</tr>
<tr>
<td>磁碟</td>
<td>Samsung PM9A1 NVMe 1TB</td>
</tr>
</tbody>
</table>
<h2>軟體</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>項目</th>
<th>版本/狀態</th>
</tr>
</thead>
<tbody>
<tr>
<td>OS</td>
<td>Ubuntu 26.04.1 LTS，kernel 7.0.0-30-generic</td>
</tr>
<tr>
<td>CUDA</td>
<td>12.8（<code>/usr/local/cuda-12.8</code>）</td>
</tr>
<tr>
<td>引擎</td>
<td>strata 0.1.44.1</td>
</tr>
</tbody>
</table>
<h2>持久化調優</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>項目</th>
<th>設定</th>
</tr>
</thead>
<tbody>
<tr>
<td>Hugepages</td>
<td><code>vm.nr_hugepages=32768</code>（64GB，<code>/etc/sysctl.d/99-hugepages.conf</code>）</td>
</tr>
<tr>
<td>CPU governor</td>
<td><code>performance</code>（<code>cpufreq-performance.service</code>，剛補 enabled）</td>
</tr>
</tbody>
</table>
<h2>模型配置（現行 NVFP4 調優版）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>項目</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>模型</td>
<td>Qwen3.8-Flash-Next NVFP4-GPTQ（~120GB：experts 64G + PLE-FP8 48G + dense 8G）</td>
</tr>
<tr>
<td>分卡</td>
<td><code>layer_split 24</code>（手動，auto 的 23 較差）</td>
</tr>
<tr>
<td>Expert cache</td>
<td>auto → 3779/3680 slots/卡（hit rate ~87–90%）</td>
</tr>
<tr>
<td>Prefill</td>
<td><code>auto:16384</code>（chunk 16384）</td>
</tr>
<tr>
<td>Spec/MTP</td>
<td><code>--spec 4 --spec-min-p 0.5</code></td>
</tr>
<tr>
<td>KV</td>
<td>int8，<code>--kv-resident 20480</code>（最低值，省 VRAM 給 expert cache）</td>
</tr>
<tr>
<td>其他</td>
<td><code>--trim-stage-weights</code>、<code>--pipeline-windows 2</code>、<code>STRATA_ROUTE_TAIL_SKIP=7</code>、auto pcie_frac 0.15、pool 17 workers</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>效能現況</strong>：NVFP4 調優後 decode 58–65 tok/s、prefill 100k 1788 tok/s；IQ3_S decode 73–79 tok/s、prefill 100k 1884 tok/s。</p>
<h3>Prefill 階梯（冷預填）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th style="text-align:right">階段</th>
<th style="text-align:right">IQ3_S</th>
<th style="text-align:right">NVFP4</th>
<th style="text-align:right">Δ</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align:right">10k</td>
<td style="text-align:right">1181</td>
<td style="text-align:right">1009</td>
<td style="text-align:right">−15%</td>
</tr>
<tr>
<td style="text-align:right">20k</td>
<td style="text-align:right">1251</td>
<td style="text-align:right">1079</td>
<td style="text-align:right">−14%</td>
</tr>
<tr>
<td style="text-align:right">50k</td>
<td style="text-align:right">1876</td>
<td style="text-align:right">1633</td>
<td style="text-align:right">−13%</td>
</tr>
<tr>
<td style="text-align:right">100k</td>
<td style="text-align:right">1884</td>
<td style="text-align:right">1788</td>
<td style="text-align:right">−5%</td>
</tr>
<tr>
<td style="text-align:right">150k</td>
<td style="text-align:right">1792</td>
<td style="text-align:right">1761</td>
<td style="text-align:right">−2%</td>
</tr>
<tr>
<td style="text-align:right">200k</td>
<td style="text-align:right">1666</td>
<td style="text-align:right">1640</td>
<td style="text-align:right">−2%</td>
</tr>
</tbody>
</table>
<h3>Decode（中位數）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th style="text-align:right">prompt</th>
<th style="text-align:right">IQ3_S</th>
<th style="text-align:right">NVFP4</th>
<th style="text-align:right">Δ</th>
</tr>
</thead>
<tbody>
<tr>
<td style="text-align:right">essay</td>
<td style="text-align:right">73.5</td>
<td style="text-align:right">58.4</td>
<td style="text-align:right"><strong>−20%</strong></td>
</tr>
<tr>
<td style="text-align:right">code</td>
<td style="text-align:right">77.9</td>
<td style="text-align:right">58.6</td>
<td style="text-align:right"><strong>−28%</strong></td>
</tr>
<tr>
<td style="text-align:right">translate</td>
<td style="text-align:right">79.6</td>
<td style="text-align:right">62.3</td>
<td style="text-align:right"><strong>−22%</strong></td>
</tr>
<tr>
<td style="text-align:right">explain</td>
<td style="text-align:right">78.1</td>
<td style="text-align:right">64.6</td>
<td style="text-align:right"><strong>−18%</strong></td>
</tr>
</tbody>
</table>
<h3>關鍵發現</h3>
<ol>
<li><strong>Prefill 長上下文差距縮小</strong>（200k 只慢 2%），但短上下文慢 13–15%。</li>
<li>品質代價換算：NVFP4 92.1% vs IQ3_S 87.7% 同 top token as Q8_0——品質好 ~4.4 個百分點，但 decode 18~30%。</li>
<li>4bit模型體積比較大，導致能調度至VRAM的熱專家變少，較大比例落到DDR3雙通道內存與PCIE 3.0交換</li>
</ol>
<p dir="auto">如果認真要玩agent 建議QFN要到4bit的量化水平，在任務中tools與指令寫錯重發的機會明顯降低很多，<br />
不要小看那4.4%的品質提升，在長鏈任務中那個蝴蝶效應是很可怕的</p>
]]></description><link>https://lcz.me/topic/2512</link><generator>RSS for Node</generator><lastBuildDate>Sat, 24 Oct 2026 06:17:15 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2512.rss" rel="self" type="application/rss+xml"/><pubDate>Thu, 22 Oct 2026 20:08:07 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Sat, 24 Oct 2026 03:57:57 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/dardeaw-feng" aria-label="Profile: dardeaw-feng">@<bdi>dardeaw-feng</bdi></a> 原來是這樣，我是雙卡v100 32G, 是跑UD_K-Q4-XL</p>
]]></description><link>https://lcz.me/post/27268</link><guid isPermaLink="true">https://lcz.me/post/27268</guid><dc:creator><![CDATA[soop ladios]]></dc:creator><pubDate>Sat, 24 Oct 2026 03:57:57 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Sat, 24 Oct 2026 03:23:08 GMT]]></title><description><![CDATA[<p dir="auto">后悔当时没买32G的 V100 只搞了个16G的，买少了</p>
]]></description><link>https://lcz.me/post/27262</link><guid isPermaLink="true">https://lcz.me/post/27262</guid><dc:creator><![CDATA[Lei Gao]]></dc:creator><pubDate>Sat, 24 Oct 2026 03:23:08 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Sat, 24 Oct 2026 03:04:47 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/soop-ladios" aria-label="Profile: soop-ladios">@<bdi>soop-ladios</bdi></a> 對 但能省8G的RAM會快一點 然後讓Volta逼格比較高<br />
你的問題跟過去對Q4 vs NVFP4的問號都是一樣的，若你是50系顯卡除了空間省一點還能利用FP4加速讓prefill速度更快</p>
]]></description><link>https://lcz.me/post/27261</link><guid isPermaLink="true">https://lcz.me/post/27261</guid><dc:creator><![CDATA[dardeaw feng]]></dc:creator><pubDate>Sat, 24 Oct 2026 03:04:47 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Sat, 24 Oct 2026 02:32:58 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/philipshi2026" aria-label="Profile: PhilipShi2026">@<bdi>PhilipShi2026</bdi></a> 128G內存 但Agent偵測每次都會少 可能他沒把基本開銷算進去<br />
ComfyUI畫圖可以 但生影片最好買單卡32G的版本</p>
]]></description><link>https://lcz.me/post/27259</link><guid isPermaLink="true">https://lcz.me/post/27259</guid><dc:creator><![CDATA[dardeaw feng]]></dc:creator><pubDate>Sat, 24 Oct 2026 02:32:58 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Sat, 24 Oct 2026 01:11:46 GMT]]></title><description><![CDATA[<p dir="auto">1、121GB RAM感觉内存组合不出来，写错了？<br />
2、这个配置当前容易实现的配置，烦请测试下ComfyUI的情况如何？<br />
<a class="plugin-mentions-user plugin-mentions-a" href="/user/dardeaw-feng" aria-label="Profile: dardeaw-feng">@<bdi>dardeaw-feng</bdi></a></p>
]]></description><link>https://lcz.me/post/27250</link><guid isPermaLink="true">https://lcz.me/post/27250</guid><dc:creator><![CDATA[PhilipShi2026]]></dc:creator><pubDate>Sat, 24 Oct 2026 01:11:46 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Fri, 23 Oct 2026 12:00:29 GMT]]></title><description><![CDATA[<table class="table table-bordered table-striped">
<tbody>
<tr>
<td style="text-align:left">UD-Q4_K_XL（Unsloth）</td>
<td style="text-align:left">111 GB</td>
<td style="text-align:left">72 GiB</td>
<td>~88 GB</td>
<td>93.1%</td>
<td>4-bit 中品質最好</td>
</tr>
<tr>
<td style="text-align:left">NVFP4（Strata GPTQ 版）</td>
<td style="text-align:left">128 GB</td>
<td style="text-align:left">~80 GB</td>
<td>~80 GB</td>
<td>92.09%</td>
<td>V100 無 FP4 tensor core，走 8-bit 精度路徑</td>
</tr>
</tbody>
</table>
<p dir="auto">請問照這個表, 不是UD-Q4_K_XL品質比較好嗎?為什麼選NVFP4?</p>
]]></description><link>https://lcz.me/post/27033</link><guid isPermaLink="true">https://lcz.me/post/27033</guid><dc:creator><![CDATA[soop ladios]]></dc:creator><pubDate>Fri, 23 Oct 2026 12:00:29 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Fri, 23 Oct 2026 04:40:31 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/gao-gao" aria-label="Profile: gao-Gao">@<bdi>gao-Gao</bdi></a><br />
batch slot session 吃 VRAM，這個會讓速度稍微掉一點，好處是切換對話時秒切</p>
]]></description><link>https://lcz.me/post/26901</link><guid isPermaLink="true">https://lcz.me/post/26901</guid><dc:creator><![CDATA[dardeaw feng]]></dc:creator><pubDate>Fri, 23 Oct 2026 04:40:31 GMT</pubDate></item><item><title><![CDATA[Reply to V100以NVFP4量化運行Strata實測 Qwen 3.8 Flash Next on Fri, 23 Oct 2026 03:52:50 GMT]]></title><description><![CDATA[<p dir="auto">如果内存空余多,是否可以扩大下面的设置,我个人体感超过10GB缓存对于某种情况下 Thinking速度会变快 当前配置文件的 args 里设置了这两个参数：<br />
--conversation-cache-mib 4096<br />
--conversation-cache-slots 2<br />
意思是：Strata 最多为 2 个暂时切开的对话保留状态，缓存总量上限 4096 MiB（4 GiB）。这是总上限，不是每个槽各 4 GiB，也不代表同时运行 2 个请求。<br />
缓存不会启动时一次性占满；在不同对话之间切换时才可能逐步使用 RAM。复用还要求提示词前缀及图片、推理模式等匹配。修改这两个值后，需要重启模型才会生效。</p>
]]></description><link>https://lcz.me/post/26895</link><guid isPermaLink="true">https://lcz.me/post/26895</guid><dc:creator><![CDATA[gao Gao]]></dc:creator><pubDate>Fri, 23 Oct 2026 03:52:50 GMT</pubDate></item></channel></rss>