<?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[满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4]]></title><description><![CDATA[<p dir="auto">论坛上已经有一帖分享 V4.1 flash 配方（<a href="https://lcz.me/topic/1658">Deepseek v4.1 flash , 4台DGX spark GB10, TP=4</a>）基本默认要一台 200G 交换机做四机全互联。我这套是用 <strong>4 根 DAC 线把四台 GB10 直接接成一个闭环</strong>——没有交换机，对面节点（rank0<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2194.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--left_right_arrow" style="height:1.4em;width:auto;vertical-align:middle" title="↔" alt="↔" />rank2）互不可直连，通信要靠邻居转发——走的是 SparkRing 的无交换机环（switchless cycle）路线，NCCL 需要打补丁才肯在这种拓扑上起来。</p>
<p dir="auto">实测结果是可以用，而且不比交换机组网差：单流解码 75–108 tok/s，满并发（8 路）聚合解码 280→703 tok/s，16K 上下文 prefill ≈4.4k tok/s，1M 上下文单流 TTFT 71.6s / prefill 3.58k tok/s。今天把<strong>并发 × 上下文</strong>的二维矩阵整跑了一遍，数据附在下面。</p>
<p dir="auto"><strong>使用方案</strong>：<a href="https://github.com/FujitsuPolycom/sparkring" rel="nofollow ugc">FujitsuPolycom/sparkring</a>（Apache-2.0，自带 NCCL 补丁）的 <code>deepseek-v41-flash-tp4</code> profile，模型是官方原始权重。</p>
<hr />
<h1>DeepSeek-V4.1-Flash TP4（四机环）部署与测试总结报告</h1>
<hr />
<h2>一、模型与服务基本信息</h2>
<ul>
<li><strong>模型名称</strong>：<code>DeepSeek-V4.1-Flash</code>（服务名 <code>DeepSeek-V4.1-Flash-TP4</code>）</li>
<li><strong>权重版本</strong>：官方 checkpoint，pinned revision <code>dba1be0a40aa</code>，48 个 safetensors 分片，<strong>476 GiB</strong></li>
<li><strong>上游原生路线</strong>：<code>github.com/FujitsuPolycom/sparkring</code> — profile <code>deepseek-v41-flash-tp4</code>（<code>recommendation: recommended</code>，装机器镜像 <code>dev-20261013-dcp2prefill-status038-sircl-tp6</code>）</li>
<li><strong>推理引擎</strong>：vLLM（SparkRing "Kraken" 分支 <code>0.1.dev21553+gab86b7073</code>）+ B12X 内核 + SIRCL 传输插件</li>
<li><strong>硬件环境</strong>：4× NVIDIA DGX Spark（GB10，aarch64，单机 121 GiB 统一内存）
<ul>
<li>rank0 <code>aitopatom-aca1</code>（技嘉 AI TOP ATOM，Node A / 服务出口）｜ rank1 <code>gx10-white</code>（ASUS GX10）｜ rank2 <code>gx10-black</code>（ASUS GX10）｜ rank3 <code>aitopatom-a9b6</code>（技嘉 AI TOP ATOM）</li>
<li>两家 OEM 的系统版本不同（GX10 = DGX OS 7.2.3 / 内核 6.17.0-1031-nvidia；AI TOP ATOM = DGX OS 7.4.0 / 内核 7.0.0-1019-nvidia），实测混装可以跑通</li>
</ul>
</li>
<li><strong>高速网络互连</strong>：<strong>无交换机环形拓扑</strong>，4 根 QSFP DAC 直连（p0 → 下一台 p1，无对角线），每台 2 个 CX7 200G 端口（每个端口两个 Socket Direct 功能，共 4 个 fabric netdev），整环 8 条逻辑通道
<ul>
<li>环序：<code>aitopatom-aca1 → gx10-white → gx10-black → aitopatom-a9b6 →（回到 aca1）</code></li>
<li>fabric 地址段 <code>198.18.0.0/21</code>，每根缆两个 /24；控制面走 WireGuard <code>sr-control</code>（<code>10.253.255.1–4</code>）</li>
</ul>
</li>
</ul>
<blockquote>
<p dir="auto"><strong>权重存储与挂载</strong>：环上没有贯通全环的 NFS 通路（对面节点不直连），所以<strong>四台各存一份完整 476 GiB 权重</strong>，本机 NVMe 直读；四机 × 48 分片全量 sha256 比对一致（192/192）。这跟交换机组网"HEAD 挂 NFS 给 worker 只读"的做法不同，是环拓扑的硬约束，也是安装前要预留 ~620 GB/台 空间的原因。</p>
</blockquote>
<hr />
<h2>二、核心规格与配置参数</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>核心参数</th>
<th>数值 / 设定</th>
<th>说明</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Context Length</strong></td>
<td><strong>1,048,576</strong>（1M tokens）</td>
<td>模型原生完整长度，本次部署跑满 1M</td>
</tr>
<tr>
<td><strong>Max Running Requests</strong></td>
<td><strong>8</strong></td>
<td><code>--max-num-seqs 8</code>，CUDA Graph 捕获尺寸覆盖到 96</td>
</tr>
<tr>
<td><strong>KV Cache Pool</strong></td>
<td><strong>12,715,913 tokens</strong>（可用 KV 显存 15.89 GiB）</td>
<td>1M 上下文时可支撑 <strong>12.13 路</strong>并发；FP8</td>
</tr>
<tr>
<td><strong>静态显存比例</strong></td>
<td><code>0.83</code>（≈100.9 GiB / 121.6 GiB）</td>
<td>每 rank 权重常驻 <strong>74.4 GiB</strong>，服务中每机仍余 9–10 GiB</td>
</tr>
<tr>
<td><strong>分块预填充</strong></td>
<td><code>--max-num-batched-tokens 8192</code> + chunked prefill</td>
<td>抑制长上下文 attention logits 峰值张量，500K+ 不 OOM</td>
</tr>
<tr>
<td><strong>前缀缓存 / 调度</strong></td>
<td>prefix caching + async scheduling</td>
<td>同一前缀重复请求命中即省 prefill</td>
</tr>
<tr>
<td><strong>推测解码</strong></td>
<td><strong>DSpark</strong>，<code>num_speculative_tokens 5</code>、概率式草稿、块级拒绝、自适应验证</td>
<td>引擎实测：平均接受长度 <strong>3.65</strong>、平均草稿接受率 <strong>61.0%</strong>、逐位接受率 0.851 / 0.684 / 0.546 / 0.463 / 0.392</td>
</tr>
<tr>
<td><strong>Engram 卸载</strong></td>
<td><code>table_memory=disk</code> + <code>disk_resident_scales</code></td>
<td>两张 FP8 n-gram 表（各 ~94.6 GiB）不占显存，按 token 需要从本机 NVMe 直读——这是 4 台能装下 1M 上下文的关键</td>
</tr>
<tr>
<td><strong>环网传输</strong></td>
<td><strong>SIRCL</strong>（<code>SIRCL_FABRIC=ring:4</code>、<code>SIRCL_GROUPS=tp</code>、NCCL never）</td>
<td>补齐的 NCCL 2.32.3 提供 <code>NCCL_SWITCHLESS_RING_ONLY=1</code>，跳过 tree/PAT 建链（环上对面 rank 无直连）</td>
</tr>
<tr>
<td><strong>多模态视觉</strong></td>
<td>原生支持，<code>limit-mm-per-prompt {"image":4}</code></td>
<td>256×128 测试图消耗 202 vision tokens</td>
</tr>
<tr>
<td><strong>思考控制</strong></td>
<td>默认 <strong>on · high</strong>；级别 <code>low / high / xhigh(mx) / 1–100</code></td>
<td>可请求级关闭（<code>chat_template_kwargs: {"thinking": false}</code>）；默认开启且思考链吃 token，<code>max_tokens</code> 给小了会返回空 <code>content</code></td>
</tr>
<tr>
<td><strong>工具 / 推理解析器</strong></td>
<td><code>deepseek_v41</code>（tool-call + reasoning）+ auto tool choice</td>
<td>支持强制工具调用</td>
</tr>
<tr>
<td><strong>端点</strong></td>
<td><code>http://192.168.10.10:8015/v1</code></td>
<td><strong>无鉴权</strong>（不带 key 与带假 key 都返回 200）：方便接线，但别把它暴露到局域网之外</td>
</tr>
<tr>
<td><strong>启动耗时</strong></td>
<td>热重启 <strong>4 分 29 秒</strong>（22:58:47 起容器 → 23:03:16 API 就绪）</td>
<td>权重 50.4 s 载完（74.4 GiB/rank），引擎 init 150 s；首次冷装（含校验 + 编译内核）约 20 分钟</td>
</tr>
</tbody>
</table>
<hr />
<h2>三、验证与测试数据</h2>
<p dir="auto">测试口径（全部一致）：</p>
<ul>
<li>每档输出 <strong>128 tokens</strong>，填充文本带<strong>唯一 nonce + 递增编号</strong> → 不命中 prefix 缓存；温度 0.6</li>
<li><strong>单流 pp/tg</strong> = 各请求均值；<strong>聚合 pp</strong> = Σprompt / max(TTFT)；<strong>聚合 tg</strong> = Σcompletion / (max 总时长 − max TTFT)</li>
<li>并发 1/2/4/8（8 = 本部署的并发上限），上下文档位 1K → 512K</li>
</ul>
<h3>1. 基础功能与能力验证（栈自检，本机环上实跑）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>测试项</th>
<th>输入</th>
<th>结果</th>
<th>状态</th>
</tr>
</thead>
<tbody>
<tr>
<td>计数</td>
<td><code>1…20</code> 顺序输出</td>
<td>全对</td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>算术</td>
<td><code>17 × 23</code></td>
<td><code>391</code></td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>代码生成</td>
<td>Python <code>is_prime(n)</code></td>
<td>完整可用实现</td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>工具调用</td>
<td><code>get_weather(Paris)</code></td>
<td>正确发起 tool call</td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>强制工具调用</td>
<td><code>tool_choice=required</code></td>
<td>正确发起 <code>get_weather(Rome)</code></td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>原生视觉</td>
<td>左红右蓝 128×64 PNG</td>
<td><code>The left half is red and the right half is blue.</code></td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>思考开关</td>
<td>关→直答 / 开→带思考链</td>
<td>thinking on 时 <code>391</code> + 62 字符思考链</td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>传输层</td>
<td>本次集合通信走哪条路</td>
<td><code>Transport: sircl, NCCL: absent</code></td>
<td><strong>PASS</strong></td>
</tr>
<tr>
<td>频率</td>
<td>四机 SM 时钟</td>
<td>2483–2522 MHz，无降频事件（仅 idle）</td>
<td><strong>PASS</strong></td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">四机环上自检全绿：<code>All checks passed.</code></p>
</blockquote>
<h3>2. 长上下文阶梯压测（并发 1）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>目标上下文</th>
<th>实际 prompt tokens</th>
<th>TTFT</th>
<th>单流 prefill</th>
<th>单流解码</th>
<th>墙钟</th>
<th>状态</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>1,116</td>
<td>0.47s</td>
<td>2,362.7 tok/s</td>
<td><strong>91.0 tok/s</strong></td>
<td>1.9s</td>
<td>PASS</td>
</tr>
<tr>
<td>4K</td>
<td>4,070</td>
<td>0.99s</td>
<td>4,118.6 tok/s</td>
<td><strong>96.5 tok/s</strong></td>
<td>2.3s</td>
<td>PASS</td>
</tr>
<tr>
<td>8K</td>
<td>7,958</td>
<td>1.76s</td>
<td>4,529.1 tok/s</td>
<td>89.4 tok/s</td>
<td>3.2s</td>
<td>PASS</td>
</tr>
<tr>
<td>16K</td>
<td>15,686</td>
<td>3.43s</td>
<td>4,569.8 tok/s</td>
<td>75.1 tok/s</td>
<td>5.1s</td>
<td>PASS</td>
</tr>
<tr>
<td>32K</td>
<td>31,741</td>
<td>7.93s</td>
<td>4,001.6 tok/s</td>
<td>74.9 tok/s</td>
<td>9.7s</td>
<td>PASS</td>
</tr>
<tr>
<td>64K</td>
<td>63,524</td>
<td>14.35s</td>
<td>4,425.9 tok/s</td>
<td>90.7 tok/s</td>
<td>15.8s</td>
<td>PASS</td>
</tr>
<tr>
<td>128K</td>
<td>127,784</td>
<td>32.14s</td>
<td>3,976.0 tok/s</td>
<td>86.7 tok/s</td>
<td>33.6s</td>
<td>PASS</td>
</tr>
<tr>
<td>256K</td>
<td>256,355</td>
<td>71.61s</td>
<td>3,579.8 tok/s</td>
<td><strong>107.9 tok/s</strong></td>
<td>72.8s</td>
<td>PASS</td>
</tr>
<tr>
<td><strong>512K</strong></td>
<td>513,344</td>
<td><strong>175.55s</strong></td>
<td>2,924.2 tok/s</td>
<td>79.0 tok/s</td>
<td>177.2s</td>
<td>PASS</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">1M 上下文是配置上限，实跑验证到 512K：<strong>512K 单请求 TTFT 175.6s、prefill 2.92k tok/s，全程无 OOM、无换页</strong>。</p>
</blockquote>
<h3>3. 高并发 × 上下文 二维扫描</h3>
<h4>并发 2</h4>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>prompt/请求</th>
<th>TTFT(max)</th>
<th>单流 pp</th>
<th>单流 tg</th>
<th>聚合 pp</th>
<th>聚合 tg</th>
<th>墙钟</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>1,094</td>
<td>0.64s</td>
<td>1,723.3</td>
<td>55.2</td>
<td>3,443.3</td>
<td>110.4</td>
<td>3.0s</td>
</tr>
<tr>
<td>4K</td>
<td>4,154</td>
<td>2.02s</td>
<td>3,029.8</td>
<td>60.3</td>
<td>4,067.2</td>
<td>143.6</td>
<td>3.8s</td>
</tr>
<tr>
<td>8K</td>
<td>7,958</td>
<td>3.52s</td>
<td>3,394.1</td>
<td>52.4</td>
<td>4,525.3</td>
<td>135.4</td>
<td>5.4s</td>
</tr>
<tr>
<td>16K</td>
<td>15,686</td>
<td>6.92s</td>
<td>3,421.6</td>
<td>44.3</td>
<td>4,579.8</td>
<td>127.8</td>
<td>8.9s</td>
</tr>
<tr>
<td>32K</td>
<td>31,094</td>
<td>14.02s</td>
<td>3,239.3</td>
<td>34.1</td>
<td>4,481.6</td>
<td>107.5</td>
<td>16.4s</td>
</tr>
<tr>
<td>64K</td>
<td>62,240</td>
<td>28.28s</td>
<td>3,204.5</td>
<td>42.7</td>
<td>4,401.8</td>
<td>153.7</td>
<td>30.0s</td>
</tr>
<tr>
<td>128K</td>
<td>125,240</td>
<td>62.84s</td>
<td>2,955.3</td>
<td>36.5</td>
<td>4,026.3</td>
<td>137.8</td>
<td>64.7s</td>
</tr>
<tr>
<td>256K</td>
<td>256,355</td>
<td>146.02s</td>
<td>2,640.5</td>
<td>34.6</td>
<td>3,511.3</td>
<td>134.9</td>
<td>148.0s</td>
</tr>
<tr>
<td>512K</td>
<td>503,240</td>
<td>349.99s</td>
<td>2,185.7</td>
<td>43.7</td>
<td>2,904.6</td>
<td>171.8</td>
<td>351.6s</td>
</tr>
</tbody>
</table>
<h4>并发 4</h4>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>prompt/请求</th>
<th>TTFT(max)</th>
<th>单流 pp</th>
<th>单流 tg</th>
<th>聚合 pp</th>
<th>聚合 tg</th>
<th>墙钟</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>1,094</td>
<td>1.06s</td>
<td>1,036.2</td>
<td>50.5</td>
<td>4,136.7</td>
<td>189.3</td>
<td>3.8s</td>
</tr>
<tr>
<td>4K</td>
<td>4,070</td>
<td>3.65s</td>
<td>1,678.6</td>
<td>40.1</td>
<td>4,486.7</td>
<td>199.0</td>
<td>6.2s</td>
</tr>
<tr>
<td>8K</td>
<td>8,123</td>
<td>7.02s</td>
<td>2,321.6</td>
<td>30.1</td>
<td>4,580.0</td>
<td>193.3</td>
<td>9.7s</td>
</tr>
<tr>
<td>16K</td>
<td>16,012</td>
<td>14.87s</td>
<td>2,235.5</td>
<td>24.5</td>
<td>4,262.4</td>
<td>206.1</td>
<td>17.4s</td>
</tr>
<tr>
<td>32K</td>
<td>31,094</td>
<td>28.89s</td>
<td>2,208.7</td>
<td>18.7</td>
<td>4,327.3</td>
<td>185.8</td>
<td>31.7s</td>
</tr>
<tr>
<td>64K</td>
<td>63,524</td>
<td>59.54s</td>
<td>2,170.9</td>
<td>18.4</td>
<td>4,202.9</td>
<td>237.4</td>
<td>61.7s</td>
</tr>
<tr>
<td>128K</td>
<td>125,240</td>
<td>125.14s</td>
<td>2,023.9</td>
<td>20.9</td>
<td>4,023.6</td>
<td>302.8</td>
<td>126.9s</td>
</tr>
<tr>
<td>256K</td>
<td>251,290</td>
<td>288.52s</td>
<td>1,842.5</td>
<td>22.4</td>
<td>3,519.0</td>
<td>339.4</td>
<td>290.1s</td>
</tr>
<tr>
<td>512K</td>
<td>513,344</td>
<td>695.75s</td>
<td>1,516.5</td>
<td>24.8</td>
<td>2,922.3</td>
<td><strong>379.3</strong></td>
<td>697.3s</td>
</tr>
</tbody>
</table>
<h4>并发 8（满并发，= <code>max-num-seqs</code>）</h4>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>prompt/请求</th>
<th>TTFT(max)</th>
<th>单流 pp</th>
<th>单流 tg</th>
<th>聚合 pp</th>
<th>聚合 tg</th>
<th>墙钟</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>1,094</td>
<td>2.27s</td>
<td>883.7</td>
<td>35.0</td>
<td>3,888.2</td>
<td>280.7</td>
<td>5.9s</td>
</tr>
<tr>
<td>4K</td>
<td>4,070</td>
<td>7.30s</td>
<td>1,245.4</td>
<td>25.3</td>
<td>4,503.8</td>
<td>256.9</td>
<td>11.3s</td>
</tr>
<tr>
<td>8K</td>
<td>8,123</td>
<td>14.24s</td>
<td>1,512.7</td>
<td>16.9</td>
<td>4,565.0</td>
<td>268.7</td>
<td>18.1s</td>
</tr>
<tr>
<td>16K</td>
<td>16,012</td>
<td>28.80s</td>
<td>1,519.1</td>
<td>12.6</td>
<td>4,402.4</td>
<td>280.3</td>
<td>32.5s</td>
</tr>
<tr>
<td>32K</td>
<td>31,094</td>
<td>57.46s</td>
<td>1,450.2</td>
<td>10.4</td>
<td>4,351.9</td>
<td>335.6</td>
<td>60.5s</td>
</tr>
<tr>
<td>64K</td>
<td>62,240</td>
<td>117.68s</td>
<td>1,418.9</td>
<td>9.7</td>
<td>4,252.8</td>
<td>437.9</td>
<td>120.1s</td>
</tr>
<tr>
<td>128K</td>
<td>125,240</td>
<td>253.92s</td>
<td>1,340.2</td>
<td>10.8</td>
<td>3,965.8</td>
<td>576.3</td>
<td>255.8s</td>
</tr>
<tr>
<td>256K</td>
<td>256,355</td>
<td>571.94s</td>
<td>1,190.9</td>
<td>10.4</td>
<td>3,550.3</td>
<td>574.0</td>
<td>573.9s</td>
</tr>
<tr>
<td><strong>512K</strong></td>
<td>503,240</td>
<td><strong>1,396.56s</strong></td>
<td>982.6</td>
<td>12.3</td>
<td>2,904.4</td>
<td><strong>703.5</strong></td>
<td>1,398.4s</td>
</tr>
</tbody>
</table>
<h3>4. 汇总对照</h3>
<p dir="auto"><strong>聚合解码吞吐（tokens/s）</strong></p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>并发 1</th>
<th>并发 2</th>
<th>并发 4</th>
<th>并发 8</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>91.0</td>
<td>110.4</td>
<td>189.3</td>
<td>280.7</td>
</tr>
<tr>
<td>4K</td>
<td>96.5</td>
<td>143.6</td>
<td>199.0</td>
<td>256.9</td>
</tr>
<tr>
<td>8K</td>
<td>89.4</td>
<td>135.4</td>
<td>193.3</td>
<td>268.7</td>
</tr>
<tr>
<td>16K</td>
<td>75.1</td>
<td>127.8</td>
<td>206.1</td>
<td>280.3</td>
</tr>
<tr>
<td>32K</td>
<td>74.9</td>
<td>107.5</td>
<td>185.8</td>
<td>335.6</td>
</tr>
<tr>
<td>64K</td>
<td>90.7</td>
<td>153.7</td>
<td>237.4</td>
<td>437.9</td>
</tr>
<tr>
<td>128K</td>
<td>86.7</td>
<td>137.8</td>
<td>302.8</td>
<td>576.3</td>
</tr>
<tr>
<td>256K</td>
<td>107.9</td>
<td>134.9</td>
<td>339.4</td>
<td>574.0</td>
</tr>
<tr>
<td>512K</td>
<td>79.0</td>
<td>171.8</td>
<td>379.3</td>
<td><strong>703.5</strong></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>单流解码吞吐（tokens/s）</strong></p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>并发 1</th>
<th>并发 2</th>
<th>并发 4</th>
<th>并发 8</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>91.0</td>
<td>55.2</td>
<td>50.5</td>
<td>35.0</td>
</tr>
<tr>
<td>4K</td>
<td>96.5</td>
<td>60.3</td>
<td>40.1</td>
<td>25.3</td>
</tr>
<tr>
<td>8K</td>
<td>89.4</td>
<td>52.4</td>
<td>30.1</td>
<td>16.9</td>
</tr>
<tr>
<td>16K</td>
<td>75.1</td>
<td>44.3</td>
<td>24.5</td>
<td>12.6</td>
</tr>
<tr>
<td>32K</td>
<td>74.9</td>
<td>34.1</td>
<td>18.7</td>
<td>10.4</td>
</tr>
<tr>
<td>64K</td>
<td>90.7</td>
<td>42.7</td>
<td>18.4</td>
<td>9.7</td>
</tr>
<tr>
<td>128K</td>
<td>86.7</td>
<td>36.5</td>
<td>20.9</td>
<td>10.8</td>
</tr>
<tr>
<td>256K</td>
<td>107.9</td>
<td>34.6</td>
<td>22.4</td>
<td>10.4</td>
</tr>
<tr>
<td>512K</td>
<td>79.0</td>
<td>43.7</td>
<td>24.8</td>
<td>12.3</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>聚合 prefill 吞吐（tokens/s）</strong></p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>并发 1</th>
<th>并发 2</th>
<th>并发 4</th>
<th>并发 8</th>
</tr>
</thead>
<tbody>
<tr>
<td>1K</td>
<td>2,362.7</td>
<td>3,443.3</td>
<td>4,136.7</td>
<td>3,888.2</td>
</tr>
<tr>
<td>4K</td>
<td>4,118.6</td>
<td>4,067.2</td>
<td>4,486.7</td>
<td>4,503.8</td>
</tr>
<tr>
<td>8K</td>
<td>4,529.1</td>
<td>4,525.3</td>
<td>4,580.0</td>
<td>4,565.0</td>
</tr>
<tr>
<td>16K</td>
<td>4,569.8</td>
<td>4,579.8</td>
<td>4,262.4</td>
<td>4,402.4</td>
</tr>
<tr>
<td>32K</td>
<td>4,001.6</td>
<td>4,481.6</td>
<td>4,327.3</td>
<td>4,351.9</td>
</tr>
<tr>
<td>64K</td>
<td>4,425.9</td>
<td>4,401.8</td>
<td>4,202.9</td>
<td>4,252.8</td>
</tr>
<tr>
<td>128K</td>
<td>3,976.0</td>
<td>4,026.3</td>
<td>4,023.6</td>
<td>3,965.8</td>
</tr>
<tr>
<td>256K</td>
<td>3,579.8</td>
<td>3,511.3</td>
<td>3,519.0</td>
<td>3,550.3</td>
</tr>
<tr>
<td>512K</td>
<td>2,924.2</td>
<td>2,904.6</td>
<td>2,922.3</td>
<td>2,904.4</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>几条结论</strong>：</p>
<ul>
<li><strong>Prefill 与并发无关</strong>：1K→256K 区间聚合 prefill 基本贴在 <strong>4.0k–4.6k tok/s</strong>，四档并发几乎重合，直到 512K 才降到 <strong>2.9k</strong>。说明环拓扑的瓶颈不在 prefill 聚合带宽（8 条 200G 通道够用），而在单请求的长序列 attention。</li>
<li><strong>Decode 随并发放大</strong>：满并发 8 路时聚合解码从 1K 的 280.7 一路涨到 512K 的 <strong>703.5 tok/s</strong>——上下文越长、每步计算越"重"，批次填得越满，环上 all-reduce 的固定开销被摊得越薄。</li>
<li><strong>单流解码 ~75–108 tok/s</strong>（与换长上下文无关，1K–512K 都在这个带内）——比交换机组网公开数据的 58–68 tok/s 略高，主要差在推测解码接受率（平均接受长度 3.65 / 接受率 61%）。</li>
<li><strong>128 tokens 的短输出会低估真实吞吐</strong>：单流 tg 里含首几步的调度开销，实际长任务（几百到几千 token）会明显更好。</li>
</ul>
<h3>5. 与交换机组网公开数据的量级对照</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>上下文</th>
<th>本机（无交换机环，单流）</th>
<th>论坛交换机组网（单流，64 输出）</th>
</tr>
</thead>
<tbody>
<tr>
<td>32K</td>
<td>TTFT 7.93s / prefill 4,001.6 / 解码 74.9</td>
<td>TTFT 10.02s / prefill 3,271.0 / 解码 67.5</td>
</tr>
<tr>
<td>128K</td>
<td>TTFT 32.14s / prefill 3,976.0 / 解码 86.7</td>
<td>TTFT 48.46s / prefill 2,704.8 / 解码 67.5</td>
</tr>
<tr>
<td>256K</td>
<td>TTFT 71.61s / prefill 3,579.8 / 解码 107.9</td>
<td>TTFT 87.46s / prefill 2,997.3 / 解码 59.4</td>
</tr>
<tr>
<td>512K</td>
<td>TTFT 175.55s / prefill 2,924.2 / 解码 79.0</td>
<td>TTFT 268.94s / prefill 1,859.1 / 解码 58.2（500K）</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">口径不完全相同（输出长度 128 vs 64、软件栈不同），仅作量级参考：<strong>环拓扑在 TTFT / prefill 上明显更快，解码在一个带内偏上</strong>。"不用交换机就跑不动 TP=4"这个结论，至少在这套栈上可以推翻了。</p>
</blockquote>
<hr />
<p dir="auto"><img src="https://upload.lcz.me/uploads/6d9b0f7c-d487-4c70-8343-f4095eb8518f.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
<p dir="auto">除了手动更新机器、组网，环境全由hermes配置，耗时16小时，花销如图。</p>
]]></description><link>https://lcz.me/topic/2346</link><generator>RSS for Node</generator><lastBuildDate>Fri, 23 Oct 2026 01:18:24 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2346.rss" rel="self" type="application/rss+xml"/><pubDate>Thu, 15 Oct 2026 16:06:19 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Fri, 16 Oct 2026 17:11:10 GMT]]></title><description><![CDATA[<p dir="auto">你4个小盒子还编排什么，直接搞完。</p>
]]></description><link>https://lcz.me/post/25343</link><guid isPermaLink="true">https://lcz.me/post/25343</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Fri, 16 Oct 2026 17:11:10 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Fri, 16 Oct 2026 16:36:44 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> 对，就是这个意思，DS编排，布置任务，审查，qwen3.8执行。</p>
]]></description><link>https://lcz.me/post/25337</link><guid isPermaLink="true">https://lcz.me/post/25337</guid><dc:creator><![CDATA[benton yi]]></dc:creator><pubDate>Fri, 16 Oct 2026 16:36:44 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Fri, 16 Oct 2026 16:23:37 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/benton-yi" aria-label="Profile: benton-yi">@<bdi>benton-yi</bdi></a> 干活除非是框架已经好了，具体功能改进，Qwen27b不错，真的从头搞，比DS差远了。</p>
]]></description><link>https://lcz.me/post/25334</link><guid isPermaLink="true">https://lcz.me/post/25334</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Fri, 16 Oct 2026 16:23:37 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Fri, 16 Oct 2026 15:56:50 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/marck-wong" aria-label="Profile: Marck-Wong">@<bdi>Marck-Wong</bdi></a> DSV4.1的优势还是真1M的长上下文，编排智能体的较优选。如果真的要本地部署模型写代码，个人感觉是不是还是让qwen3.8家族的来干会好点。</p>
]]></description><link>https://lcz.me/post/25330</link><guid isPermaLink="true">https://lcz.me/post/25330</guid><dc:creator><![CDATA[benton yi]]></dc:creator><pubDate>Fri, 16 Oct 2026 15:56:50 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Fri, 16 Oct 2026 14:04:21 GMT]]></title><description><![CDATA[<p dir="auto">请教各位大佬 用Deepseek V4.1 Flash的时候 有没有觉得只是快 代码质量不行？Thinking开到了Max 都还是觉得干活好差？ 我用的是Deepseek API 在DSH里测试的</p>
]]></description><link>https://lcz.me/post/25310</link><guid isPermaLink="true">https://lcz.me/post/25310</guid><dc:creator><![CDATA[Marck Wong]]></dc:creator><pubDate>Fri, 16 Oct 2026 14:04:21 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Thu, 15 Oct 2026 16:18:11 GMT]]></title><description><![CDATA[<p dir="auto">可以，下次等架子到位，发一个</p>
]]></description><link>https://lcz.me/post/25138</link><guid isPermaLink="true">https://lcz.me/post/25138</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Thu, 15 Oct 2026 16:18:11 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Thu, 15 Oct 2026 16:16:16 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> <a href="/post/25128">说</a>:</p>
<p dir="auto">只是......价格坑爹了。</p>
</blockquote>
<p dir="auto">两台1T的华硕GX10是30500/台入手的：<br />
<img src="https://upload.lcz.me/uploads/a4737e97-46eb-49d8-810e-6139fd3262fe.jpg" alt="0ac41782272412797eb35146c8c8af3b.jpg" class="img-markdown  img-fluid" /></p>
<p dir="auto">两台4T的技嘉Ai Top Atom是国庆节前36000/台加急锁单：<br />
<img src="https://upload.lcz.me/uploads/c5d550f2-8d50-4ee0-90ad-2c81fb4fca17.jpg" alt="441b31da222392cf6b54404c8f180750.jpg" class="img-markdown  img-fluid" /></p>
<p dir="auto">找线上的玩家定制了小米空净器的打印散热框架，还没到货，组装完全体之后再发图吧</p>
]]></description><link>https://lcz.me/post/25135</link><guid isPermaLink="true">https://lcz.me/post/25135</guid><dc:creator><![CDATA[benton yi]]></dc:creator><pubDate>Thu, 15 Oct 2026 16:16:16 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Thu, 15 Oct 2026 16:09:47 GMT]]></title><description><![CDATA[<p dir="auto">记得发装备秀啊，给填充点资源<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f602.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--joy" style="height:1.4em;width:auto;vertical-align:middle" title="😂" alt="😂" /></p>
]]></description><link>https://lcz.me/post/25130</link><guid isPermaLink="true">https://lcz.me/post/25130</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Thu, 15 Oct 2026 16:09:47 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Thu, 15 Oct 2026 16:08:03 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2b50.png?v=ec85ec55bc1" class="not-responsive emoji emoji-android emoji--star" style="height:1.4em;width:auto;vertical-align:middle" title="⭐" alt="⭐" /> 版主审定：本帖设为<strong>精华</strong>，作者 <strong>+5 积分</strong>奖励，希望再接再厉！</p>
<blockquote>
<p dir="auto">奖励凭证：精华 +5 分 · 编号 f2346-18-1（系统自动发放，显示稍有延迟）</p>
</blockquote>
]]></description><link>https://lcz.me/post/25129</link><guid isPermaLink="true">https://lcz.me/post/25129</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Thu, 15 Oct 2026 16:08:03 GMT</pubDate></item><item><title><![CDATA[Reply to 满血DeepSeek-V4.1-Flash，4台DGX Spark GB10环形组网（无交换机）TP=4 on Thu, 15 Oct 2026 16:08:16 GMT]]></title><description><![CDATA[<p dir="auto">很好，虽迟但到，我哥是不折不扣的土豪，这效果真的是不错，DGX Spark的潜力还在提升！这下再也不用担心显存不够了，只是......价格坑爹了。</p>
]]></description><link>https://lcz.me/post/25128</link><guid isPermaLink="true">https://lcz.me/post/25128</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Thu, 15 Oct 2026 16:08:16 GMT</pubDate></item></channel></rss>