<?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[【实测】Thinkpad P1 Gen4-3080显卡15款大模型测试报告摘要]]></title><description><![CDATA[<p dir="auto">#Thinkpad P1 Gen4-3080显卡15款大模型测试报告摘要</p>
<h2>1 系统信息与测试环境</h2>
<h3>1.1 硬件平台</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>规格</th>
</tr>
</thead>
<tbody>
<tr>
<td>操作系统</td>
<td>Microsoft Windows 11  10.0.22631 64 λ</td>
</tr>
<tr>
<td>CPU</td>
<td>11th Gen Intel(R) Core(TM) i7-11850H @ 2.50GHz（8 核 / 16 线程）</td>
</tr>
<tr>
<td>内存</td>
<td>32 GB</td>
</tr>
<tr>
<td>显卡</td>
<td>NVIDIA GeForce RTX 3080 Laptop GPU /  16384 MiB /  596.47</td>
</tr>
</tbody>
</table>
<h3>1.2 推理后端</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>版本 / 参数</th>
</tr>
</thead>
<tbody>
<tr>
<td>推理框架</td>
<td>llama.cpp CUDA 版</td>
</tr>
<tr>
<td>版本</td>
<td>version: 0.5.0-dev (build 11413, commit b3daa077a) / built with Clang 20.1.8 for Windows x86_64</td>
</tr>
<tr>
<td>CUDA 后端</td>
<td>ggml-cuda.dll（NVIDIA CUDA 13.2，驱动 596.47）</td>
</tr>
</tbody>
</table>
<h3>1.3 测试方法与口径</h3>
<ol>
<li><strong>服务配置</strong>：<code>llama-server -ngl 99 -fa on -np 1 -t 16 --jinja -rea off</code>，KV 缓存按配置阶梯取 <code>q8_0</code>/<code>q4_0</code>。</li>
<li><strong>上下文场景</strong>：按 64K(65536)、96K(98304)、128K(131072)、160K(163840)、192K(196608)、224K(229376) 六档，逐模型升序实测。</li>
<li><strong>负载构造</strong>：提示词按 <code>0.875 × 上下文</code> 填充（逐模型校准字符/token 比，实测比值 5.681），生成上限 300 token，温度 0，确保长上下文场景真实生效。</li>
<li><strong>首次响应时间（TTFT）</strong>：流式请求发出至收到第一个有效内容块的墙钟秒数，其中预填充（prompt eval）占主导。</li>
<li><strong>每秒生成 Token 数（TPS）</strong>：取服务端 <code>timings.predicted_per_second</code>（不含预填充）。</li>
<li><strong>多路径配置阶梯</strong>：T1 = 默认 KV（GPU）→ T2 = KV q4_0 → T3 = 加深 <code>--n-cpu-moe</code> 卸层 → T4 = <code>-nkvo</code>（KV 落内存）。GPU 路径已达标（≥17 t/s）则不再跑 T4；未达标则强制实测对比，按真实结果择优。</li>
<li><strong>剪枝</strong>：某档所有配置均无法启动即判定更高档位不支持（KV 随上下文单调增长），记录为“不支持（推断）”。</li>
</ol>
<h3>1.4 判定标准</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>判定</th>
<th>条件</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>不合适</strong></td>
<td>可支持上下文低于 64K，<strong>或</strong> 64K 场景每秒生成 Token 数 &lt; 17</td>
</tr>
<tr>
<td><strong>合适</strong></td>
<td>可支持上下文 ≥ 64K <strong>且</strong> 64K 场景 TPS ≥ 17</td>
</tr>
<tr>
<td><strong>可入启动菜单</strong></td>
<td>存在某个场景同时满足 上下文 &gt; 100K <strong>且</strong> TPS ≥ 17（取满足条件的最大场景）</td>
</tr>
</tbody>
</table>
<h2>17 全模型测试结果总表</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>序号</th>
<th>模型</th>
<th>最大可支持上下文</th>
<th>64K TTFT(s)</th>
<th>64K TPS</th>
<th>模型判定</th>
<th>可入菜单场景</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td>Qwen3.6-35B-A3B 官方 IQ2_M</td>
<td>224K</td>
<td>55.38</td>
<td>38.88</td>
<td>合适</td>
<td>224K</td>
</tr>
<tr>
<td>2</td>
<td>Qwen3.6-35B-A3B 无审查 Aggressive IQ3_M</td>
<td>224K</td>
<td>73.02</td>
<td>33.01</td>
<td>合适</td>
<td>160K</td>
</tr>
<tr>
<td>3</td>
<td>Qwen3.6-35B-A3B 无审查 Aggressive Q4_K_M</td>
<td>192K</td>
<td>161.47</td>
<td>26.70</td>
<td>合适</td>
<td>128K</td>
</tr>
<tr>
<td>4</td>
<td>Qwen3.8-35B-A3B 蒸馏 Q2_K</td>
<td>224K</td>
<td>84.75</td>
<td>34.58</td>
<td>合适</td>
<td>224K</td>
</tr>
<tr>
<td>5</td>
<td>Qwen3-Coder-30B-A3B UD-Q4_K_XL</td>
<td>160K</td>
<td>165.90</td>
<td>9.50</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>6</td>
<td>Qwen3.8-27B-Uncensored IQ4_XS_4BPW</td>
<td>128K</td>
<td>179.96</td>
<td>12.01</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>7</td>
<td>Qwen3.8-27B-ZB4.14-MIN IQ4_XS</td>
<td>128K</td>
<td>180.48</td>
<td>11.70</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>8</td>
<td>Jack-3.8-27B-Coder</td>
<td>128K</td>
<td>36.88</td>
<td>11.04</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>9</td>
<td>Qwen3.8-27B-Heretic-Ara 3.0 MTP</td>
<td>128K</td>
<td>183.30</td>
<td>11.57</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>10</td>
<td>Ternary-Bonsai-2-27B PQ2_0</td>
<td>无</td>
<td>—</td>
<td>—</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>11</td>
<td>Ternary-Bonsai-2-27B PTQ1_0</td>
<td>无</td>
<td>—</td>
<td>—</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>12</td>
<td>Holo4-27B Q4_K_S</td>
<td>64K</td>
<td>183.68</td>
<td>11.26</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>13</td>
<td>Qwen3.8-27B GSQ-RCO IQ3_S MTP</td>
<td>128K</td>
<td>188.97</td>
<td>11.11</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>14</td>
<td>Qwen3.8-Flash-Next GSQ-RCO IQ3_S</td>
<td>无</td>
<td>—</td>
<td>—</td>
<td>不合适</td>
<td>—</td>
</tr>
<tr>
<td>15</td>
<td>POCKET-Darwin-180B UD-Q4_K_XL</td>
<td>无</td>
<td>—</td>
<td>—</td>
<td>不合适</td>
<td>—</td>
</tr>
</tbody>
</table>
]]></description><link>https://lcz.me/topic/2261</link><generator>RSS for Node</generator><lastBuildDate>Mon, 12 Oct 2026 13:28:57 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2261.rss" rel="self" type="application/rss+xml"/><pubDate>Sun, 11 Oct 2026 01:29:46 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 【实测】Thinkpad P1 Gen4-3080显卡15款大模型测试报告摘要 on Mon, 12 Oct 2026 11:06:08 GMT]]></title><description><![CDATA[<p dir="auto">太偏门了....主要是35B干活能力不行。</p>
]]></description><link>https://lcz.me/post/24259</link><guid isPermaLink="true">https://lcz.me/post/24259</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 12 Oct 2026 11:06:08 GMT</pubDate></item></channel></rss>