<rss xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title>Strata - 标签 -</title><link>https://oklife.me/tags/strata/</link><description>Strata - 标签 -</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><managingEditor>contact@oklife.me (梦行志)</managingEditor><webMaster>contact@oklife.me (梦行志)</webMaster><copyright>© 2026 梦行志 | 保留所有权利</copyright><lastBuildDate>Wed, 21 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://oklife.me/tags/strata/" rel="self" type="application/rss+xml"/><item><title>12G 本地推理生产线 ① 总览：怎么在 12G 显卡上跑出 128K 与 40 t/s</title><link>https://oklife.me/2026/10/21/sc117-series-00-overview/</link><pubDate>Wed, 21 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/21/sc117-series-00-overview/</guid><description>&lt;div class="featured-image">
                &lt;img src="/images/Code-Art-Studio-images/sc117-series-00-overview/sc117-series-00-overview.webp" referrerpolicy="no-referrer">
            &lt;/div>一台 RTX 4070 12G + 62G 内存的消费机，怎么把 SC117（Qwen3.8-Flash-Next MoE）跑到 128K 上下文、42 tok/s，并接进 Hermes Agent 当零成本推理底座。八篇实录，全部数字可复现。</description></item><item><title>12G 本地推理生产线 ② 三维 OOM 账本：host RAM、VRAM、磁盘三道账怎么列</title><link>https://oklife.me/2026/10/21/sc117-series-01-oom-ledger/</link><pubDate>Wed, 21 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/21/sc117-series-01-oom-ledger/</guid><description>&lt;div class="featured-image">
                &lt;img src="/images/Code-Art-Studio-images/sc117-series-01-oom-ledger/sc117-series-01-oom-ledger.webp" referrerpolicy="no-referrer">
            &lt;/div>本地 MoE 部署的三种 OOM 死法（宿主内存 exit -9、显存 0 slots 拒启、磁盘软降速）：每维的账怎么算、死结长什么样、怎么修，附真实日志证据。</description></item><item><title>12G 本地推理生产线 ③ 显存调优深水区：expert cache auto 怎么算的（146 experts 事故）</title><link>https://oklife.me/2026/10/21/sc117-series-02-vram-tuning/</link><pubDate>Wed, 21 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/21/sc117-series-02-vram-tuning/</guid><description><![CDATA[<div class="featured-image">
                <img src="/images/Code-Art-Studio-images/sc117-series-02-vram-tuning/sc117-series-02-vram-tuning.webp" referrerpolicy="no-referrer">
            </div>拆开 expert cache auto 的计算逻辑（字节预算、verify/draft head 预留池、启动瞬间决定论），复盘 146 experts 事故，给出&rsquo;重启前 nvidia-smi 零检查&rsquo;这条用事故换来的启动纪律。]]></description></item><item><title>12G 本地推理生产线 ④ 32K 墙是怎么死的：死旋案 + KV 流式 + 128K 窗口决策</title><link>https://oklife.me/2026/10/21/sc117-series-03-32k-wall-128k/</link><pubDate>Wed, 21 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/21/sc117-series-03-32k-wall-128k/</guid><description>&lt;div class="featured-image">
                &lt;img src="/images/Code-Art-Studio-images/sc117-series-03-32k-wall-128k/sc117-series-03-32k-wall-128k.webp" referrerpolicy="no-referrer">
            &lt;/div>agent 长会话在 31K 撞死旋墙、32K 版在 32.4–32.7K prompt 打回答塌缩：窗口截断机制、KV 驻留与流式的真实行为、int8 vs q4_0 取舍，以及为什么最终推到 128K。</description></item><item><title>12G 本地推理生产线 ⑤ 量化位怎么选：IQ2_XS vs IQ3_S 四方台账与主线决策</title><link>https://oklife.me/2026/10/21/sc117-series-04-quantization-selection/</link><pubDate>Wed, 21 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/21/sc117-series-04-quantization-selection/</guid><description><![CDATA[<div class="featured-image">
                <img src="/images/Code-Art-Studio-images/sc117-series-04-quantization-selection/sc117-series-04-quantization-selection.webp" referrerpolicy="no-referrer">
            </div>同一张 4070、同一套调优块下官方 IQ2_XS / SC117-IQ2_XS / IQ3_S 四方实测台账：速度差距根因是 expert blob 大小。主线/备用决策与四个&rsquo;触发前不动&rsquo;的参数触发器。]]></description></item><item><title>在 Ubuntu 上用 Strata 部署 SC117：从下载、独立打包到文本与视觉实测的完整教程</title><link>https://oklife.me/2026/10/17/strata-sc117-deployment-guide/</link><pubDate>Sat, 17 Oct 2026 00:00:00 +0000</pubDate><author>contact@oklife.me (梦行志)</author><guid>https://oklife.me/2026/10/17/strata-sc117-deployment-guide/</guid><description>&lt;div class="featured-image">
                &lt;img src="/images/Code-Art-Studio-images/strata-sc117-deployment-guide/strata-sc117-deployment-guide.webp" referrerpolicy="no-referrer">
            &lt;/div>在原版 Strata + Qwen3.8-Flash-Next 已跑通的机器上部署 SC117 的实操：分片与 MTP 运行时成套下载、独立 native pack、独立配置与启动脚本、131072-token 上下文服务启动，并完成两轮文本与一轮视觉任务实测。</description></item></channel></rss>