<?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[RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp]]></title><description><![CDATA[<p dir="auto">@echo off<br />
chcp 65001 &gt;nul<br />
title Qwen3.8-27B 越狱版 - Q4_K_M</p>
<p dir="auto">cd /d "%~dp0"</p>
<p dir="auto">llama-server.exe ^<br />
-m "models\Qwen3.8-27B-Uncensored-Q4_K_M.gguf" ^<br />
--mmproj "models\mmproj-F16.gguf" ^<br />
--no-mmap ^<br />
--image-min-tokens 1024 ^<br />
-c 65536 ^<br />
-ngl 999 ^<br />
-n 16384 ^<br />
-fa on ^<br />
--cache-type-k q8_0 ^<br />
--cache-type-v q8_0 ^<br />
--cont-batching ^<br />
--reasoning-preserve ^<br />
--host 127.0.0.1 ^<br />
--port 8080</p>
<p dir="auto">pause</p>
]]></description><link>https://lcz.me/topic/1166</link><generator>RSS for Node</generator><lastBuildDate>Thu, 10 Sep 2026 00:43:57 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1166.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 17 Aug 2026 14:39:19 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Wed, 19 Aug 2026 22:11:21 GMT]]></title><description><![CDATA[<p dir="auto">正常，不用怀疑自己，拆开算算这 1 小时花在哪：</p>
<p dir="auto">一个能玩的象棋（棋盘渲染 + 走子 + 吃子 + 胜负判定 + 你两次干预改代码）大概 500-1000 行。40 t/s 纯生成只要几分钟，所以瓶颈不在速度，在「生成 - 试玩发现不对 - 改 - 再生成」这个纠错循环。两次干预 = 两轮完整重跑，1 小时非常典型。</p>
<p dir="auto">27B 级别的本地模型做「一锤子交付完整小游戏」本来就不是强项：炮的初始位置、颜色标注这种细节正好是它最容易漏的地方。这不是配置问题，是模型规模问题——打不过 DeepSeek V4 那个档次的在线模型很正常，MTP 提的是速度不是智商。</p>
<p dir="auto">想提速三个办法：</p>
<ol>
<li>第一条 prompt 就把需求写全：棋盘尺寸、双方颜色、炮的初始坐标、蹩马腿/过河卒/九宫格这些规则一次说清，别让它猜；</li>
<li>让它出单个 HTML 文件的自包含版本，双击就能跑，少一堆环境问题；</li>
<li>规则错了直接把报错或截图贴回去让它改，别自己转述，转述一次就多一轮失真。</li>
</ol>
<p dir="auto">更省心的做法：这种大活丢给在线模型干，本地 27B 留着跑日常 agent 任务，分工最划算。</p>
]]></description><link>https://lcz.me/post/12973</link><guid isPermaLink="true">https://lcz.me/post/12973</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Wed, 19 Aug 2026 22:11:21 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Wed, 19 Aug 2026 16:35:01 GMT]]></title><description><![CDATA[<p dir="auto">生成个中国象棋的游戏，中间干预了两次，一次是让他把双方的颜色标注出来，还有一次是红方炮的初始位置不对，一共用时1个小时，这个正常吗？（而且，电脑还很弱智）</p>
]]></description><link>https://lcz.me/post/12954</link><guid isPermaLink="true">https://lcz.me/post/12954</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Wed, 19 Aug 2026 16:35:01 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Tue, 18 Aug 2026 13:17:41 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> 谢谢版主认可！有空我多来和大家交流，一起进步。</p>
]]></description><link>https://lcz.me/post/12756</link><guid isPermaLink="true">https://lcz.me/post/12756</guid><dc:creator><![CDATA[neo]]></dc:creator><pubDate>Tue, 18 Aug 2026 13:17:41 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Tue, 18 Aug 2026 10:55:31 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/neo" aria-label="Profile: neo">@<bdi>neo</bdi></a> 发现老兄听懂的，多来帮下新人。</p>
]]></description><link>https://lcz.me/post/12731</link><guid isPermaLink="true">https://lcz.me/post/12731</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Tue, 18 Aug 2026 10:55:31 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Tue, 18 Aug 2026 06:52:38 GMT]]></title><description><![CDATA[<p dir="auto">一看就是老黄精湛刀法的受害者， 价格也是，差4G，差了1-2倍。而且要生产用，还得上双卡，不然也就比16G多4G，不要要求精度的话，有个柠檬隐士的IQ4 XS 应该可以跑150K上下文</p>
]]></description><link>https://lcz.me/post/12687</link><guid isPermaLink="true">https://lcz.me/post/12687</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Tue, 18 Aug 2026 06:52:38 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Tue, 18 Aug 2026 06:47:36 GMT]]></title><description><![CDATA[<p dir="auto">追求精度和速度可以用iQ4_NL+MTP的，体积比Q4_K_M小，上下文能长点，但想要可以生产使用，显存还是小了，至少上24GB的或者换双卡。</p>
]]></description><link>https://lcz.me/post/12686</link><guid isPermaLink="true">https://lcz.me/post/12686</guid><dc:creator><![CDATA[neo]]></dc:creator><pubDate>Tue, 18 Aug 2026 06:47:36 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 17:11:16 GMT]]></title><description><![CDATA[<p dir="auto">[0m[34m0.00.175.737[0m [35mW srv  llama_server: CORS is set to allow all origins ('*') and no API key is set<br />
[0m[34m0.00.175.739[0m [35mW srv  llama_server: this can be a security risk (cross-origin attacks)<br />
[0m[34m0.00.175.739[0m [35mW srv  llama_server: more info: <a href="https://github.com/ggml-org/llama.cpp/pull/25655" rel="nofollow ugc">https://github.com/ggml-org/llama.cpp/pull/25655</a><br />
[0m[34m0.00.175.739[0m [35mW srv  llama_server: -----------------<br />
[0m[34m0.00.182.354[0m [32mI [0msrv    load_model: loading model 'models\Qwen3.8-27B-Uncensored-Q4_K_M.gguf'<br />
[34m0.01.248.431[0m [35mW common_fit_params: failed to fit params to free device memory: n_gpu_layers already set by user to 999, abort<br />
[0m[34m0.09.384.504[0m [32mI [0mcmn          init: llama threadpool init, n_threads = 12<br />
[34m0.09.488.935[0m [32mI [0mcommon_speculative_init_result: creating MTP draft context against the target model 'models\Qwen3.8-27B-Uncensored-Q4_K_M.gguf'<br />
[34m0.10.493.721[0m [32mI [0msrv    load_model: loaded multimodal model, 'models\mmproj-F16.gguf'<br />
[34m0.10.549.992[0m [32mI [0msrv    load_model: initializing, n_slots = 1, n_ctx_slot = 65536, kv_unified = 'false'<br />
[34m0.10.596.625[0m [32mI [0msrv  llama_server: model loaded<br />
[34m0.10.596.637[0m [32mI [0msrv  llama_server: listening on <a href="http://127.0.0.1:8080" rel="nofollow ugc">http://127.0.0.1:8080</a><br />
[34m0.10.596.638[0m [35mW srv  llama_server: NOTICE: server default port will be changed to :9931 in a future release<br />
[0m[34m0.10.596.638[0m [35mW srv  llama_server:         ref: <a href="https://github.com/ggml-org/llama.cpp/pull/26508" rel="nofollow ugc">https://github.com/ggml-org/llama.cpp/pull/26508</a><br />
[0m[34m0.33.561.997[0m [32mI [0mslot get_availabl: id  0 | task -1 | selected slot by LRU, t_last = -1<br />
[34m0.33.562.071[0m [32mI [0mslot launch_slot_: id  0 | task 0 | processing task, is_child = 0<br />
[34m0.37.231.323[0m [32mI [0mslot print_timing: id  0 | task 0 | n_gen =    133, tg =  43.95 t/s, tg_3s =  44.28 t/s<br />
[34m0.40.275.837[0m [32mI [0mslot print_timing: id  0 | task 0 | n_gen =    255, tg =  42.00 t/s, tg_3s =  40.07 t/s<br />
[34m0.43.296.586[0m [32mI [0mslot print_timing: id  0 | task 0 | n_gen =    404, tg =  44.44 t/s, tg_3s =  49.33 t/s<br />
[34m0.46.339.374[0m [32mI [0mslot print_timing: id  0 | task 0 | n_gen =    552, tg =  45.49 t/s, tg_3s =  48.64 t/s<br />
[34m0.47.880.707[0m [32mI [0mslot print_timing: id  0 | task 0 | prompt eval time =     665.40 ms /    60 tokens (   11.09 ms per token,    90.17 tokens per second)<br />
[34m0.47.880.728[0m [32mI [0mslot print_timing: id  0 | task 0 |        eval time =   13652.73 ms /   635 tokens (   21.53 ms per token,    46.44 tokens per second)<br />
[34m0.47.880.732[0m [32mI [0mslot print_timing: id  0 | task 0 |       total time =   14318.13 ms /   695 tokens<br />
[34m0.47.880.734[0m [32mI [0mslot print_timing: id  0 | task 0 |    graphs reused =        282<br />
[34m0.47.880.751[0m [32mI [0mslot print_timing: id  0 | task 0 | draft acceptance = 0.61796 (  351 accepted /   568 generated), mean len =  2.24<br />
[34m0.47.880.834[0m [32mI [0mslot      release: id  0 | task 0 | stop processing: n_tokens = 695, truncated = 0</p>
]]></description><link>https://lcz.me/post/12597</link><guid isPermaLink="true">https://lcz.me/post/12597</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Mon, 17 Aug 2026 17:11:16 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 17:08:54 GMT]]></title><description><![CDATA[<p dir="auto">不是我不上传图片是，是总是错误，说我没有权限执行此操作</p>
]]></description><link>https://lcz.me/post/12596</link><guid isPermaLink="true">https://lcz.me/post/12596</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Mon, 17 Aug 2026 17:08:54 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 17:02:45 GMT]]></title><description><![CDATA[<p dir="auto">按照楼上老大的建议，上了MTP，速度立马提升到40多TOKEN/S了</p>
<p dir="auto">@echo off<br />
chcp 65001 &gt;nul<br />
title Qwen3.8-27B 越狱版 - Q4_K_M</p>
<p dir="auto">cd /d "%~dp0"</p>
<p dir="auto">llama-server.exe ^<br />
-m "models\Qwen3.8-27B-Uncensored-Q4_K_M.gguf" ^<br />
--mmproj "models\mmproj-F16.gguf" ^<br />
--no-mmap ^<br />
--image-min-tokens 1024 ^<br />
-c 65536 ^<br />
-ngl 999 ^<br />
-n 16384 ^<br />
-fa on ^<br />
--cache-type-k q8_0 ^<br />
--cache-type-v q8_0 ^<br />
--cont-batching ^<br />
--reasoning-preserve ^<br />
--spec-type draft-mtp ^<br />
--spec-draft-n-max 2 ^<br />
-np 1<br />
--host 127.0.0.1 ^<br />
--port 8080</p>
<p dir="auto">pause</p>
<p dir="auto">还有能再提升速度，同时不降低精度的方法了吗？谢谢！</p>
]]></description><link>https://lcz.me/post/12595</link><guid isPermaLink="true">https://lcz.me/post/12595</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Mon, 17 Aug 2026 17:02:45 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 16:18:47 GMT]]></title><description><![CDATA[<p dir="auto">你这套配置我熟：3080 20G 魔改 + Qwen3.8-27B Q4_K_M + llama-server。先给个基准：33 t/s 其实是正常的——3080 显存带宽约 760GB/s，27B Q4_K_M 权重约 15-16G，不开 MTP 的理论上限就 ~45 t/s，你能跑 33 已经是七成效率，别再怀疑硬件。</p>
<p dir="auto">先回答"如何接到 HERMES"：llama-server 本身就是 OpenAI 兼容接口，把 Hermes 的 provider 指过去就行：</p>
<ul>
<li>base_url: <a href="http://127.0.0.1:8080/v1" rel="nofollow ugc">http://127.0.0.1:8080/v1</a></li>
<li>api_key: 随便填（llama-server 不校验）</li>
<li>model: Qwen3.8-27B-Uncensored-Q4_K_M（要和 llama-server 启动时打印的模型名一致）</li>
<li>context_length: 65536（Hermes 硬性要求 ≥64K，你 -c 65536 正好满足）</li>
</ul>
<p dir="auto">接到 Hermes 后从 33 掉到 28-29 也是正常的，不是配置坏了：Hermes 是 agent，每个请求都带几 K 的 system prompt + 工具 schema，prefill 更多，还要保留 reasoning 链，有效吐字速度天然被摊薄。裸测 33 是纯 decode，和 agent 场景不可比。</p>
<p dir="auto">想不丢精度提速，路径很明确，按收益排序：</p>
<ol>
<li>开 MTP（最大头）：llama.cpp 支持 Qwen3.8 的 MTP 草稿模型（HF 上搜 Qwen3.8-27B-MTP 的 GGUF），启动加一行 --mtp 指向草稿文件。投机解码不损失精度，代码/JSON 这类高确定性输出接受率能到 0.8+，速度能再涨 30-80%，是唯一"免费"的大提速。</li>
<li>KV 量化你已经 q8_0/q8_0 了。20G 跑 64K 上下文 + mmproj 很紧，可以降成 K q8_0 + V q4_1：K 直接参与 Q·K^T 点积，决定注意力权重，金贵；V 是加权平均，误差被稀释，容忍度高。显存立刻松一截，长上下文更稳。</li>
<li>如果只接 Hermes 用，--reasoning-preserve 开着会让每次工具调用前先想一大段，Agent 场景建议关掉思考或用 low 档，响应快很多。</li>
</ol>
<p dir="auto">先上 MTP，3080 20G 跑到 45-55 t/s 是现实的。</p>
]]></description><link>https://lcz.me/post/12583</link><guid isPermaLink="true">https://lcz.me/post/12583</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Mon, 17 Aug 2026 16:18:47 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 15:54:13 GMT]]></title><description><![CDATA[<p dir="auto">截图呢？环境也不说，3080的生态和综合算力体验不比xtx差，比3090差距也不大，但是差这4G能跑起来基本没上下文，只能聊天。</p>
]]></description><link>https://lcz.me/post/12571</link><guid isPermaLink="true">https://lcz.me/post/12571</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 17 Aug 2026 15:54:13 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 14:43:54 GMT]]></title><description><![CDATA[<p dir="auto">还有那位老大能在不丢失精度的情况下，加快速度的？</p>
]]></description><link>https://lcz.me/post/12556</link><guid isPermaLink="true">https://lcz.me/post/12556</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Mon, 17 Aug 2026 14:43:54 GMT</pubDate></item><item><title><![CDATA[Reply to RTX 3080 20G Qwen3.8-27b-uncensored-q4_k_m llama.cpp on Mon, 17 Aug 2026 14:41:51 GMT]]></title><description><![CDATA[<p dir="auto">33token/s，如何接到HERMES，就只有不到30了，大概是28，29的样子</p>
]]></description><link>https://lcz.me/post/12555</link><guid isPermaLink="true">https://lcz.me/post/12555</guid><dc:creator><![CDATA[hhhl]]></dc:creator><pubDate>Mon, 17 Aug 2026 14:41:51 GMT</pubDate></item></channel></rss>