{"id":77,"date":"2026-08-27T16:23:00","date_gmt":"2026-08-27T16:23:00","guid":{"rendered":"https:\/\/jyy7998.tk\/?p=77"},"modified":"2026-09-18T00:57:21","modified_gmt":"2026-09-18T00:57:21","slug":"%f0%9f%9a%80-windows-11-qwen3-8-27b-%e6%a8%a1%e5%9e%8b%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2","status":"publish","type":"post","link":"https:\/\/wp.jyy7998.tk\/?p=77","title":{"rendered":"\ud83d\ude80 Windows 11 Qwen3.8-27B \u6a21\u578b\u672c\u5730\u90e8\u7f72"},"content":{"rendered":"<div class=\"doc\" style=\"--doc-accent:#1e8cbe;--doc-accent-soft:#edf6fa\">\n<p class=\"doc-lead\">RTX 5090 24GB + Qwen3.8-27B Q4_K_M + DeepSeek Harness<\/p>\n<p>\u672c\u6559\u7a0b\u9762\u5411 <strong>Windows 11<\/strong> \u7cfb\u7edf\uff0c\u663e\u5361\u4e3a <strong>NVIDIA RTX 5090 24GB<\/strong>\u3002\u5047\u8bbe\u4f60\u5df2\u5b8c\u6210 Windows 11 \u7cfb\u7edf\u5b89\u88c5\u548c NVIDIA \u663e\u5361\u9a71\u52a8\u5b89\u88c5\uff08\u9a71\u52a8\u7248\u672c \u2265 570\uff09\u3002\u6211\u4eec\u5c06\u4ece CUDA \u5de5\u5177\u5305\u5b89\u88c5\u5f00\u59cb\uff0c\u9010\u6b65\u5b8c\u6210 llama.cpp \u9884\u7f16\u8bd1\u5305\u90e8\u7f72\u3001Qwen3.8-27B \u6a21\u578b\u4e0b\u8f7d\uff0c\u4ee5\u53ca DeepSeek Harness \u7684\u5b89\u88c5\u4e0e\u914d\u7f6e\uff0c\u6700\u7ec8\u6784\u5efa\u4e00\u5957\u5b8c\u6574\u7684\u672c\u5730 AI \u5f00\u53d1\u73af\u5883\u3002<\/p>\n<nav class=\"doc-toc\">\n  <strong>\ud83d\udcd6 \u76ee\u5f55<\/strong><\/p>\n<ul>\n<li><a href=\"#sec1\">1. \u5b89\u88c5 CUDA Toolkit<\/a><\/li>\n<li><a href=\"#sec2\">2. \u5b89\u88c5 Git for Windows<\/a><\/li>\n<li><a href=\"#sec3\">3. \u4e0b\u8f7d llama.cpp \u9884\u7f16\u8bd1\u5305<\/a><\/li>\n<li><a href=\"#sec4\">4. \u4e0b\u8f7d Qwen3.8-27B Q4_K_M \u6a21\u578b<\/a><\/li>\n<li><a href=\"#sec5\">5. \u6d4b\u8bd5\u63a8\u7406<\/a><\/li>\n<li><a href=\"#sec6\">6. \u521b\u5efa\u542f\u52a8\u811a\u672c\uff08\u6279\u5904\u7406\uff09<\/a><\/li>\n<li><a href=\"#sec7\">7. \u521b\u5efa\u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/a><\/li>\n<li><a href=\"#sec8\">8. \u5b89\u88c5 DeepSeek Harness<\/a><\/li>\n<li><a href=\"#sec9\">9. \u914d\u7f6e\u672c\u5730 Qwen \u6a21\u578b\u63a5\u5165 Harness<\/a><\/li>\n<li><a href=\"#sec10\">10. \u521b\u5efa Harness \u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/a><\/li>\n<li><a href=\"#sec11\">11. \u5e38\u89c1\u95ee\u9898<\/a><\/li>\n<\/ul>\n<\/nav>\n<hr>\n<h2 id=\"sec1\">1. \u5b89\u88c5 CUDA Toolkit<\/h2>\n<p><strong>RTX 5090 \u9700\u8981 CUDA 12.8 \u6216\u66f4\u9ad8\u7248\u672c<\/strong>\u624d\u80fd\u5b8c\u6574\u652f\u6301 Blackwell \u67b6\u6784\u3002<\/p>\n<p>\u8bbf\u95ee <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" target=\"_blank\" rel=\"noopener\">NVIDIA CUDA Toolkit \u4e0b\u8f7d\u9875\u9762<\/a>\uff0c\u9009\u62e9\uff1a<\/p>\n<ul>\n<li><strong>Operating System<\/strong>\uff1aWindows<\/li>\n<li><strong>Architecture<\/strong>\uff1ax86_64<\/li>\n<li><strong>Version<\/strong>\uff1a11\uff08\u6216\u5bf9\u5e94\u4f60\u7684 Windows \u7248\u672c\uff09<\/li>\n<li><strong>Installer Type<\/strong>\uff1aexe (local)<\/li>\n<\/ul>\n<p>\u4e0b\u8f7d\u5b89\u88c5\u5305\uff08\u5982 <code>cuda_13.3.0_xxx_windows.exe<\/code>\uff09\u3002<\/p>\n<p><strong>\u4ee5\u7ba1\u7406\u5458\u8eab\u4efd\u8fd0\u884c<\/strong>\u5b89\u88c5\u7a0b\u5e8f\uff0c\u5efa\u8bae\uff1a<\/p>\n<ul>\n<li>\u5b89\u88c5\u8def\u5f84\u4fdd\u6301\u9ed8\u8ba4\uff1a<code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v13.3<\/code><\/li>\n<li>\u7ec4\u4ef6\u9009\u62e9\uff1a\u786e\u4fdd\u52fe\u9009 <strong>CUDA<\/strong> \u6838\u5fc3\u7ec4\u4ef6<\/li>\n<li>\u81ea\u5b9a\u4e49\u5b89\u88c5\u65f6\u53ef\u53ea\u52fe\u9009 Runtime \u548c Development \u76f8\u5173\u7ec4\u4ef6<\/li>\n<\/ul>\n<p>\u5b89\u88c5\u5b8c\u6210\u540e\uff0c<strong>\u91cd\u542f\u7535\u8111<\/strong>\uff0c\u7136\u540e\u6253\u5f00\u65b0\u7684 PowerShell \u9a8c\u8bc1\uff1a<\/p>\n<p>\u68c0\u67e5\u7248\u672c\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>nvcc --version<\/code><\/pre>\n<p>\u5e94\u663e\u793a CUDA \u7248\u672c \u2265 13.3\u3002<\/p>\n<div class=\"doc-note doc-note--warn\">\n  \u26a0\ufe0f <strong>\u91cd\u8981\uff1a<\/strong> \u4e0d\u8981\u901a\u8fc7 <code>conda install cudatoolkit<\/code> \u5b89\u88c5 CUDA\uff0cRTX 5090 \u9700\u8981\u5b8c\u6574\u7684\u672c\u5730 CUDA Toolkit\u3002\n<\/div>\n<hr>\n<h2 id=\"sec2\">2. \u5b89\u88c5 Git for Windows<\/h2>\n<p>DeepSeek Harness \u7684\u6e90\u7801\u5b89\u88c5\u9700\u8981 <code>git<\/code> \u547d\u4ee4\u6765\u514b\u9686\u4ed3\u5e93\u3002\u5982\u679c\u4f60\u7684\u7cfb\u7edf\u4e2d\u5c1a\u672a\u5b89\u88c5 Git\uff0c\u8bf7\u6309\u4ee5\u4e0b\u6b65\u9aa4\u64cd\u4f5c\uff1a<\/p>\n<ol>\n<li>\u8bbf\u95ee <a href=\"https:\/\/git-scm.com\/download\/win\" target=\"_blank\" rel=\"noopener\">Git \u5b98\u7f51\u4e0b\u8f7d\u9875\u9762<\/a>\uff0c\u4e0b\u8f7d 64 \u4f4d Windows \u5b89\u88c5\u5305\uff08\u5982 <code>Git-2.45.0-64-bit.exe<\/code>\uff09\u3002<\/li>\n<li>\u8fd0\u884c\u5b89\u88c5\u7a0b\u5e8f\uff0c\u5efa\u8bae\u4fdd\u7559\u9ed8\u8ba4\u9009\u9879\uff0c\u4e00\u8def <strong>Next<\/strong> \u5373\u53ef\uff08\u53ef\u52fe\u9009 \u201cGit from the command line and also from 3rd-party software\u201d \u4ee5\u4fbf\u5728 PowerShell \u4e2d\u4f7f\u7528\uff09\u3002<\/li>\n<li>\u5b89\u88c5\u5b8c\u6210\u540e\uff0c<strong>\u5173\u95ed\u5e76\u91cd\u65b0\u6253\u5f00<\/strong> PowerShell\uff0c\u6267\u884c\u4ee5\u4e0b\u547d\u4ee4\u9a8c\u8bc1\uff1a<\/li>\n<\/ol>\n<pre class=\"prettyprint linenums\" ><code>git --version<\/code><\/pre>\n<p>\u5e94\u663e\u793a Git \u7248\u672c\u53f7\uff08\u5982 <code>git version 2.45.0.windows.1<\/code>\uff09\u3002<\/p>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 \u5982\u679c\u4e0d\u60f3\u5b89\u88c5 Git\uff0c\u4f60\u4e5f\u53ef\u4ee5\u76f4\u63a5\u4e0b\u8f7d Harness \u7684 <strong>\u684c\u9762\u5e94\u7528\u7248<\/strong>\uff08\u89c1\u7b2c 8 \u8282\u65b9\u5f0f\u4e8c\uff09\uff0c\u65e0\u9700\u4f7f\u7528 Git\u3002\n<\/div>\n<hr>\n<h2 id=\"sec3\">3. \u4e0b\u8f7d llama.cpp \u9884\u7f16\u8bd1\u5305<\/h2>\n<p>Windows \u4e0b\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u5b98\u65b9\u63d0\u4f9b\u7684\u9884\u7f16\u8bd1\u4e8c\u8fdb\u5236\u6587\u4ef6\uff0c\u7701\u53bb\u5b89\u88c5 Visual Studio\u3001CMake \u7b49\u7f16\u8bd1\u5de5\u5177\u7684\u7e41\u7410\u6b65\u9aa4\u3002<\/p>\n<p>\u8bbf\u95ee <a href=\"https:\/\/github.com\/ggml-org\/llama.cpp\/releases\" target=\"_blank\" rel=\"noopener\">llama.cpp Releases<\/a>\uff0c\u5728 <strong>Assets<\/strong> \u4e2d\u627e\u5230\u6700\u65b0\u7248\u672c\uff08<code>bXXXX<\/code>\uff09\uff0c\u4e0b\u8f7d\u5bf9\u5e94\u4f60\u663e\u5361\u7684\u9884\u7f16\u8bd1\u5305\u3002<\/p>\n<h3>\u663e\u5361\u4e0e\u9884\u7f16\u8bd1\u5305\u9009\u62e9\uff08\u6027\u80fd\u4f18\u5148\uff09<\/h3>\n<table class=\"doc-table\">\n<thead>\n<tr>\n<th>\u663e\u5361\u7cfb\u5217<\/th>\n<th>\u4ee3\u8868\u578b\u53f7<\/th>\n<th>\u63a8\u8350\u9884\u7f16\u8bd1\u5305<\/th>\n<th>\u6027\u80fd\u8bf4\u660e<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>RTX 50 \u7cfb\u5217<\/strong><\/td>\n<td>RTX 5090\u30015080\u30015070<\/td>\n<td><code>cuda-13.3<\/code> \u6216\u66f4\u9ad8<\/td>\n<td>\u5fc5\u987b\uff0c\u5426\u5219\u65e0\u6cd5\u9a71\u52a8\u663e\u5361<\/td>\n<\/tr>\n<tr>\n<td><strong>RTX 40 \u7cfb\u5217<\/strong><\/td>\n<td>RTX 4090\u30014080\u30014070<\/td>\n<td><strong><code>cuda-13.3<\/code><\/strong>\uff08\u6027\u80fd\u9996\u9009\uff09<br \/>\u6216 <code>cuda-12.8<\/code><\/td>\n<td>CUDA 13.3 \u5e26\u6765 <strong>12-14%<\/strong> \u6574\u4f53\u6027\u80fd\u63d0\u5347\uff0cPrefill \u9636\u6bb5\u63d0\u5347\u9ad8\u8fbe <strong>38%<\/strong>\uff0c\u5e76\u652f\u6301 FP8 \u63a8\u7406\uff08\u901f\u5ea6\u7ffb\u500d\uff09<\/td>\n<\/tr>\n<tr>\n<td><strong>RTX 30 \u7cfb\u5217<\/strong><\/td>\n<td>RTX 3090\u30013080\u30013070<\/td>\n<td><strong><code>cuda-13.3<\/code><\/strong>\uff08\u6027\u80fd\u9996\u9009\uff09<br \/>\u6216 <code>cuda-12.8<\/code><\/td>\n<td>CUDA 13.3 \u5728 Prefill \u9636\u6bb5\u6709 <strong>6-9%<\/strong> \u63d0\u5347\uff0cGeneration \u9636\u6bb5\u65e0\u53d8\u5316<\/td>\n<\/tr>\n<tr>\n<td><strong>RTX 20 \u7cfb\u5217<\/strong><\/td>\n<td>RTX 2080\u30012070\u30012060<\/td>\n<td><code>cuda-12.x<\/code><\/td>\n<td>Turing \u67b6\u6784\uff0cCUDA 13 \u65e0\u6027\u80fd\u63d0\u5347\uff0c\u7528 12.x \u5373\u53ef<\/td>\n<\/tr>\n<tr>\n<td><strong>GTX 16 \u7cfb\u5217<\/strong><\/td>\n<td>GTX 1660\u30011650<\/td>\n<td><code>cuda-12.x<\/code><\/td>\n<td>Turing \u67b6\u6784\uff08\u65e0 Tensor Core\uff09\uff0cCUDA 13 \u65e0\u63d0\u5347<\/td>\n<\/tr>\n<tr>\n<td><strong>GTX 10 \u7cfb\u5217<\/strong><\/td>\n<td>GTX 1080 Ti\u30011070<\/td>\n<td><code>cuda-12.x\uff08legacy\uff09<\/code><\/td>\n<td>Pascal \u67b6\u6784\uff0c\u4e0d\u652f\u6301 CUDA 13<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 \u8ffd\u6c42\u6027\u80fd\u5c31\u9009 <strong>cuda-13.3<\/strong>\uff08RTX 30\/40\/50 \u7cfb\uff09\u3002CUDA 13.3 \u7684\u4f18\u5316\u91cd\u5fc3\u5728 RTX 40\/50 \u7cfb\u5217\uff0cRTX 30 \u7cfb\u6709\u5c11\u91cf\u63d0\u5347\uff0cRTX 20 \u7cfb\u53ca\u66f4\u65e9\u65e0\u63d0\u5347\u3002\n<\/div>\n<h3>\u5fc5\u987b\u540c\u65f6\u4e0b\u8f7d\u914d\u5957 DLL<\/h3>\n<p>\u5728 llama.cpp Releases \u9875\u9762\u4e2d\uff0c\u4f60\u9700\u8981\u4e0b\u8f7d <strong>\u4e24\u4e2a\u6587\u4ef6<\/strong>\uff0c\u7f3a\u4e00\u4e0d\u53ef\uff1a<\/p>\n<ol>\n<li><strong>\u4e3b\u7a0b\u5e8f\u5305<\/strong>\uff1a<code>llama-bXXXX-bin-win-cuda-13.3-x64.zip<\/code>\uff08\u6216 <code>cuda-12<\/code> \u7248\u672c\uff09<\/li>\n<li><strong>DLL \u5305<\/strong>\uff1a<code>cudart-llama-bin-win-cuda-13.3-x64.zip<\/code>\uff08\u6216\u5bf9\u5e94\u7248\u672c\uff09<\/li>\n<\/ol>\n<h3>\u5b89\u88c5\u65b9\u6cd5<\/h3>\n<ol>\n<li>\u5c06\u4e24\u4e2a\u538b\u7f29\u5305\u90fd\u4e0b\u8f7d\u5230\u672c\u5730<\/li>\n<li>\u5148\u89e3\u538b\u4e3b\u7a0b\u5e8f\u5305\u5230\u4f60\u7684\u76ee\u5f55\uff0c\u4f8b\u5982 <code>D:\\dev\\llm\\llama.cpp<\/code><\/li>\n<li><strong>\u518d\u89e3\u538b DLL \u5305\uff0c\u5c06\u5176\u4e2d\u7684\u6240\u6709 <code>.dll<\/code> \u6587\u4ef6\u590d\u5236\u5230\u4e0e <code>llama-server.exe<\/code> \u76f8\u540c\u7684\u76ee\u5f55\u4e0b<\/strong><\/li>\n<\/ol>\n<p>\u6700\u7ec8\u76ee\u5f55\u7ed3\u6784\u5e94\u7c7b\u4f3c\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>D:\\dev\\llm\\llama.cpp\\\r\n\u251c\u2500\u2500 llama-server.exe\r\n\u251c\u2500\u2500 llama-cli.exe\r\n\u251c\u2500\u2500 cudart64_13.dll      \u2190 \u6765\u81ea DLL \u5305\r\n\u251c\u2500\u2500 cublas64_13.dll      \u2190 \u6765\u81ea DLL \u5305\r\n\u251c\u2500\u2500 cublasLt64_13.dll    \u2190 \u6765\u81ea DLL \u5305\r\n\u251c\u2500\u2500 ...\uff08\u5176\u4ed6 DLL \u6587\u4ef6\uff09\r\n\u2514\u2500\u2500 models\\\r\n    \u2514\u2500\u2500 Qwen3.8-27B\\\r\n        \u2514\u2500\u2500 qwen3.8-27b-q4_k_m.gguf<\/code><\/pre>\n<div class=\"doc-note doc-note--warn\">\n  \u26a0\ufe0f \u5982\u679c\u4e0d\u653e\u5165\u8fd9\u4e9b DLL \u6587\u4ef6\uff0c<code>llama-server.exe<\/code> \u5c06<strong>\u65e0\u6cd5\u8bc6\u522b GPU<\/strong>\uff0c\u53ea\u80fd\u56de\u9000\u5230 CPU \u8fd0\u884c\uff0c\u901f\u5ea6\u4f1a\u6162\u5f88\u591a\u3002\n<\/div>\n<h3>\u9a71\u52a8\u7248\u672c\u8981\u6c42<\/h3>\n<p>\u4f7f\u7528 <code>cuda-13.3<\/code> \u9884\u7f16\u8bd1\u5305\u8981\u6c42 <strong>NVIDIA \u663e\u5361\u9a71\u52a8 \u2265 570<\/strong>\u3002\u5728 PowerShell \u4e2d\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u68c0\u67e5\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>C:\\Windows\\System32\\nvidia-smi.exe<\/code><\/pre>\n<p>\u67e5\u770b\u5217\u8868\u6700\u4e0a\u65b9\u7684 <strong>Driver Version<\/strong>\u3002\u5982\u679c\u4f4e\u4e8e 570\uff0c\u8bf7\u524d\u5f80 <a href=\"https:\/\/www.nvidia.com\/Download\/index.aspx\" target=\"_blank\" rel=\"noopener\">NVIDIA \u9a71\u52a8\u5b98\u7f51<\/a> \u66f4\u65b0\u3002<\/p>\n<hr>\n<h2 id=\"sec4\">4. \u4e0b\u8f7d Qwen3.8-27B Q4_K_M \u6a21\u578b<\/h2>\n<p>\u9009\u62e9 4-bit \u91cf\u5316\u7248\u672c\uff08Q4_K_M\uff09\uff0c\u6587\u4ef6\u5927\u5c0f\u7ea6 17-18 GB\uff0c\u5b8c\u7f8e\u9002\u914d 24GB \u663e\u5b58\u3002<\/p>\n<p>\u5728 <code>llama.cpp<\/code> \u76ee\u5f55\u4e0b\u521b\u5efa\u6a21\u578b\u6587\u4ef6\u5939\u5e76\u4e0b\u8f7d\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>cd D:\\dev\\llm\\llama.cpp\r\nmkdir models\\Qwen3.8-27B\r\ncd models\\Qwen3.8-27B\r\ncurl -L -O https:\/\/huggingface.co\/Qwen\/Qwen3.8-27B-GGUF\/resolve\/main\/qwen3.8-27b-q4_k_m.gguf<\/code><\/pre>\n<p>\u6216\u4f7f\u7528 wget\uff08\u5982\u5df2\u5b89\u88c5\uff09\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>wget https:\/\/huggingface.co\/Qwen\/Qwen3.8-27B-GGUF\/resolve\/main\/qwen3.8-27b-q4_k_m.gguf<\/code><\/pre>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 \u5982\u679c\u94fe\u63a5\u5931\u6548\uff0c\u8bf7\u8bbf\u95ee <a href=\"https:\/\/huggingface.co\/Qwen\/Qwen3.8-27B-GGUF\" target=\"_blank\">Qwen3.8-27B-GGUF \u9875\u9762<\/a> \u83b7\u53d6\u6700\u65b0\u4e0b\u8f7d\u5730\u5740\u3002\n<\/div>\n<div class=\"doc-note doc-note--warn\">\n  \u26a0\ufe0f \u4e0b\u8f7d\u8fc7\u7a0b\u4e2d\u8bf7\u786e\u4fdd\u7f51\u7edc\u7a33\u5b9a\u3002\u5982\u679c\u4e0b\u8f7d\u4e2d\u65ad\uff0c\u53ef\u6dfb\u52a0 <code>-C -<\/code> \u53c2\u6570\uff08curl\uff09\u6216 <code>-c<\/code> \u53c2\u6570\uff08wget\uff09\u7ee7\u7eed\u4e0b\u8f7d\u3002\n<\/div>\n<hr>\n<h2 id=\"sec5\">5. \u6d4b\u8bd5\u63a8\u7406<\/h2>\n<pre class=\"prettyprint linenums\" ><code>cd D:\\dev\\llm\\llama.cpp\r\n.\\llama-cli.exe -m .\\models\\Qwen3.8-27B\\qwen3.8-27b-q4_k_m.gguf -ngl 999 -p \"\u4f60\u597d\" -n 32<\/code><\/pre>\n<p>\u82e5\u80fd\u8f93\u51fa\u4e2d\u6587\u56de\u590d\uff0c\u5219\u7f16\u8bd1\u548c\u6a21\u578b\u52a0\u8f7d\u6210\u529f\u3002<\/p>\n<hr>\n<h2 id=\"sec6\">6. \u521b\u5efa\u542f\u52a8\u811a\u672c\uff08\u6279\u5904\u7406\uff09<\/h2>\n<p>\u5728\u684c\u9762\u521b\u5efa <code>run-qwen.bat<\/code> \u542f\u52a8\u811a\u672c\uff0c\u65b9\u4fbf\u4e00\u952e\u542f\u52a8 API \u670d\u52a1\u3002<\/p>\n<p>\u5728\u684c\u9762\u7a7a\u767d\u5904\u53f3\u952e \u2192 <strong>\u65b0\u5efa<\/strong> \u2192 <strong>\u6587\u672c\u6587\u6863<\/strong>\uff0c\u547d\u540d\u4e3a <code>run-qwen.bat<\/code>\uff08\u786e\u4fdd\u6269\u5c55\u540d\u4e3a <code>.bat<\/code>\uff09\uff0c\u53f3\u952e\u7f16\u8f91\uff0c\u7c98\u8d34\u4ee5\u4e0b\u5185\u5bb9\uff08\u8bf7\u6839\u636e\u5b9e\u9645\u8def\u5f84\u4fee\u6539\uff09\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>@echo off\r\ncd \/d D:\\dev\\llm\\llama.cpp\r\n\r\nllama-server.exe ^\r\n  -m \"D:\\dev\\llm\\llama.cpp\\models\\Qwen3.8-27B\\qwen3.8-27b-q4_k_m.gguf\" ^\r\n  --no-mmap ^\r\n  --ctx-size 32768 ^\r\n  --flash-attn ^\r\n  --batch-size 4096 ^\r\n  --ubatch-size 1024 ^\r\n  --parallel 1 ^\r\n  --host 0.0.0.0 ^\r\n  --port 8080 ^\r\n  -ngl 999 ^\r\n  --threads 16 ^\r\n  --threads-batch 16 ^\r\n  --cache-type-k q8_0 ^\r\n  --cache-type-v q8_0 ^\r\n  --temp 0.6 ^\r\n  --top-p 0.95 ^\r\n  --top-k 20 ^\r\n  --min-p 0.00 ^\r\n  --repeat-penalty 1.1 ^\r\n  --jinja ^\r\n  --props ^\r\n  --metrics ^\r\n  --perf\r\n\r\npause<\/code><\/pre>\n<p><strong>\u53c2\u6570\u8bf4\u660e\uff1a<\/strong><\/p>\n<ul>\n<li><code>--ctx-size 32768<\/code>\uff1a\u4e0a\u4e0b\u6587 32K\uff0c\u5b89\u5168\u8d77\u70b9\uff0c\u53ef\u6839\u636e\u663e\u5b58\u4f59\u91cf\u8c03\u9ad8\u81f3 64K<\/li>\n<li><code>--threads 16<\/code>\uff1a\u6839\u636e\u4f60\u7684 CPU \u6838\u5fc3\u6570\u9002\u5f53\u8c03\u6574<\/li>\n<li><code>--no-mmap<\/code>\uff1a\u5b8c\u5168\u52a0\u8f7d\u5230\u5185\u5b58\/\u663e\u5b58\uff0c\u907f\u514d\u78c1\u76d8 I\/O<\/li>\n<li><code>-ngl 999<\/code>\uff1a\u6240\u6709\u5c42\u5378\u8f7d\u5230 GPU<\/li>\n<li><code>--host 0.0.0.0<\/code>\uff1a\u5141\u8bb8\u672c\u673a\u548c\u5c40\u57df\u7f51\u8bbf\u95ee<\/li>\n<li><code>--flash-attn<\/code>\uff1a\u542f\u7528 Flash Attention \u52a0\u901f<\/li>\n<li><code>--jinja<\/code>\uff1a\u542f\u7528 Jinja \u6a21\u677f\u652f\u6301\uff08\u7528\u4e8e\u804a\u5929\u6a21\u677f\uff09<\/li>\n<\/ul>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 Windows \u6279\u5904\u7406\u4e2d\u4f7f\u7528 <code>^<\/code> \u4f5c\u4e3a\u884c\u7eed\u884c\u7b26\uff0c\u6ce8\u610f <code>^<\/code> \u540e\u9762\u4e0d\u80fd\u6709\u7a7a\u683c\u6216\u6ce8\u91ca\u3002\n<\/div>\n<hr>\n<h2 id=\"sec7\">7. \u521b\u5efa\u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/h2>\n<p>\u53f3\u952e\u70b9\u51fb\u684c\u9762\u4e0a\u7684 <code>run-qwen.bat<\/code> \u2192 <strong>\u53d1\u9001\u5230<\/strong> \u2192 <strong>\u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/strong>\u3002<\/p>\n<p>\u53f3\u952e\u8be5\u5feb\u6377\u65b9\u5f0f \u2192 <strong>\u5c5e\u6027<\/strong> \u2192 <strong>\u5feb\u6377\u65b9\u5f0f<\/strong>\u9009\u9879\u5361\uff0c\u53ef\u66f4\u6539\u56fe\u6807\u6216\u8c03\u6574\u8fd0\u884c\u65b9\u5f0f\uff08\u5982\u8bbe\u7f6e\u4e3a\u201c\u6700\u5c0f\u5316\u201d\u542f\u52a8\uff09\u3002<\/p>\n<p>\u53cc\u51fb\u5feb\u6377\u65b9\u5f0f\u5373\u53ef\u542f\u52a8 Qwen API \u670d\u52a1\uff0c\u670d\u52a1\u8fd0\u884c\u65f6<strong>\u7ec8\u7aef\u7a97\u53e3\u5fc5\u987b\u4fdd\u6301\u6253\u5f00<\/strong>\uff0c\u5173\u95ed\u5373\u7ec8\u6b62\u670d\u52a1\u3002<\/p>\n<hr>\n<h2 id=\"sec8\">8. \u5b89\u88c5 DeepSeek Harness<\/h2>\n<p>DeepSeek Harness \u662f DeepSeek \u5f00\u6e90\u7684 AI \u667a\u80fd\u4f53\u6846\u67b6\uff0c\u91c7\u7528\u201c\u4e00\u5207\u7686\u63d2\u4ef6\u201d\u67b6\u6784\uff0c\u652f\u6301\u63a5\u5165\u591a\u79cd\u6a21\u578b\uff08\u5305\u62ec\u672c\u5730 Ollama\u3001llama.cpp \u548c\u4e91\u7aef API\uff09\u3002\u5728 Windows \u4e0a\u6709\u4e24\u79cd\u5b89\u88c5\u65b9\u5f0f\uff1a<\/p>\n<h3>\u65b9\u5f0f\u4e00\uff1a\u6e90\u7801\u5b89\u88c5\uff08\u63a8\u8350\uff0c\u529f\u80fd\u6700\u65b0\uff09<\/h3>\n<h4>8.1 \u5b89\u88c5 Node.js<\/h4>\n<p>DeepSeek Harness <strong>\u8981\u6c42 Node.js \u2265 v22.19<\/strong>\u3002\u4ece <a href=\"https:\/\/nodejs.org\/\" target=\"_blank\" rel=\"noopener\">Node.js \u5b98\u7f51<\/a> \u4e0b\u8f7d LTS \u7248\u672c\uff08v22.x \u6216\u66f4\u9ad8\uff09\u5b89\u88c5\u3002<\/p>\n<p>\u5b89\u88c5\u5b8c\u6210\u540e\u9a8c\u8bc1\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>node --version   # \u5e94\u663e\u793a v22.x.x \u6216\u66f4\u9ad8<\/code><\/pre>\n<h4>8.2 \u5b89\u88c5 pnpm<\/h4>\n<p>DeepSeek Harness \u4f7f\u7528 pnpm \u4f5c\u4e3a\u5305\u7ba1\u7406\u5668\uff0c<strong>\u8981\u6c42\u7248\u672c \u2265 11<\/strong>\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>npm install -g pnpm\r\npnpm --version   # \u5e94\u663e\u793a 11.x.x \u6216\u66f4\u9ad8<\/code><\/pre>\n<h4>8.3 \u514b\u9686\u5e76\u5b89\u88c5\u4f9d\u8d56<\/h4>\n<pre class=\"prettyprint linenums\" ><code>cd D:\\dev\\llm\r\ngit clone https:\/\/github.com\/deepseek-ai\/deepseek-harness.git\r\ncd deepseek-harness\r\npnpm install<\/code><\/pre>\n<h4>8.4 \u6784\u5efa\u5e76\u542f\u52a8<\/h4>\n<pre class=\"prettyprint linenums\" ><code>pnpm run build\r\npnpm dsh web<\/code><\/pre>\n<p>\u542f\u52a8\u6210\u529f\u540e\uff0c\u7ec8\u7aef\u4f1a\u663e\u793a\u8bbf\u95ee\u5730\u5740\uff08\u9ed8\u8ba4 <code>http:\/\/127.0.0.1:3080<\/code>\uff09\u3002<\/p>\n<div class=\"doc-note doc-note--warn\">\n  \u26a0\ufe0f <strong>\u6ce8\u610f\uff1a<\/strong> \u6b64\u7ec8\u7aef\u7a97\u53e3\u9700\u8981\u4fdd\u6301\u6253\u5f00\uff0c\u5173\u95ed\u4f1a\u5bfc\u81f4\u670d\u52a1\u7ec8\u6b62\u3002\n<\/div>\n<h3>\u65b9\u5f0f\u4e8c\uff1a\u684c\u9762\u5e94\u7528\uff08\u66f4\u7b80\u4fbf\uff09<\/h3>\n<p>\u793e\u533a\u63d0\u4f9b\u4e86\u5c01\u88c5\u597d\u7684 Windows \u684c\u9762\u5e94\u7528\uff0c\u65e0\u9700\u914d\u7f6e Node.js \u73af\u5883\uff1a<\/p>\n<ol>\n<li>\u8bbf\u95ee <a href=\"https:\/\/github.com\/Links2008\/DeepSeek-Harness-Desktop\/releases\/latest\" target=\"_blank\" rel=\"noopener\">DeepSeek Harness Desktop Releases<\/a><\/li>\n<li>\u4e0b\u8f7d <code>DeepSeekHarness-Setup-&lt;\u7248\u672c&gt;.exe<\/code><\/li>\n<li>\u53cc\u51fb\u8fd0\u884c\u5b89\u88c5\u5411\u5bfc\uff0c\u6309\u63d0\u793a\u5b8c\u6210\u5b89\u88c5\uff08\u4f1a\u521b\u5efa\u684c\u9762\u5feb\u6377\u65b9\u5f0f\uff09<\/li>\n<li>\u4ece\u684c\u9762\u5feb\u6377\u65b9\u5f0f\u6216\u5f00\u59cb\u83dc\u5355\u542f\u52a8<\/li>\n<\/ol>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 \u684c\u9762\u5e94\u7528\u4e3a\u793e\u533a\u7ef4\u62a4\u7684\u975e\u5b98\u65b9\u7248\u672c\uff0c\u66f4\u65b0\u53ef\u80fd\u6ede\u540e\u4e8e\u6e90\u7801\u3002\u8ffd\u6c42\u6700\u65b0\u529f\u80fd\u8bf7\u4f7f\u7528\u6e90\u7801\u7f16\u8bd1\u3002\n<\/div>\n<hr>\n<h2 id=\"sec9\">9. \u914d\u7f6e\u672c\u5730 Qwen \u6a21\u578b\u63a5\u5165 Harness<\/h2>\n<p>\u5982\u679c\u4f60\u5df2\u901a\u8fc7\u65b9\u5f0f\u4e00\uff08\u6e90\u7801\uff09\u542f\u52a8 Harness\uff0c\u6216\u901a\u8fc7\u684c\u9762\u5e94\u7528\u542f\u52a8\uff0c\u53ef\u6309\u4ee5\u4e0b\u6b65\u9aa4\u5c06\u672c\u5730 Qwen \u6a21\u578b\u63a5\u5165\uff1a<\/p>\n<ol>\n<li><strong>\u786e\u4fdd llama-server \u6b63\u5728\u8fd0\u884c<\/strong>\uff08\u53cc\u51fb\u684c\u9762 <code>run-qwen.bat<\/code> \u5feb\u6377\u65b9\u5f0f\uff09<\/li>\n<li>\u6253\u5f00\u6d4f\u89c8\u5668\u8bbf\u95ee <code>http:\/\/127.0.0.1:3080<\/code>\uff0c\u8fdb\u5165 Harness Web UI\n<ul>\n<li><strong>\u9996\u6b21\u6253\u5f00\u65f6\uff0c\u4f1a\u5f39\u51fa\u4e00\u4e2a\u8981\u6c42\u8f93\u5165 DeepSeek API Key \u7684\u5bf9\u8bdd\u6846<\/strong><\/li>\n<li><strong>\u76f4\u63a5\u70b9\u51fb\u5bf9\u8bdd\u6846\u53f3\u4e0a\u89d2\u7684 \u00d7 \u6216 Cancel \u5c06\u5176\u5173\u95ed<\/strong>\uff08\u5982\u679c\u53ea\u9700\u8981\u672c\u5730\u6a21\u578b\uff09<\/li>\n<\/ul>\n<\/li>\n<li>\u70b9\u51fb\u5de6\u4e0b\u89d2\u7684 <strong>Settings<\/strong>\uff08\u9f7f\u8f6e\u56fe\u6807\uff09\uff0c\u9009\u62e9 <strong>Models<\/strong> \u9009\u9879\u5361<\/li>\n<li>\u70b9\u51fb <strong>Add Custom Model<\/strong>\uff08\u6216\u7c7b\u4f3c\u6309\u94ae\uff09\uff0c\u586b\u5199\u4ee5\u4e0b\u4fe1\u606f\uff1a\n<ul>\n<li><strong>Name<\/strong>\uff1a<code>Qwen3.8-27B (Local)<\/code><\/li>\n<li><strong>API Base URL<\/strong>\uff1a<code>http:\/\/127.0.0.1:8080\/v1<\/code>\uff08\u6ce8\u610f\u672b\u5c3e\u7684 <code>\/v1<\/code>\uff09<\/li>\n<li><strong>API Key<\/strong>\uff1a\u7559\u7a7a\u6216\u8f93\u5165\u4efb\u610f\u5360\u4f4d\u7b26\uff08\u5982 <code>dummy<\/code>\uff09<\/li>\n<li><strong>Model ID<\/strong>\uff1a<code>qwen3.8-27b-q4_k_m<\/code><\/li>\n<\/ul>\n<\/li>\n<li>\u70b9\u51fb <strong>Save<\/strong> \u4fdd\u5b58<\/li>\n<\/ol>\n<p>\u4e4b\u540e\u5728\u804a\u5929\u754c\u9762\u9009\u62e9\u8be5\u6a21\u578b\u5373\u53ef\u4f7f\u7528\u672c\u5730 Qwen \u8fdb\u884c\u5bf9\u8bdd\u3002<\/p>\n<div class=\"doc-note doc-note--info\">\n  \ud83d\udca1 \u5982\u679c llama-server \u672a\u542f\u52a8\uff0cHarness \u65e0\u6cd5\u8fde\u63a5\u3002\u53ef\u5728 Harness \u4e2d\u540c\u65f6\u4fdd\u7559 DeepSeek \u4e91\u7aef\u6a21\u578b\u548c\u672c\u5730\u6a21\u578b\uff0c\u6309\u9700\u5207\u6362\u3002\n<\/div>\n<hr>\n<h2 id=\"sec10\">10. \u521b\u5efa Harness \u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/h2>\n<h3>\u6e90\u7801\u5b89\u88c5\u7248<\/h3>\n<p>\u5728\u684c\u9762\u7a7a\u767d\u5904\u53f3\u952e \u2192 <strong>\u65b0\u5efa<\/strong> \u2192 <strong>\u6587\u672c\u6587\u6863<\/strong>\uff0c\u547d\u540d\u4e3a <code>dsh-web.bat<\/code>\uff0c\u53f3\u952e\u7f16\u8f91\uff0c\u7c98\u8d34\uff1a<\/p>\n<pre class=\"prettyprint linenums\" ><code>@echo off\r\ncd \/d D:\\dev\\llm\\deepseek-harness\r\npnpm dsh web\r\npause<\/code><\/pre>\n<p>\u53f3\u952e <code>dsh-web.bat<\/code> \u2192 <strong>\u53d1\u9001\u5230<\/strong> \u2192 <strong>\u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/strong>\uff0c\u53cc\u51fb\u5373\u53ef\u542f\u52a8 Harness Web UI\u3002<\/p>\n<h3>\u684c\u9762\u5e94\u7528\u7248<\/h3>\n<p>\u5b89\u88c5\u5b8c\u6210\u540e\u4f1a\u81ea\u52a8\u5728\u684c\u9762\u548c\u5f00\u59cb\u83dc\u5355\u521b\u5efa\u5feb\u6377\u65b9\u5f0f\uff0c\u76f4\u63a5\u53cc\u51fb\u542f\u52a8\u5373\u53ef\u3002<\/p>\n<hr>\n<h2 id=\"sec11\">11. \u5e38\u89c1\u95ee\u9898<\/h2>\n<table class=\"doc-table\">\n<thead>\n<tr>\n<th>\u62a5\u9519\u73b0\u8c61<\/th>\n<th>\u53ef\u80fd\u539f\u56e0<\/th>\n<th>\u89e3\u51b3\u65b9\u6848<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>nvcc fatal: Unsupported gpu architecture 'sm_120'<\/code><\/td>\n<td>CUDA \u7248\u672c &lt; 12.8<\/td>\n<td>\u5347\u7ea7 CUDA \u81f3 12.8 \u6216\u66f4\u9ad8<\/td>\n<\/tr>\n<tr>\n<td><code>llama-server.exe<\/code> \u95ea\u9000\u6216\u62a5\u9519<\/td>\n<td>CUDA \u7248\u672c\u4e0d\u5339\u914d\u6216\u7f3a\u5c11 DLL<\/td>\n<td>\u786e\u8ba4\u4e0b\u8f7d\u4e86\u914d\u5957 DLL \u5305\uff0c\u5c06\u6240\u6709 <code>.dll<\/code> \u6587\u4ef6\u653e\u5165 <code>llama-server.exe<\/code> \u540c\u7ea7\u76ee\u5f55<\/td>\n<\/tr>\n<tr>\n<td><code>CUDA error: out of memory<\/code><\/td>\n<td>\u6a21\u578b + KV Cache \u8d85\u51fa 24GB<\/td>\n<td>\u964d\u4f4e <code>--ctx-size<\/code>\uff08\u5982 16384\uff09\u6216 <code>--batch-size<\/code>\uff08\u5982 2048\uff09<\/td>\n<\/tr>\n<tr>\n<td>\u7aef\u53e3 8080 \u88ab\u5360\u7528<\/td>\n<td>\u5176\u4ed6\u8fdb\u7a0b\u4f7f\u7528\u8be5\u7aef\u53e3<\/td>\n<td>\u66f4\u6539 <code>--port<\/code> \u4e3a\u5176\u4ed6\u7aef\u53e3\uff08\u5982 8081\uff09<\/td>\n<\/tr>\n<tr>\n<td><code>pnpm<\/code> \u547d\u4ee4\u65e0\u6cd5\u8bc6\u522b<\/td>\n<td>pnpm \u672a\u6b63\u786e\u5b89\u88c5\u6216\u672a\u52a0\u5165 PATH<\/td>\n<td><code>npm install -g pnpm<\/code> \u540e\u91cd\u542f\u7ec8\u7aef<\/td>\n<\/tr>\n<tr>\n<td><code>git<\/code> \u547d\u4ee4\u65e0\u6cd5\u8bc6\u522b<\/td>\n<td>\u672a\u5b89\u88c5 Git<\/td>\n<td>\u5b89\u88c5 <a href=\"https:\/\/git-scm.com\/download\/win\" target=\"_blank\">Git for Windows<\/a><\/td>\n<\/tr>\n<tr>\n<td>Harness \u5f39\u51fa API Key \u5bf9\u8bdd\u6846<\/td>\n<td>\u9996\u6b21\u542f\u52a8\u9ed8\u8ba4\u8981\u6c42\u4e91\u7aef Key<\/td>\n<td>\u76f4\u63a5\u5173\u95ed\u5bf9\u8bdd\u6846\uff0c\u7528\u672c\u5730\u6a21\u578b\u5373\u53ef<\/td>\n<\/tr>\n<tr>\n<td>Harness \u8fde\u63a5\u4e0d\u4e0a Qwen<\/td>\n<td>llama-server \u672a\u542f\u52a8\u6216\u5730\u5740\u9519\u8bef<\/td>\n<td>\u786e\u8ba4\u670d\u52a1\u5df2\u542f\u52a8\uff0c\u5730\u5740\u4e3a <code>:8080\/v1<\/code><\/td>\n<\/tr>\n<tr>\n<td>\u684c\u9762\u5feb\u6377\u65b9\u5f0f\u65e0\u53cd\u5e94<\/td>\n<td>\u8def\u5f84\u9519\u8bef\u6216\u6743\u9650\u95ee\u9898<\/td>\n<td>\u68c0\u67e5 <code>.bat<\/code> \u6587\u4ef6\u4e2d\u7684\u8def\u5f84\u662f\u5426\u6b63\u786e<\/td>\n<\/tr>\n<tr>\n<td>\u6a21\u578b\u4e0b\u8f7d\u4e2d\u65ad<\/td>\n<td>\u7f51\u7edc\u4e0d\u7a33\u5b9a<\/td>\n<td>\u4f7f\u7528 <code>curl -C -<\/code> \u6216 <code>wget -c<\/code> \u7eed\u4f20<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<hr>\n<h2>\u4e0e Linux \u7248\u672c\u7684\u4e3b\u8981\u5dee\u5f02<\/h2>\n<table class=\"doc-table\">\n<thead>\n<tr>\n<th>\u5bf9\u6bd4\u9879<\/th>\n<th>Linux (Ubuntu)<\/th>\n<th>Windows 11<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><strong>llama.cpp \u5b89\u88c5<\/strong><\/td>\n<td>\u6e90\u7801\u7f16\u8bd1\uff08git clone + cmake + make\uff09<\/td>\n<td>\u76f4\u63a5\u4e0b\u8f7d\u9884\u7f16\u8bd1\u5305\uff08.zip \u89e3\u538b\u5373\u7528\uff09<\/td>\n<\/tr>\n<tr>\n<td><strong>\u7f16\u8bd1\u4f9d\u8d56<\/strong><\/td>\n<td>build-essential, cmake, git<\/td>\n<td>\u65e0\u9700\u5b89\u88c5\uff08\u4ec5\u9700 Git \u7528\u4e8e Harness\uff09<\/td>\n<\/tr>\n<tr>\n<td><strong>\u542f\u52a8\u811a\u672c<\/strong><\/td>\n<td><code>.sh<\/code>\uff08Shell \u811a\u672c\uff09<\/td>\n<td><code>.bat<\/code>\uff08\u6279\u5904\u7406\u6587\u4ef6\uff09<\/td>\n<\/tr>\n<tr>\n<td><strong>\u8def\u5f84\u683c\u5f0f<\/strong><\/td>\n<td><code>\/home\/ai\/...<\/code><\/td>\n<td><code>D:\\dev\\llm\\...<\/code><\/td>\n<\/tr>\n<tr>\n<td><strong>CUDA \u5b89\u88c5<\/strong><\/td>\n<td>APT \u6e90\u5b89\u88c5<\/td>\n<td>\u5b98\u7f51\u4e0b\u8f7d .exe \u5b89\u88c5\u5305<\/td>\n<\/tr>\n<tr>\n<td><strong>DeepSeek Harness<\/strong><\/td>\n<td>\u6e90\u7801\u5b89\u88c5\uff08Node.js + pnpm\uff09<\/td>\n<td>\u6e90\u7801\u5b89\u88c5 \u6216 \u684c\u9762\u5e94\u7528\uff08.exe\uff09<\/td>\n<\/tr>\n<tr>\n<td><strong>\u684c\u9762\u5feb\u6377\u65b9\u5f0f<\/strong><\/td>\n<td><code>.desktop<\/code> \u6587\u4ef6<\/td>\n<td><code>.bat<\/code> + \u5feb\u6377\u65b9\u5f0f<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<hr>\n<p>\ud83c\udf89 <strong>\u606d\u559c\uff01<\/strong> \u4f60\u7684 Windows 11 + RTX 5090 + Qwen3.8-27B + DeepSeek Harness \u73af\u5883\u5df2\u914d\u7f6e\u5b8c\u6210\u3002<\/p>\n<p><strong>\u65e5\u5e38\u4f7f\u7528\u4e09\u6b65\u8d70\uff1a<\/strong><\/p>\n<ol>\n<li>\u53cc\u51fb\u684c\u9762 <code>run-qwen.bat<\/code> \u5feb\u6377\u65b9\u5f0f \u2192 \u542f\u52a8 Qwen API \u670d\u52a1\uff08<code>8080<\/code> \u7aef\u53e3\uff09<\/li>\n<li>\u53cc\u51fb\u684c\u9762 <code>DeepSeek Harness<\/code> \u56fe\u6807\uff08\u6e90\u7801\u7248\u4e3a <code>dsh-web.bat<\/code>\uff09\u2192 \u542f\u52a8 Web UI\uff08<code>3080<\/code> \u7aef\u53e3\uff09<\/li>\n<li>\u6d4f\u89c8\u5668\u8bbf\u95ee <code>http:\/\/127.0.0.1:3080<\/code> \u2192 \u9009\u62e9\u672c\u5730\u6a21\u578b\u5f00\u59cb\u5bf9\u8bdd<\/li>\n<\/ol>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>RTX 5090 24GB + Qwen3.8-27B Q4_K_M + DeepSeek Harness \u672c\u6559\u7a0b\u9762\u5411 Windows 11 \u7cfb\u7edf\uff0c\u663e\u5361\u4e3a NVIDIA RTX 5090 24GB\u3002\u5047\u8bbe\u4f60\u5df2\u5b8c\u6210 Windows 11 \u7cfb\u7edf\u5b89\u88c5\u548c NVIDIA \u663e\u5361\u9a71\u52a8\u5b89\u88c5\uff08\u9a71\u52a8\u7248\u672c \u2265 570\uff09\u3002\u6211\u4eec\u5c06\u4ece CUDA \u5de5\u5177\u5305\u5b89\u88c5\u5f00\u59cb\uff0c\u9010\u6b65\u5b8c\u6210 llama.cpp \u9884\u7f16\u8bd1\u5305\u90e8\u7f72\u3001Qwen3.8-27B [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[18,12],"tags":[31,29,27,28,30,23,26],"class_list":["post-77","post","type-post","status-publish","format-standard","hentry","category-local-ai-windows","category-local-ai","tag-cuda","tag-deepseek-harness","tag-llama-cpp","tag-qwen","tag-rtx-5090","tag-windows-11","tag-local-llm"],"_links":{"self":[{"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/posts\/77","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=77"}],"version-history":[{"count":1,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/posts\/77\/revisions"}],"predecessor-version":[{"id":118,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=\/wp\/v2\/posts\/77\/revisions\/118"}],"wp:attachment":[{"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=77"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=77"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wp.jyy7998.tk\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=77"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}