{"id":4151,"date":"2025-04-02T00:11:45","date_gmt":"2025-04-01T16:11:45","guid":{"rendered":"https:\/\/10ccc.cc\/?p=4151"},"modified":"2026-01-03T18:42:14","modified_gmt":"2026-01-03T10:42:14","slug":"%e6%a8%a1%e5%9e%8b%e5%be%ae%e8%b0%83%ef%bc%9a%e4%bd%bf%e7%94%a8-unsloth-%e5%be%ae%e8%b0%83-deepseek_r1","status":"publish","type":"post","link":"https:\/\/10ccc.cc\/?p=4151","title":{"rendered":"\u6a21\u578b\u5fae\u8c03\uff1a\u4f7f\u7528 Unsloth \u5fae\u8c03 DeepSeek_R1"},"content":{"rendered":"

\u91cd\u8981\u58f0\u660e<\/h1>\n

\u5f53\u65f6\u5c1d\u8bd5\u5fae\u8c03\u6a21\u578b\u662f\u4e3a\u4e86\u4e86\u89e3\u3010\u7279\u5b9a\u9886\u57df\u6a21\u578b\u5fae\u8c03\u7684\u5b9e\u9645\u5de5\u4f5c\u6d41\u7a0b\u3011\uff0c\u5728\u8fdb\u884c\u5b9e\u9645\u7684\u5fae\u8c03\u5de5\u4f5c\u65f6\uff0c\u4e3a\u4e86\u65b9\u4fbf\u548c\u7f51\u53cb\u4ea4\u6d41\uff0c\u56e0\u6b64 jupyter notebook \u4e2d\u5747\u4f7f\u7528\u82f1\u6587\u3002\u4f46\u9063\u8bcd\u9020\u53e5\u5747\u6bd4\u8f83\u57fa\u7840\uff0c\u5982\u679c\u9605\u8bfb\u5b9e\u5728\u56f0\u96be\uff0c\u53ef\u4ee5\u4f7f\u7528\u7ffb\u8bd1\u63d2\u4ef6\u3002\u5c31\u6539\u56de\u6765\u4e86\uff0c\u5b9e\u5728\u5b9e\u5728\u538c\u6076\u91cd\u590d\u5de5\u4f5c\u3002<\/p><\/blockquote>\n

 <\/p>\n

\u51c6\u5907\u5de5\u4f5c<\/h1>\n

\u9996\u5148\u58f0\u660e\u4e00\u4e9b\u91cd\u8981\u4e14\u5fc5\u987b\u8981\u505a\u7684\u51c6\u5907\u57fa\u7840\u5de5\u4f5c\u5982\u4e0b\u3002\u4e14\u9700\u8981\u6ce8\u610f\uff0c\u5728\u5b89\u88c5\u8f6f\u4ef6\u65f6\uff0c\u9700\u8981\u4e25\u683c\u6ce8\u610f\u3010\u8f6f\u4ef6\u7248\u672c\u3011\uff0c\u4e4b\u524d\u8bd5\u8fc7\u7531\u4e8e\u8f6f\u4ef6\u7248\u672c\u4e0d\u540c\u5bfc\u81f4\u7684\u8ba1\u7b97\u6570\u503c\u5dee\u5f02\uff1a<\/p>\n

    \n
  1. Nvidia Driver <\/strong>and Cuda toolkit<\/strong>: choose a proper driver for your Nvidia GPU <\/strong>and installcuda 12.4 toolkit<\/a><\/strong><\/li>\n
  2. Anaconda<\/strong>: we need this to create virtual environment <\/strong>with python3.10<\/strong>, and install PyTorch <\/strong>and Unsloth<\/strong><\/li>\n
  3. git <\/strong>and cmake<\/strong>: when we save finetuned model into gguf <\/strong>format, we need to callsave_pretrained_gguf()method and the method will clonellama.cppfrom github and compile it. you can download it from thegit offical site<\/a>andcmake offical site<\/a>and append their bin folders after windows'path<\/li>\n
  4. ollama <\/strong>and Page Assist<\/strong>: we use ollama <\/strong>to run finetuned model and Page Assist <\/strong>for a friendly UI. You can download ollama ->here<\/a>and install Page Assist from ->here<\/a><\/li>\n
  5. Triton wheel for python 3.10<\/strong>: if you are using a windows pc<\/strong>, you will need to download thetriton-windows-builds<\/a>and install it later viapip install triton-3.0.0-cp310-cp310-win_amd64.whl<\/li>\n<\/ol>\n

     <\/p>\n

    Jupyter Note \u7b14\u8bb0\u5185\u5bb9\uff1a<\/h1>\n

    \u7b14\u8bb0\u4e0b\u8f7d\uff1a\r\n<\/i>Download<\/span><\/a><\/p>\n

    \u6211\u5728\u7b14\u8bb0\u4e2d\u5df2\u7ecf\u5199\u5f97\u975e\u5e38\u8be6\u7ec6\uff0c\u5177\u4f53\u5185\u5bb9\u8bf7\u3010\u5c55\u5f00\u3011\u540e\u8be6\u7ec6\u9605\u8bfb\uff1a<\/p>\n

    \n