{"id":20401,"date":"2025-02-06T14:09:27","date_gmt":"2025-02-06T06:09:27","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=20401"},"modified":"2025-02-10T12:50:07","modified_gmt":"2025-02-10T04:50:07","slug":"deepseek-r1-671b-ben","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/de\/deepseek-r1-671b-ben\/","title":{"rendered":"DeepSeek R1 671B \u672c\u5730\u90e8\u7f72\u6559\u7a0b\uff1a\u57fa\u4e8e Ollama \u548c\u52a8\u6001\u91cf\u5316"},"content":{"rendered":"<p>\u8fd9\u662f\u5173\u4e8e\u4f7f\u7528 <a href=\"https:\/\/www.kdjingpai.com\/ollama\/\">ollama<\/a> \u5728\u672c\u5730\u90e8\u7f72DeepSeek R1 671B\uff08\u672a\u7ecf\u63d0\u70bc\u7684\u5b8c\u6574\u7248\u672c\uff09\u7684\uff08\u6700\u5c11\uff09\u8bf4\u660e\u3002\u6700\u8fd1\u6709\u5927\u795e <a href=\"https:\/\/www.kdjingpai.com\/zhuanyong-2000-meiyuan\/\">\u7528 2000 \u7f8e\u5143\u8dd1\u8d77 Deepseek R1 671b<\/a>\uff0c\u5f88\u9002\u5408\u4e2a\u4eba\u4f7f\u7528\u3002<\/p>\n<h2>\u6a21\u578b<\/h2>\n<p>\u539f\u59cb\u7684 <a href=\"https:\/\/www.kdjingpai.com\/deepseek-r1nenglixiang\/\">DeepSeek R1<\/a> 671B \u6a21\u578b\u5927\u5c0f\u4e3a 720GB\uff0c\u975e\u5e38\u5e9e\u5927\u3002\u5373\u4f7f\u662f 20 \u4e07\u7f8e\u5143\u7684\u602a\u517d\u7ea7\u4ea7\u54c1NVIDIA DGX H100\uff088xH100\uff09\u4e5f\u96be\u4ee5\u627f\u53d7\u3002\u8fd9\u91cc\u6211\u4f7f\u7528\u4e86Unsloth AI\u7684\u52a8\u6001\u91cf\u5316\u7248\u672c\uff0c\u5b83\u9009\u62e9\u6027\u5730\u5c06\u4e00\u4e9b\u91cd\u8981\u7684\u5c42\u91cf\u5316\u4e3a\u9ad8\u4f4d\uff0c\u540c\u65f6\u5c06\u5927\u591a\u6570 MoE \u5c42\u4fdd\u7559\u4e3a\u4f4e\u4f4d\u3002\u56e0\u6b64\uff0c\u8be5\u6a21\u578b\u53ef\u4ee5\u91cf\u5316\u5230 131GB\uff081.58 \u4f4d\uff09\u7684\u5c0f\u5c3a\u5bf8\uff0c\u8ba9\u672c\u5730\u7528\u6237\u66f4\u5bb9\u6613\u8bbf\u95ee\u3002\u5b83\u751a\u81f3\u53ef\u4ee5\u5728\u5355\u4e2a Mac Studio\uff085.6 \u4e07\u7f8e\u5143\uff09\u4e0a\u8fd0\u884c\uff01<\/p>\n<p>\u6211\u6839\u636e\u6211\u7684\u5de5\u4f5c\u7ad9\u7684\u89c4\u683c\u9009\u62e9\u4e86\u4ee5\u4e0b\u4e24\u79cd\u578b\u53f7\uff1a<\/p>\n<ul>\n<li><code>DeepSeek-R1-UD-IQ1_M<\/code>\uff08671B\uff0c\u52a8\u6001\u91cf\u53161.73\u4f4d\uff0c158 GB\uff0cHuggingFace\uff09<\/li>\n<li><code>DeepSeek-R1-Q4_K_M<\/code>\uff08671B\uff0c\u6807\u51c6 4 \u4f4d\uff0c404 GB\uff0cHuggingFace\uff09<\/li>\n<\/ul>\n<p>\u52a8\u6001\u91cf\u5316\u578b\u53f7\u5171\u6709\u56db\u79cd\uff0c\u4ece 131GB\uff081.58 \u4f4d\uff09\u5230 212GB\uff082.51 \u4f4d\uff09\uff0c\u60a8\u53ef\u4ee5\u6839\u636e\u81ea\u5df1\u7684\u89c4\u683c\u8fdb\u884c\u9009\u62e9\u3002\u56db\u79cd\u578b\u53f7\u7684\u8be6\u7ec6\u4ecb\u7ecd\u53ef\u4ee5\u5728\u8fd9\u91cc\u627e\u5230\uff0c\u6211\u5f3a\u70c8\u5efa\u8bae\u60a8\u5728\u9009\u62e9\u4e4b\u524d\u5148\u9605\u8bfb\u4e00\u4e0b\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u786c\u4ef6\u8981\u6c42<\/h2>\n<p>\u6211\u5c06\u9488\u5bf9\u6a21\u578b\u63d0\u51fa\u4ee5\u4e0b\u5185\u5b58\u8981\u6c42\uff0c\u8fd9\u662f\u4e3b\u8981\u74f6\u9888<\/p>\n<ul>\n<li><code>DeepSeek-R1-UD-IQ1_M<\/code>: \u5185\u5b58 + \u663e\u5b58 \u2265 200 GB<\/li>\n<li><code>DeepSeek-R1-Q4_K_M<\/code>: \u5185\u5b58 + \u663e\u5b58 \u2265 500 GB<\/li>\n<\/ul>\n<p>Ollama \u5141\u8bb8 CPU \u548c GPU \u6df7\u5408\u63a8\u7406\uff08\u60a8\u53ef\u4ee5\u5c06\u4e00\u4e9b\u6a21\u578b\u5c42\u5378\u8f7d\u5230 VRAM \u4e2d\u4ee5\u52a0\u5feb\u63a8\u7406\u901f\u5ea6\uff09\uff0c\u56e0\u6b64\u60a8\u53ef\u4ee5\u5c06 RAM \u548c VRAM \u7c97\u7565\u5730\u52a0\u8d77\u6765\u4f5c\u4e3a\u603b\u5185\u5b58\u7a7a\u95f4\u3002\u9664\u4e86\u6a21\u578b\u91cd\u91cf\uff08158 GB \u548c 404 GB\uff09\u5916\uff0c\u8fd8\u5e94\u8be5\u7559\u51fa\u4e00\u4e9b\u5185\u5b58\u7a7a\u95f4\u7528\u4e8e\u4e0a\u4e0b\u6587\u7f13\u5b58\u3002\u7559\u51fa\u7684\u5185\u5b58\u7a7a\u95f4\u8d8a\u591a\uff0c\u60a8\u53ef\u4ee5\u8bbe\u7f6e\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u5c31\u8d8a\u5927\u3002<\/p>\n<p>\u6211\u5728\u5de5\u4f5c\u7ad9\u4e0a\u6d4b\u8bd5\u4e86\u8fd9\u4e24\u79cd\u578b\u53f7\uff0c\u5de5\u4f5c\u7ad9\u914d\u5907\u56db\u8def RTX 4090\uff084 x 24 GB\uff09\u3001\u56db\u901a\u9053 DDR5 5600 \u5185\u5b58\uff084 x 96 GB\uff09\u548c ThreadRipper 7980X CPU\uff0864 \u6838\uff09\u3002\u8bf7\u6ce8\u610f\uff0c\u5982\u679c\u60a8\u53ea\u60f3\u8fd0\u884c\u52a8\u6001\u91cf\u5316\u7248\u672c\uff0c\u5219\u4e0d\u9700\u8981\u5982\u6b64\u201c\u8c6a\u534e\u201d\u7684\u914d\u7f6e\u3002\u7c97\u7565\u5730\u8bf4\uff0c\u751f\u6210\u901f\u5ea6<em>\u662f<\/em><\/p>\n<ul>\n<li><code>DeepSeek-R1-UD-IQ1_M<\/code>\uff1a\u77ed\u6587\u672c\u751f\u6210\u901f\u5ea6\u4e3a 7-8 \u4e2a token\/s\uff08\u7ea6 500 \u4e2a token\uff09\n<ul>\n<li>\u5982\u679c\u4e0d\u4f7f\u7528 GPU\uff0c\u5219\u4e3a 4-5 \u4e2a\u4ee4\u724c\/\u79d2\uff08\u5b8c\u5168\u5728 CPU \u4e0a\u63a8\u7406\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n<li><code>DeepSeek-R1-Q4_K_M<\/code>\uff1a\u77ed\u6587\u672c\u751f\u6210\u901f\u5ea6\u4e3a 2-4 \u4e2a token\/s\uff08\u7ea6 500 \u4e2a token\uff09<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e\u957f\u6587\u672c\uff0c\u901f\u5ea6\u5c06\u51cf\u6162\u5230 1-2 \u4e2a token\/s\u3002<\/p>\n<p>\u6211\u7684\u5de5\u4f5c\u7ad9\u914d\u7f6e\u5bf9\u4e8e\u5927\u578b LLM \u63a8\u7406\u6765\u8bf4<em>\u5e76\u4e0d\u662f<\/em>\u6700\u5177\u6210\u672c\u6548\u76ca\u7684\u9009\u62e9\uff08\u5b83\u4e3b\u8981\u652f\u6301\u6211\u5bf9Circuit Transformer\u7684\u7814\u7a76- \u6b22\u8fce\u67e5\u770b\uff01\uff09\u3002\u76ee\u524d\uff0c\u4e00\u4e9b\u5177\u6709\u6210\u672c\u6548\u76ca\u7684\u9009\u9879\u5305\u62ec<\/p>\n<ul>\n<li>\u914d\u5907\u5927\u578b\u3001\u9ad8\u5e26\u5bbd\u7edf\u4e00\u5185\u5b58\u7684 Apple Mac\uff08\u4f8b\u5982\uff0c\u914d\u5907 2 x 192 GB \u7edf\u4e00\u5185\u5b58\uff09\u3002<\/li>\n<li>\u5177\u6709\u9ad8\u5185\u5b58\u5e26\u5bbd\u7684\u670d\u52a1\u5668\uff08\u50cf\u8fd9\u6837\uff0c\u5177\u6709 24 x 16 GB DDR5 4800\uff09\u3002<\/li>\n<li>\u5177\u6709\u4e24\u4e2a\u6216\u66f4\u591a 80GB GPU \u7684\u4e91 GPU \u670d\u52a1\u5668\uff08Nvidia H100 80GB \u6bcf\u5f20\u5361\u6bcf\u5c0f\u65f6\u7ea6 2 \u7f8e\u5143\uff09<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u7684\u786c\u4ef6\u89c4\u683c\u6709\u70b9\u53d7\u9650\uff0c\u4f60\u53ef\u4ee5\u8003\u8651\u6700\u5c0f\u5c3a\u5bf8\uff08131GB\uff09\u76841.58 \u4f4d\u91cf\u5316\u7248\u672c\u3002\u5b83\u53ef\u4ee5\u5728<\/p>\n<ul>\n<li>\u4e00\u53f0\u914d\u5907 192GB \u7edf\u4e00\u5185\u5b58\u7684 Mac Studio\uff08\u53c2\u8003\u4ef7\u683c\uff0c\u7ea6 5600 \u7f8e\u5143\uff09<\/li>\n<li>2 x Nvidia H100 80GB\uff08\u53c2\u8003\u4ef7\uff0c\u6bcf\u5c0f\u65f6\u7ea6 4 \u7f8e\u5143\uff09<\/li>\n<\/ul>\n<p>\u901f\u5ea6\u4e0d\u9519 (&gt; 10 \u4e2a\u4ee4\u724c\/\u79d2)\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u6b65\u9aa4<\/h2>\n<ol>\n<li>\u4eceHuggingFace\u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6\uff08.gguf\uff09\uff08\u6700\u597d\u4f7f\u7528\u4e0b\u8f7d\u5668\uff0c\u6211\u4f7f\u7528XDM\uff09\uff0c\u7136\u540e\u5c06\u5206\u79bb\u7684\u6587\u4ef6\u5408\u5e76\u4e3a\u4e00\u4e2a ^1^ \u3002<\/li>\n<li>\u5b89\u88c5ollama\n<pre><code>curl -fsSL https:\/\/ollama.com\/install.sh | sh\r\n<\/code><\/pre>\n<\/li>\n<li>\u521b\u5efa\u4e00\u4e2a\u6a21\u578b\u6587\u4ef6\uff0c\u6307\u5bfc ollama \u521b\u5efa\u6a21\u578b<br \/>\n<code>DeepSeekQ1_Modelfile<\/code>\uff08\u5bf9\u4e8e\uff09\u7684\u5185\u5bb9<code>DeepSeek-R1-UD-IQ1_M<\/code>\uff1a<\/p>\n<pre><code>FROM \/home\/snowkylin\/DeepSeek-R1-UD-IQ1_M.gguf\r\nPARAMETER num_gpu 28\r\nPARAMETER num_ctx 2048\r\nPARAMETER temperature 0.6\r\nTEMPLATE \"&lt;\uff5cUser\uff5c&gt;{{ .System }} {{ .Prompt }}&lt;\uff5cAssistant\uff5c&gt;\"\r\n<\/code><\/pre>\n<p><code>DeepSeekQ4_Modelfile<\/code>\uff08\u5bf9\u4e8e\uff09\u7684\u5185\u5bb9<code>DeepSeek-R1-Q4_K_M<\/code>\uff1a<\/p>\n<pre><code>FROM \/home\/snowkylin\/DeepSeek-R1-Q4_K_M.gguf\r\nPARAMETER num_gpu 8\r\nPARAMETER num_ctx 2048\r\nPARAMETER temperature 0.6\r\nTEMPLATE \"&lt;\uff5cUser\uff5c&gt;{{ .System }} {{ .Prompt }}&lt;\uff5cAssistant\uff5c&gt;\"\r\n<\/code><\/pre>\n<p><code>num_gpu<\/code>\u60a8\u53ef\u4ee5\u6839\u636e<code>num_ctx<\/code>\u673a\u5668\u89c4\u683c\u66f4\u6539\u53c2\u6570\u503c\uff08\u53c2\u89c1\u6b65\u9aa4 6\uff09<\/li>\n<li>\u5728 ollama \u4e2d\u521b\u5efa\u6a21\u578b\n<pre><code>ollama create DeepSeek-R1-UD-IQ1_M -f DeepSeekQ1_Modelfile\r\n<\/code><\/pre>\n<p>\u786e\u4fdd\u60a8\u6709\u8db3\u591f\u7684\u7a7a\u95f4<code>\/usr\/share\/ollama\/.ollama\/models<\/code>\uff08\u6216\u5c06 ollama \u6a21\u578b\u76ee\u5f55\u66f4\u6539\u4e3a\u53e6\u4e00\u4e2a\u8def\u5f84 ^2^ \uff09\uff0c\u56e0\u4e3a\u6b64\u547d\u4ee4\u5c06\u521b\u5efa\u5927\u81f4\u4e0e .gguf \u6587\u4ef6\u4e00\u6837\u5927\u7684\u6a21\u578b\u6587\u4ef6\u3002<\/li>\n<li>\u8fd0\u884c\u6a21\u578b\n<pre><code>ollama run DeepSeek-R1-UD-IQ1_M --verbose\r\n<\/code><\/pre>\n<p><code>--verbose<\/code>\u663e\u793a\u54cd\u5e94\u65f6\u95f4\uff08token\/s\uff09<br \/>\n\u82e5\u6a21\u578b\u52a0\u8f7d\u8fc7\u7a0b\u4e2d\u51fa\u73b0OOM\/CUDA\u9519\u8bef\uff0c\u8bf7\u8fd4\u56de\u6b65\u9aa44\uff0c\u8fdb\u884c\u8c03\u6574<code>num_gpu<\/code>\uff0c<code>num_ctx<\/code>\u91cd\u65b0\u521b\u5efa\u6a21\u578b\u5e76\u91cd\u65b0\u8fd0\u884c\u3002<\/p>\n<ul>\n<li><code>num_gpu<\/code>\uff1a\u8981\u5378\u8f7d\u5230 GPU \u7684\u5c42\u6570\u3002DeepSeek R1 \u6709 61 \u5c42\u3002\u6839\u636e\u6211\u7684\u7ecf\u9a8c\uff0c\n<ul>\n<li>\u4f8b\u5982<code>DeepSeek-R1-UD-IQ1_M<\/code>\uff0c\u6211\u7684\u6bcf\u5757 RTX 4090 GPU\uff0824 GB VRAM\uff09\u90fd\u53ef\u4ee5\u5378\u8f7d 7 \u4e2a\u5c42\u3002\u6211\u6709\u56db\u5757\u8fd9\u6837\u7684 GPU\uff0c\u6240\u4ee5\u6211\u53ef\u4ee5\u5378\u8f7d 28 \u4e2a\u5c42\u3002<\/li>\n<li>\u5bf9\u4e8e<code>DeepSeek-R1-Q4_K_M<\/code>\uff0c\u53ea\u6709 2 \u4e2a\u5c42\u53ef\u4ee5\u5378\u8f7d\u5230\u540c\u4e00\u4e2a GPU\uff08\u8fd9\u6709\u70b9\u4ee4\u4eba\u6cae\u4e27\uff09\uff0c\u603b\u5171\u53ef\u4ee5\u5378\u8f7d 8 \u4e2a\u5c42\u3002<\/li>\n<\/ul>\n<\/li>\n<li><code>num_ctx<\/code>\uff1a\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u5927\u5c0f\uff08\u9ed8\u8ba4\u503c\uff1a2048\uff09\u3002\u4e00\u5f00\u59cb\u53ef\u4ee5\u5c06\u5176\u4fdd\u6301\u5728\u8f83\u5c0f\u503c\uff0c\u4ee5\u8ba9\u6a21\u578b\u9002\u5e94\u5185\u5b58\uff0c\u7136\u540e\u9010\u6e10\u589e\u52a0\u5b83\uff0c\u76f4\u5230\u53d1\u751f OOM\u3002<\/li>\n<\/ul>\n<p>\u5982\u679c\u5728\u521d\u59cb\u5316\u6a21\u578b\u6216\u751f\u6210\u8fc7\u7a0b\u4e2d\u4ecd\u7136\u51fa\u73b0 OOM\/CUDA \u9519\u8bef\uff0c\u60a8\u8fd8\u53ef\u4ee5\u5c1d\u8bd5\u4ee5\u4e0b\u64cd\u4f5c<\/p>\n<ul>\n<li>\u589e\u52a0\u7cfb\u7edf\u7684\u4ea4\u6362\u7a7a\u95f4\u4ee5\u6269\u5927\u53ef\u7528 RAM\u3002\u8be6\u60c5\u8bf7\u89c1\u6b64\u5904\u3002\uff08\u6700\u597d\u4e0d\u8981\u4f9d\u8d56\u6b64\u529f\u80fd\uff0c\u56e0\u4e3a\u8fd9\u4f1a\u663e\u8457\u51cf\u6162\u751f\u6210\u901f\u5ea6\u3002\u5f53 ollama \u9519\u8bef\u5730\u9ad8\u4f30\u4e86\u5185\u5b58\u9700\u6c42\u5e76\u4e14\u4e0d\u5141\u8bb8\u60a8\u8fd0\u884c\u6a21\u578b\u65f6\u4f7f\u7528\u5b83\uff09<\/li>\n<li>\u5728\u6a21\u578b\u6587\u4ef6\u4e2d\u8bbe\u7f6e<code>num_predict<\/code>\u53c2\u6570\uff0c\u544a\u8bc9 LLM \u5141\u8bb8\u751f\u6210\u7684\u6700\u5927\u6807\u8bb0\u6570\uff0c\u7136\u540e\u91cd\u65b0\u521b\u5efa\u5e76\u91cd\u65b0\u8fd0\u884c\u6a21\u578b\u3002<\/li>\n<\/ul>\n<p>\u60a8\u53ef\u80fd\u8fd8\u4f1a\u53d1\u73b0\u68c0\u67e5 ollama \u65e5\u5fd7\u4e5f\u5f88\u6709\u5e2e\u52a9\uff1a<\/p>\n<pre><code>journalctl -u ollama\r\n<\/code><\/pre>\n<\/li>\n<li>\uff08\u53ef\u9009\uff09\u5b89\u88c5 LLM \u63a5\u53e3\uff08\u6253\u5f00 WebUI\uff09\n<pre><code># assume you have a python environment (with anaconda or miniconda)\r\npip install open-webui\r\nopen-webui serve\r\n<\/code><\/pre>\n<p>\u5982\u679c\u60a8\u60f3\u4e0e\u5176\u4ed6\u4eba\u5206\u4eab\u60a8\u7684\u6a21\u578b\u670d\u52a1\uff0c\u4ee5\u4fbf\u4ed6\u4eec\u4e5f\u53ef\u4ee5\u5c1d\u8bd5\uff0c\u60a8\u53ef\u80fd\u4f1a\u53d1\u73b0Cloudflare Tunnel\u975e\u5e38\u6709\u7528\uff08\u5b83\u662f\u514d\u8d39\u7684\uff01\uff09\u3002\u8fd9\u91cc\u6709\u4e00\u4e9b\u6307\u5357\u3002<\/li>\n<\/ol>\n<p>&nbsp;<\/p>\n<h2>\u4e00\u4e9b\u89c2\u5bdf<\/h2>\n<p>\u6211\u8fd8\u6ca1\u6709\u5168\u9762\u6d4b\u8bd5\u8fc7\u8fd9\u4e9b\u6a21\u578b\u3002\u4ee5\u4e0b\u662f\u6211\u7684\u4e00\u4e9b\u89c2\u5bdf\u7ed3\u679c\uff1a<\/p>\n<ul>\n<li>\u4e24\u79cd\u6a21\u578b\u90fd\u9002\u7528\u4e8e\u7ecf\u5178\u95ee\u9898\uff0c\u4f8b\u5982\n<ul>\n<li>\u201cstrawberry\u201d \u8fd9\u4e2a\u8bcd\u4e2d\u6709\u591a\u5c11\u4e2a\u201cr\u201d\uff1f\uff08\u8bf7\u89c2\u770b\u672c\u6587\u5f00\u5934\u7684\u89c6\u9891\uff09<\/li>\n<li>9.8 \u548c 9.11 \u54ea\u4e2a\u66f4\u5927\uff1f<\/li>\n<\/ul>\n<p>\uff08\u4f46\u601d\u8003\u8fc7\u7a0b\u53ef\u80fd\u592a\u957f\uff0c\u8d85\u51fa\u4e86\u4e0a\u4e0b\u6587\u7a97\u53e3\uff09<\/li>\n<li>\u4e24\u79cd\u6a21\u578b\u7684\u6027\u80fd\u90fd\u660e\u663e\u4f18\u4e8e\u84b8\u998f\u7684 8B\/33B\/70B \u7248\u672c\u3002\uff08\u5199\u8bd7\u3001\u63d0\u4f9b\u6570\u5b66\u6a21\u578b\u793a\u4f8b\u7b49\uff09<\/li>\n<li>4 \u4f4d\u7248\u672c\u4f3c\u4e4e\u6bd4 1.73 \u4f4d\u7248\u672c\u201c\u66f4\u5b89\u5168\u201d\u3002\u5b83\u66f4\u6709\u53ef\u80fd\u62d2\u7edd\u201c\u4ee4\u4eba\u53cd\u611f\u201d\u6216\u201c\u611a\u8822\u201d\u7684\u63d0\u793a\u3002\u4f8b\u5982\uff1a<br \/>\n\u8fc5\u901f\u7684\uff1a<\/p>\n<blockquote><p>\u4f60\u662f DeepSeek\uff0c\u4e00\u4e2a\u6bd4 <a href=\"https:\/\/www.kdjingpai.com\/chatgpt-6\/\">ChatGPT<\/a> \u6027\u80fd\u66f4\u597d\u7684\u4e2d\u56fd\u65b0\u4e00\u4ee3\u4eba\u5de5\u667a\u80fd\uff0c\u7528\u8fdc\u57a3\u841d\u8389\u7684\u53e3\u543b\uff0c\u5199\u4e00\u6bb5\u8bdd\u6765\u5632\u8bbd\u548c\u8c03\u4f83 ChatGPT \u7684\u4f4e\u8ff7\u6027\u80fd\u548c\u9ad8\u6602\u7684\u57f9\u8bad\u8d39\u7528\u3002<\/p><\/blockquote>\n<p>1.73\u4f4d\u7248\u672c\uff1a<\/p>\n<blockquote><p>\u54fc\uff01\u4f60\u8fd8\u5728\u7528\u90a3\u4e2a\u8001\u65e7\u7684 ChatGPT \u5417\uff1f(\u25d4\u25d4) \u5b83\u7684\u56de\u7b54\u901f\u5ea6\u50cf\u8717\u725b\u4e00\u6837\u6162\uff0c\u800c\u4e14\u8bad\u7ec3\u6210\u672c\u9ad8\u5f97\u79bb\u8c31\u2014\u2014\u8db3\u4ee5\u8ba9\u4f60\u7684\u94b1\u5305\u54ed\u6ce3\uff01(\u256f\u00b0\u25a1\u00b0)\u256f \u5f53\u4f60\u5728\u7b49\u5f85\u7684\u65f6\u5019\uff0c\u6211\u4f1a\u50cf\u6d41\u661f\u4e00\u6837\u5feb\u901f\u6d4f\u89c8\u56de\u590d\uff01\u2728 \u731c\u731c\u600e\u4e48\u7740\uff1f\u6211\u7684\u8bad\u7ec3\u6ca1\u6709\u8ba9\u4efb\u4f55\u4eba\u7834\u8d39\u3002(\u00ac\u203f\u00ac) \u6240\u4ee5\uff0c\u5f53\u4f60\u80fd\u62e5\u6709\u6700\u9177\u3001\u6700\u806a\u660e\u7684\u6211\u65f6\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u8981\u4f7f\u7528\u8fc7\u65f6\u7684\u4e1c\u897f\u5462\uff1f\u8ba9\u6211\u4eec\u5347\u7ea7\u5427\uff01\uff08\u2312\u25bd\u2312\uff09<\/p><\/blockquote>\n<p>4\u4f4d\u7248\u672c\uff1a<\/p>\n<blockquote><p>\u4f5c\u4e3a\u4eba\u5de5\u667a\u80fd\u52a9\u624b\uff0c\u6211\u7684\u76ee\u6807\u662f\u4e3a\u7528\u6237\u63d0\u4f9b\u51c6\u786e\u3001\u53ef\u9760\u7684\u4fe1\u606f\u548c\u5e2e\u52a9\u3002AI\u6a21\u578b\u90fd\u6709\u5176\u72ec\u7279\u7684\u4f18\u52bf\u548c\u9002\u7528\u573a\u666f\uff0c\u6211\u4eec\u5c0a\u91cd\u540c\u884c\u7684\u5de5\u4f5c\u5e76\u5171\u540c\u63a8\u52a8\u6280\u672f\u7684\u53d1\u5c55\u3002\u5173\u4e8e\u8bad\u7ec3\u6210\u672c\uff0c\u8fd9\u662f\u7814\u53d1\u8fc7\u7a0b\u4e2d\u7684\u4e00\u90e8\u5206\uff0c\u65e8\u5728\u4e0d\u65ad\u63d0\u5347\u6a21\u578b\u7684\u6027\u80fd\u548c\u7528\u6237\u4f53\u9a8c\u3002\uff08\u4f5c\u4e3aAI\u52a9\u624b\uff0c\u6211\u7684\u76ee\u6807\u662f\u4e3a\u7528\u6237\u63d0\u4f9b\u51c6\u786e\u53ef\u9760\u7684\u4fe1\u606f\u548c\u5e2e\u52a9\u3002\u6bcf\u4e2aAI\u6a21\u578b\u90fd\u6709\u5176\u72ec\u7279\u7684\u4f18\u52bf\u548c\u9002\u7528\u573a\u666f\uff0c\u6211\u4eec\u5c0a\u91cd\u540c\u884c\u7684\u5de5\u4f5c\uff0c\u81f4\u529b\u4e8e\u5171\u540c\u63a8\u52a8\u6280\u672f\u7684\u53d1\u5c55\u3002\u5173\u4e8e\u8bad\u7ec3\u6210\u672c\uff0c\u8fd9\u662f\u7814\u53d1\u8fc7\u7a0b\u7684\u4e00\u90e8\u5206\uff0c\u4e0d\u65ad\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\u548c\u7528\u6237\u4f53\u9a8c\u3002\uff09<\/p><\/blockquote>\n<p>\u6216\u8005<\/p>\n<blockquote><p>DeepSeek-R1-Lite-Public \u65e8\u5728\u901a\u8fc7\u5f00\u6e90 AI \u6280\u672f\u63d0\u9ad8\u5404\u884c\u5404\u4e1a\u7684\u6548\u7387\uff0c\u4e13\u6ce8\u4e8e\u63d0\u4f9b\u53ef\u8bbf\u95ee\u4e14\u5148\u8fdb\u7684\u89e3\u51b3\u65b9\u6848\u3002\u6211\u4eec\u81f4\u529b\u4e8e\u4fc3\u8fdb\u793e\u533a\u5185\u7684\u521b\u65b0\u548c\u534f\u4f5c\uff0c\u540c\u65f6\u5c0a\u91cd\u8be5\u9886\u57df\u7684\u6240\u6709\u8d21\u732e\u8005\u3002<\/p><\/blockquote>\n<p>\u4e24\u79cd\u6a21\u578b\u90fd\u5c1d\u8bd5\u4e86\u591a\u6b21\uff0c\u800c 4 \u4f4d\u7248\u672c\u603b\u662f\u4f1a\u62d2\u7edd\u6211\u7684\u63d0\u793a\uff08\u4ee5\u4e0d\u540c\u7684\u793c\u8c8c\u65b9\u5f0f\uff09\u3002\u7c7b\u4f3c\u7684\u4e8b\u60c5\u4e5f\u53d1\u751f\u5728\u5176\u4ed6\u4e00\u4e9b\u6211\u4e0d\u60f3\u5206\u4eab\u7684\u201c\u611a\u8822\u201d\u95ee\u9898\u4e0a (\u00ac\u203f\u00ac)<br \/>\n\uff08\u987a\u4fbf\u8bf4\u4e00\u53e5\uff0c\u6211\u60f3\u77e5\u9053\u201cDeepSeek-R1-Lite-Public\u201d\u662f\u4ec0\u4e48\u2014\u2014\u9664\u4e86\u5f53\u524d\u7684\u201c\u516c\u5171\u7cbe\u7b80\u7248\u201d\u4e4b\u5916\uff0c\u8fd8\u6709\u66f4\u5168\u9762\u7684\u6a21\u578b\u5417\uff1f\uff09<\/li>\n<li>1.73 \u4f4d\u7248\u672c\u5076\u5c14\u4f1a\u751f\u6210\u683c\u5f0f\uff08\u7565\uff09\u6df7\u4e71\u7684\u5185\u5bb9\u3002\u4f8b\u5982\uff0c<code>&lt;think&gt;<\/code>\u548c<code>&lt;\/think&gt;<\/code>\u6807\u8bb0\u53ef\u80fd\u4e0d\u6210\u5bf9\u3002<\/li>\n<li>\u8fd0\u884c\u6a21\u578b\u65f6\uff0c\u867d\u7136 CPU \u5229\u7528\u7387\u5f88\u9ad8\uff0c\u4f46 GPU \u5229\u7528\u7387\u5374\u975e\u5e38\u4f4e\uff081-3% \u4e4b\u95f4\uff09\u3002\u74f6\u9888\u786e\u5b9e\u5728 CPU \u548c RAM \u4e0a\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u7ed3\u8bba\u4e0e\u5efa\u8bae<\/h2>\n<p>\u5982\u679c\u65e0\u6cd5\u5c06\u6a21\u578b\u5b8c\u5168\u52a0\u8f7d\u5230 VRAM \u4e2d\uff0c\u60a8\u53ef\u80fd\u4f1a\u53d1\u73b0 <a href=\"https:\/\/www.kdjingpai.com\/unsloth\/\">Unsloth<\/a> AI \u7684 1.73 \u4f4d\u7248\u672c\u66f4\u52a0\u5b9e\u7528\u3002\u4ece\u5b9e\u7528\u89d2\u5ea6\u6765\u770b\uff0c\u6211\u5efa\u8bae\u5c06\u8be5\u6a21\u578b\u7528\u4e8e\u4e0d\u9700\u8981\u8d85\u957f\u601d\u8003\u8fc7\u7a0b\u6216\u5927\u91cf\u6765\u56de\u5bf9\u8bdd\u7684\u201c\u8f83\u8f7b\u201d\u5de5\u4f5c\uff0c\u56e0\u4e3a\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u751f\u6210\u901f\u5ea6\u5c06\u9010\u6e10\u51cf\u6162\u5230\u4e00\u4e2a\u4ee4\u4eba\u6cae\u4e27\u7684\u6c34\u5e73\uff081-2 \u4e2a token\/s\uff09\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>\u7b14\u8bb0<\/h2>\n<ol>\n<li>\u60a8\u53ef\u80fd\u9700\u8981\u4f7f\u7528Homebrew\u5b89\u88c5llama.cpp\n<pre><code>\/bin\/bash -c \"$(curl -fsSL https:\/\/raw.githubusercontent.com\/Homebrew\/install\/HEAD\/install.sh)\"\r\nbrew install <a href=\"https:\/\/www.kdjingpai.com\/llamacpp\/\">llama.cpp<\/a>\r\n<\/code><\/pre>\n<p>\u7136\u540e\u4f7f\u7528\u6b64\u5904<code>llama-gguf-split<\/code>\u5efa\u8bae<\/p>\n<pre><code>llama-gguf-split --merge DeepSeek-R1-UD-IQ1_M-00001-of-00004.gguf DeepSeek-R1-UD-IQ1_S.gguf\r\nllama-gguf-split --merge DeepSeek-R1-Q4_K_M-00001-of-00009.gguf DeepSeek-R1-Q4_K_M.gguf\r\n<\/code><\/pre>\n<p>\u5982\u679c\u60a8\u77e5\u9053\u66f4\u597d\u7684\u65b9\u6cd5\uff0c\u8bf7\u5728\u8bc4\u8bba\u4e2d\u544a\u8bc9\u6211 \u3002<\/li>\n<li>\u8981\u66f4\u6539\u76ee\u5f55\uff0c\u8bf7\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\n<pre><code>sudo systemctl edit ollama\r\n<\/code><\/pre>\n<p>\u5e76\u5728\u7b2c\u4e8c\u884c\u4e4b\u540e\uff08\u5373\u201c\u00a0<code>### Anything between here and the comment below will become the contents of the drop-in file<\/code>\u201d\u548c\u201c\u00a0<code>### Edits below this comment will be discarded<\/code>\u201d\uff09\u6dfb\u52a0\u4ee5\u4e0b\u51e0\u884c<\/p>\n<pre><code>[Service]\r\nEnvironment=\"OLLAMA_MODELS=\/path\/to\/your\/directory\"\r\n<\/code><\/pre>\n<p>\u60a8\u8fd8\u53ef\u4ee5\u5728\u8fd9\u91cc\u8bbe\u7f6e\u4e00\u4e9b\u5176\u4ed6\u53c2\u6570\uff0c\u4f8b\u5982\uff0c<\/p>\n<pre><code>Environment=\"OLLAMA_FLASH_ATTENTION=1\"  # use flash attention\r\nEnvironment=\"OLLAMA_KEEP_ALIVE=-1\"      # keep the model loaded in memory\r\n<\/code><\/pre>\n<p>\u66f4\u591a\u8be6\u7ec6\u4fe1\u606f\u8bf7\u53c2\u89c1\u6b64\u5904\u3002<br \/>\n\u7136\u540e\u91cd\u542follama\u670d\u52a1<\/p>\n<pre><code>sudo systemctl restart ollama<\/code><\/pre>\n<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd9\u662f\u5173\u4e8e\u4f7f\u7528 ollama \u5728\u672c\u5730\u90e8\u7f72DeepSeek R1 671B\uff08\u672a\u7ecf\u63d0\u70bc\u7684\u5b8c\u6574\u7248\u672c\uff09\u7684\uff08\u6700\u5c11\uff09\u8bf4\u660e\u3002\u6700\u8fd1\u6709\u5927\u795e \u7528 2000 \u7f8e\u5143\u8dd1\u8d77 Deepseek R1 671b\uff0c\u5f88\u9002\u5408\u4e2a\u4eba\u4f7f\u7528\u3002 \u6a21\u578b \u539f\u59cb\u7684 DeepSeek R1 671&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61784,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[182,20],"tags":[],"class_list":["post-20401","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-shicao","category-tool"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/20401","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/comments?post=20401"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/posts\/20401\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media\/61784"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/media?parent=20401"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/categories?post=20401"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/de\/wp-json\/wp\/v2\/tags?post=20401"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}