{"id":10323,"date":"2025-02-11T16:07:01","date_gmt":"2025-02-11T08:07:01","guid":{"rendered":"https:\/\/yanjingang.com\/blog\/?p=10323"},"modified":"2025-09-02T18:58:51","modified_gmt":"2025-09-02T10:58:51","slug":"deepseek%e6%9c%ac%e5%9c%b0%e5%8c%96%e9%83%a8%e7%bd%b2","status":"publish","type":"post","link":"https:\/\/yanjingang.com\/blog\/?p=10323","title":{"rendered":"DeepSeek\u2014\u672c\u5730\u5316\u90e8\u7f72"},"content":{"rendered":"
\n

\u751f\u6210\u5f0f AI\uff08Artificial Intelligence Generated Content\uff0cAIGC\uff09\u7684\u5174\u8d77\u6807\u5fd7\u7740\u4fe1\u606f\u6280\u672f\u884c\u4e1a\u8fdb\u5165\u4e86\u5d2d\u65b0\u7684\u53d1\u5c55\u9636\u6bb5\uff0c\u4e3a\u793e\u4f1a\u8fdb\u6b65\u548c\u7ecf\u6d4e\u53d1\u5c55\u6253\u5f00\u4e86\u5de8\u5927\u7684\u60f3\u8c61\u7a7a\u95f4\u3002<\/p>\n

\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval Augmented Generation\uff0cRAG\uff09\u6280\u672f\u901a\u8fc7\u6574\u5408\u5185\u90e8\u77e5\u8bc6\u5e93\u7684\u4e30\u5bcc\u4fe1\u606f\u8d4b\u80fd\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Model\uff0c\u7b80\u79f0 LLM\uff09\uff0c\u4e3a\u751f\u6210\u5f0f AI \u5e26\u6765\u4e86\u9769\u547d\u6027\u7684\u6539\u8fdb\u3002\u5728\u751f\u6210\u7b54\u6848\u6216\u5185\u5bb9\u65f6\uff0cRAG \u6280\u672f\u80fd\u591f\u53c2\u7167\u7cbe\u786e\u7684\u77e5\u8bc6\u6e90\uff0c\u4e3a\u7528\u6237\u63d0\u4f9b\u9ad8\u5ea6\u5408\u7406\u548c\u51c6\u786e\u7684\u54cd\u5e94\u3002<\/p>\n<\/div>\n

\n

RAG \u6280\u672f\u4e0e LLM \u6a21\u578b\u76f8\u7ed3\u5408\u7684\u5e94\u7528\u65b9\u5f0f\uff0c\u6210\u529f\u5f25\u8865\u4e86 LLM \u6a21\u578b\u5728\u5904\u7406\u7279\u5b9a\u9886\u57df\u8bf7\u6c42\u65f6\u5b58\u5728\u5e7b\u89c9\u3001\u77e5\u8bc6\u6cdb\u5316\u3001\u53ef\u89e3\u91ca\u6027\u5dee\u7b49\u95ee\u9898\u3002\u57fa\u4e8e LLM \u548c RAG \u6280\u672f\u7684\u77e5\u8bc6\u5e93\u5e94\u7528\uff0c\u53ef\u4ee5\u5e2e\u52a9\u4f01\u4e1a\u63d0\u5347\u77e5\u8bc6\u7ba1\u7406\u4e0e\u670d\u52a1\u6c34\u5e73\uff0c\u63d0\u9ad8\u5de5\u4f5c\u6548\u7387\uff0c\u5b9e\u73b0\u4e1a\u52a1\u7684\u63d0\u8d28\u589e\u6548\u3002<\/p>\n

\u672c\u6587\u57fa\u4e8eDeepSeek\u6a21\u578b+Ollama\u5927\u6a21\u578b\u6846\u67b6+MaxKB\u77e5\u8bc6\u5e93\u95ee\u7b54\u7cfb\u7edf\uff0c\u6765\u5728\u672c\u5730\u642d\u5efa\u4e00\u4e2a\u79c1\u6709\u7684\u3001\u5177\u5907\u4f01\u4e1a\u79c1\u6709\u65b0\u77e5\u8bc6\u5b66\u4e60\u80fd\u529b\u7684\u77e5\u8bc6\u95ee\u7b54\u7cfb\u7edf\u3002<\/p>\n<\/div>\n

\u4e00\u3001\u672c\u5730\u90e8\u7f72<\/h1>\n

1. \u663e\u5361\u8981\u6c42<\/h4>\n

\u9996\u5148\u4f60\u9700\u8981\u6709\u4e00\u4e2aNidia\u6216AMD GPU\u7684Linux\u8bbe\u5907\uff0c\u6a21\u578b\u53c2\u6570\u5927\u5c0f\u4e0e\u663e\u5b58\u5927\u5c0f\u7684\u5bf9\u5e94\u5173\u7cfb\uff1a<\/p>\n

\u6a21\u578b\u663e\u5b58\u5360\u7528\uff1a<2b\u96c6\u6210\u663e\u5361\u53ef\u4ee5\u8dd1\uff0c7b\u6a21\u578b\u9700\u89814G\u663e\u5b58\uff0c8b\u6a21\u578b\u9700\u89816-8G\u663e\u5b58\uff0c14b\u9700\u898112-16G\u663e\u5b58\uff0c32b\u9700\u898116-24G\u663e\u5b58\uff0c671b\u57fa\u672c\u9700\u89811.3T\u663e\u5b58\uff0c\u53ef\u4ee5\u75288\u4e2aH200\uff08141G*8=1128G\u663e\u5b58\uff09\u62164\u7ec48\u4e2aA100(80G*8*4=2560G)\u3002<\/p>\n

2. \u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6<\/h4>\n

\u6a21\u578b\u6587\u4ef6\u5360\u7528\u78c1\u76d8\u7a7a\u95f4\uff1a1.5b\u7ea6\u5360\u7528\u7a7a\u95f41.1GB\uff0c7b\u6a21\u578b4.7GB\uff0c8b\u6a21\u578b4.9GB\uff0c14b\u6a21\u578b9.0GB\uff0c32b\u6a21\u578b20GB\uff0c70b\u6a21\u578b43GB\uff0c671b\u6a21\u578b404GB\u3002<\/p>\n

\u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6\uff1a\uff08\u8fd9\u91cc\u6211\u4eec\u4f7f\u7528ollama\uff0c\u4f7f\u7528SGLang\u3001vLLM\u7b49\u4e5f\u53ef\u4ee5\uff09<\/p>\n

# \u5b89\u88c5ollama\r\ncurl -fsSL https:\/\/ollama.com\/install.sh | sh\r\n\r\n# \u4fee\u6539\u6a21\u578b\u4fdd\u5b58\u4f4d\u7f6e\r\nmkdir -p \/home\/work\/ollama\/models\r\nsudo vim \/etc\/systemd\/system\/ollama.service\r\n    # \u5728\u539f\u6709Environment\u53d8\u91cf\u4e0b\u589e\u52a0\u884c\uff0c\u5185\u5bb9\u5982\u4e0b\r\n    Environment=\"OLLAMA_MODELS=\/home\/work\/ollama\/models\"\r\n    Environment=\"OLLAMA_HOST=0.0.0.0\"\r\n\r\n# \u79fb\u52a8\u65e7\u6a21\u578b\r\nsudo ls -al \/usr\/share\/ollama\/.ollama\/models\r\nsudo mv \/usr\/share\/ollama\/.ollama\/models \/home\/work\/ollama\/\r\nll \/home\/work\/ollama\/models\/\r\n\r\n# \u91cd\u542follama\r\nsudo systemctl daemon-reload\r\nsudo systemctl restart ollama\r\nsudo systemctl status ollama\r\nollama list\r\n    NAME ID SIZE MODIFIED \r\n    deepseek-r1:1.5b a42b25d8c10a 1.1 GB 12 minutes ago\r\n\r\n# \u4e0b\u8f7d&\u8fd0\u884c\u6a21\u578b\r\nollama run deepseek-r1:1.5b\r\nollama run deepseek-r1:8b\r\nollama run deepseek-r1:32b\r\nollama run deepseek-v2:16b\r\nollama run deepseek-coder:1.3b\r\n# ollama run deepseek-v3:671b<\/code><\/pre>\n
ollama\u5e38\u7528\u547d\u4ee4\uff1a<\/p>\n\n\n\n\n\n\n\n\n\n
\u547d\u4ee4<\/td>\n\u8bf4\u660e<\/td>\n<\/tr>\n
serve<\/td>\n\u542f\u52a8ollama\u3002<\/td>\n<\/tr>\n
run<\/td>\n\u8fd0\u884c\u4e00\u4e2a\u6307\u5b9a\u6a21\u578b\uff0c\u672c\u5730\u4e0d\u5b58\u5728\u65f6\u4f1a\u4ece\u670d\u52a1\u5668\u4e0b\u8f7d\u3002<\/td>\n<\/tr>\n
stop<\/td>\n\u505c\u6b62\u4e00\u4e2a\u6307\u5b9a\u6a21\u578b\u3002<\/td>\n<\/tr>\n
list<\/td>\n\u5217\u51fa\u672c\u5730\u6240\u6709\u7684\u6a21\u578b\u3002<\/td>\n<\/tr>\n
ps<\/td>\n\u5217\u51fa\u672c\u5730\u771f\u6b63\u8fd0\u884c\u7684\u6a21\u578b\u3002<\/td>\n<\/tr>\n
rm<\/td>\n\u5220\u9664\u4e00\u4e2a\u6307\u5b9a\u7684\u6a21\u578b\u3002<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
3. \u8fd0\u884c\u6a21\u578b<\/h4>\n
3.1 \u547d\u4ee4\u884c\u65b9\u5f0f<\/h6>\n
ollama run\u4f1a\u81ea\u52a8\u4e0b\u8f7d\u5e76\u8fd0\u884c\u6a21\u578b\u3002<\/p>\n
\u4f60\u662f\u540d\u4fa6\u63a2\u4f73\u8dc3\u6469\u65af\uff0c\u5c45\u4f4f\u5728\u5317\u4eac\u8d1d\u514b\u6c49\u59c6\u8857666\u53f7\uff0c\u6839\u636e\u4f60\u7684\u4e30\u5bcc\u7684\u63a2\u6848\u7ecf\u9a8c\uff0c\u8be6\u7ec6\u63cf\u8ff0\u4e00\u4e2a\u4f60\u9047\u5230\u8fc7\u7684\u6700\u53cd\u8f6c\u7684\u6848\u4ef6\u8fc7\u7a0b<\/p>\n
\"\"<\/a><\/p>\n
\"\"<\/a><\/p>\n
\u4f60\u662f\u56fd\u9645\u8c61\u68cb\u5927\u5e08\u5361\u5c14\u68ee\uff0c\u73b0\u5728\u5728\u53c2\u52a0\u4e16\u754c\u51a0\u519b\u8d5b\u7b2c\u4e00\u8f6e\u5bf9\u9635\u56fd\u9645\u7279\u7ea7\u5927\u5e08\u95eb\u4f73\u8dc3\uff0c\u4f60\u6267\u767d\uff0c\u7b2c\u4e00\u6b65\u8d70\u4ec0\u4e48\uff1f<\/p>\n
\"\"<\/a><\/p>\n
3.2 Python\u65b9\u5f0f<\/h6>\n
# \u5b89\u88c5\u4f9d\u8d56\r\npip3 install ollama\r\n\r\n# \u7f16\u5199\u7a0b\u5e8f\r\nvim test.py\r\nimport ollama\r\n# res\r\nres = ollama.chat(\r\n    model='deepseek-r1:32b',\r\n    messages=[{'role': 'user', 'content': '\u4f60\u662f\u8c01?'}],\r\n    options={\"temperature\": 0},\r\n    stream=False,\r\n)\r\nprint(res['message']['content'])\r\n# stream\r\nstream = ollama.chat(\r\n    model='deepseek-r1:32b',\r\n    messages=[{'role': 'user', 'content': '\u5929\u7a7a\u4e3a\u4ec0\u4e48\u662f\u84dd\u8272\u7684?'}],\r\n    options={\"temperature\": 0},\r\n    stream=True,\r\n)\r\nfor chunk in stream:\r\n  print(chunk['message']['content'], end='', flush=True)\r\n\r\n# \u8fd0\u884c\r\npython3 test.py<\/code><\/pre>\n
\u53c2\u6570\u4ecb\u7ecd\u5982\u4e0b\uff1a<\/p>\n