{"id":215,"date":"2026-08-08T13:14:26","date_gmt":"2026-08-08T13:14:26","guid":{"rendered":"https:\/\/leetap.com\/blog\/2026\/08\/08\/2000-yuan-ai-reasoning-server\/"},"modified":"2026-08-08T13:14:26","modified_gmt":"2026-08-08T13:14:26","slug":"2000-yuan-ai-reasoning-server","status":"publish","type":"post","link":"https:\/\/leetap.com\/blog\/2026\/08\/08\/2000-yuan-ai-reasoning-server\/","title":{"rendered":"\u88ab\u73a9\u660e\u767d\u4e86\uff012000\u5757\u62fc\u4e00\u53f0\u80fd\u8dd1320\u4ebf\u53c2\u6570\u7684AI\u63a8\u7406\u670d\u52a1\u5668\uff0c\u53cc\u663e\u5361\u5f20\u91cf\u5e76\u884c\u6bd4llama.cpp\u5feb6\u500d\uff0c\u96f6\u57fa\u7840\u5b8c\u6574\u6559\u7a0b"},"content":{"rendered":"

\u5144\u5f1f\u4eec\uff0c\u60f3\u8dd1\u5927\u6a21\u578b\u4f46\u88ab\u663e\u5361\u4ef7\u683c\u529d\u9000\u7684\u770b\u8fc7\u6765\uff01\u4e00\u5f204090\u8981\u4e00\u4e07\u591a\uff0c\u4e00\u5f20A100\u4e8c\u624b\u90fd\u8981\u597d\u51e0\u4e07\uff0c\u666e\u901a\u4eba\u6839\u672c\u73a9\u4e0d\u8d77\u3002\u4f46\u4eca\u5929\u6211\u8981\u544a\u8bc9\u4f60\u4e00\u4e2a\u9a9a\u64cd\u4f5c\uff1a\u82b12000\u591a\u5757\uff0c\u5c31\u80fd\u62fc\u4e00\u53f0\u80fd\u8dd1320\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u7684AI\u63a8\u7406\u670d\u52a1\u5668<\/strong>\uff0c8\u4e2a\u4eba\u540c\u65f6\u63d0\u95ee\u90fd\u4e0d\u5361\uff0c\u901f\u5ea6\u6bd4llama.cpp\u5feb6\u500d\uff01<\/p>\n

\u8fd9\u5957\u65b9\u6848\u7684\u6838\u5fc3\u5c31\u4fe9\u5b57\uff1a\u6361\u6f0f<\/strong>\u3002\u7528\u4e24\u5f20\u4e8c\u624bAMD Radeon VII\u663e\u5361\uff08\u6bcf\u5f2016GB\u663e\u5b58\uff0c\u4e24\u5f20\u51d132GB\uff09\uff0c\u914d\u4e0a\u6dd8\u6c70\u7684\u6d0b\u5783\u573e\u670d\u52a1\u5668\u4e3b\u677f\u548cCPU\uff0c\u603b\u4ef72500\u5757\u641e\u5b9a\u3002\u6ca1\u9519\uff0c\u5c31\u662f\u5f53\u5e74AMD\u62ff\u6765\u6253RTX 2080 Ti\u7684\u65d7\u8230\u5361\uff0c\u73b0\u5728\u4e8c\u624b\u53ea\u8981800\u5757\u4e00\u5f20\uff0c\u663e\u5b58\u6bd42080 Ti\u8fd8\u5927\uff01<\/p>\n

\u8fd9\u7bc7\u6587\u7ae0\u4ece\u786c\u4ef6\u6e05\u5355\u3001\u7ec4\u88c5\u3001\u88c5\u7cfb\u7edf\u5230\u90e8\u7f72vLLM\u670d\u52a1\uff0c\u4e00\u6b65\u6b65\u6559\u4f60\u5168\u6d41\u7a0b\uff0c\u770b\u5b8c\u5c31\u80fd\u7167\u7740\u62fc\u4e00\u53f0\u3002\u9002\u5408\u60f3\u672c\u5730\u8dd1\u5927\u6a21\u578b\u3001\u53c8\u4e0d\u60f3\u82b1\u5927\u4ef7\u94b1\u7684\u5144\u5f1f\uff0c\u4e5f\u9002\u5408\u62ff\u6765\u5f53\u5f00\u6e90\u9879\u76ee\u7ec3\u624b\u3002<\/p>\n

\u4e00\u3001\u5148\u641e\u61c2\u539f\u7406\uff1a\u4e3a\u4ec0\u4e48\u4e24\u5f20\u65e7\u5361\u80fd\u8dd1320\u4ebf\u53c2\u6570\uff1f<\/h2>\n

\u4f60\u53ef\u80fd\u60f3\u95ee\uff1a320\u4ebf\u53c2\u6570=32B\u6a21\u578b\uff0c\u5149\u6743\u91cd\u6587\u4ef6\u5c3120\u591a\u4e2aG\uff0c\u4e24\u5f2016G\u7684\u5361\u52a0\u8d77\u676532G\uff0c\u600e\u4e48\u585e\u5f97\u4e0b\uff1f\u4e24\u4e2a\u5173\u952e\u70b9\uff1a<\/p>\n

\u7b2c\u4e00\uff0c\u91cf\u5316\u3002<\/strong>\u6a21\u578b\u75284-bit AWQ\u91cf\u5316\u540e\uff0c\u4f53\u79ef\u7f29\u5230\u539f\u6765\u76841\/4\u5de6\u53f3\uff0c32B\u6a21\u578b\u538b\u523020G\u4e0a\u4e0b\uff0c32G\u663e\u5b58\u521a\u597d\u88c5\u4e0b\u3002<\/p>\n

\u7b2c\u4e8c\uff0c\u5f20\u91cf\u5e76\u884c\u3002<\/strong>\u8fd9\u662f\u91cd\u70b9\uff01vLLM\u7684\u5f20\u91cf\u5e76\u884c\u4f1a\u628a\u6a21\u578b\u77e9\u9635\u4ece\u4e2d\u95f4\u5207\u5f00\uff0c\u4e24\u5f20\u663e\u5361\u540c\u65f6<\/strong>\u5904\u7406\u540c\u4e00\u4e2a\u5c42\u7684\u4e0d\u540c\u90e8\u5206\uff0c1+1=2\u3002<\/p>\n

\u5bf9\u6bd4\u4e00\u4e0b\u4f20\u7edf\u65b9\u6848llama.cpp\uff1a\u5b83\u7528\u7684\u662f\u6d41\u6c34\u7ebf\u5e76\u884c\uff0c\u6a21\u578b\u4e00\u5c42\u5c42\u987a\u5e8f\u8dd1\uff0c\u540c\u4e00\u65f6\u95f4\u53ea\u6709\u4e00\u5f20\u5361\u5728\u5e72\u6d3b\uff0c\u53e6\u4e00\u5f20\u5728\u7b49\u7740\u3002\u6240\u4ee5vLLM\u5f20\u91cf\u5e76\u884c\u6bd4llama.cpp\u5feb\u7ea66\u500d\u2014\u2014\u5b9e\u6d4b8\u4e2a\u7528\u6237\u5e76\u53d1\u63d0\u95ee\uff0c\u603b\u541e\u5410\u91cf\u8d85\u8fc7100 tokens\/s\uff0c\u6240\u6709\u7a97\u53e3\u6d41\u7545\u8fd0\u884c\u4e0d\u6392\u961f\u3002<\/p>\n

<\/p>\n

\u26a0\ufe0f \u6ce8\u610f\uff1a<\/strong>\u8fd9\u5957\u65b9\u6848\u5bf9MoE\uff08\u6df7\u5408\u4e13\u5bb6\uff09\u6a21\u578b\u652f\u6301\u6709\u9650\uff0cAWQ\u91cf\u5316\u7684\u7a20\u5bc6\u6a21\u578b\u662f\u5b83\u7684\u6700\u4f73\u5e94\u7528\u573a\u666f\u3002\u522b\u62ff\u5b83\u8dd1DeepSeek-V3\u90a3\u79cdMoE\u6a21\u578b\uff0c\u8dd1\u4e0d\u52a8\u7684\u3002<\/p>\n

\u4e8c\u3001\u51c6\u5907\u5de5\u4f5c\uff1a\u786c\u4ef6\u6e05\u5355\uff08\u5168\u59572500\u5143\uff09<\/h2>\n

\u5148\u770b\u6e05\u5355\uff0c\u5168\u662f\u4e8c\u624b\u4ef6\uff0c\u53bb\u54ea\u4e70\u6211\u90fd\u7ed9\u4f60\u6807\u597d\u4e86\uff1a<\/p>\n\n\n\n\n\n\n\n\n
\u914d\u4ef6<\/th>\n\u578b\u53f7<\/th>\n\u4e8c\u624b\u4ef7<\/th>\n\u53bb\u54ea\u4e70<\/th>\n<\/tr>\n
\u663e\u5361\u00d72<\/td>\nAMD Radeon VII\uff0816GB HBM2\uff09<\/td>\n800\u5143\/\u5f20<\/td>\n\u95f2\u9c7c<\/td>\n<\/tr>\n
CPU<\/td>\nIntel E5-2680v4\uff0814\u683828\u7ebf\u7a0b\uff09<\/td>\n56\u5143<\/td>\n\u95f2\u9c7c\/\u6dd8\u5b9d\u6d0b\u5783\u573e\u5e97<\/td>\n<\/tr>\n
\u4e3b\u677f<\/td>\n\u534e\u5357X99M Plus<\/td>\n300\u5143\u5de6\u53f3<\/td>\n\u6dd8\u5b9d<\/td>\n<\/tr>\n
\u5185\u5b58<\/td>\nDDR4\u670d\u52a1\u5668\u5185\u5b588GB\u00d74\uff08\u517132G\uff09<\/td>\n\u51e0\u5341\u5143<\/td>\n\u95f2\u9c7c<\/td>\n<\/tr>\n
\u786c\u76d8<\/td>\n512GB NVMe SSD<\/td>\n100\u5143<\/td>\n\u95f2\u9c7c<\/td>\n<\/tr>\n
\u7535\u6e90<\/td>\n800W\u7535\u6e90<\/td>\n155\u5143<\/td>\n\u95f2\u9c7c\/\u6dd8\u5b9d<\/td>\n<\/tr>\n<\/table>\n

\u5907\u9009\u663e\u5361\uff1a<\/strong>\u5982\u679cRadeon VII\u6da8\u4ef7\u4e86\uff0c\u53ef\u4ee5\u770bAMD M50\u8ba1\u7b97\u5361\u2014\u2014\u540c\u6b3eGFX906\u6838\u5fc3\uff0c\u6027\u80fd\u63a5\u8fd1\u8fd8\u66f4\u4fbf\u5b9c\uff0c\u4f46\u9700\u8981\u81ea\u5df1\u52a8\u624b\u6539\u6563\u70ed\uff0c\u52a8\u624b\u80fd\u529b\u5f3a\u7684\u53ef\u4ee5\u4e0a\u3002<\/p>\n

\u26a0\ufe0f \u6ce8\u610f\uff1a<\/strong>\u4e24\u5f20Radeon VII\u6ee1\u8f7d\u529f\u8017\u63a5\u8fd1600W\uff0c\u7535\u6e90\u5fc5\u987b\u4e0a800W\u4ee5\u4e0a\uff0c\u522b\u7701\u8fd9\u4e2a\u94b1\u3002Radeon VII\u53d1\u70ed\u5927\uff0c\u673a\u7bb1\u901a\u98ce\u8981\u505a\u597d\uff0c\u5efa\u8bae\u5f00\u653e\u5f0f\u673a\u67b6\u6216\u8005\u52a0\u88c5\u98ce\u6247\u3002<\/p>\n

\u4e09\u3001\u7ec4\u88c5\u670d\u52a1\u5668\uff08\u534a\u5c0f\u65f6\u641e\u5b9a\uff09<\/h2>\n

\u8fd9\u5957X99\u5e73\u53f0\u7ec4\u88c5\u6bd4\u666e\u901a\u7535\u8111\u8fd8\u7b80\u5355\uff0c\u90fd\u662f\u670d\u52a1\u5668\u62c6\u673a\u4ef6\uff0c\u63a5\u53e3\u6807\u51c6\u7edf\u4e00\uff1a<\/p>\n

\u7b2c1\u6b65\uff1a\u88c5CPU\u548c\u5185\u5b58\u3002<\/strong>E5-2680v4\u5bf9\u51c6\u4e3b\u677fCPU\u63d2\u69fd\u7684\u4e09\u89d2\u6807\u8bb0\u653e\u8fdb\u53bb\uff0c\u538b\u597d\u6263\u5177\u30024\u6761\u5185\u5b58\u63d2\u6ee1\u56db\u901a\u9053\uff0c\u6ce8\u610f\u63d2\u69fd\u987a\u5e8f\uff0c\u6309\u4e3b\u677f\u8bf4\u660e\u4e66\u6765\u3002<\/p>\n

\u7b2c2\u6b65\uff1a\u88c5\u786c\u76d8\u548c\u7535\u6e90\u3002<\/strong>NVMe SSD\u63d2\u5230\u4e3b\u677fM.2\u63a5\u53e3\u3002\u7535\u6e90\u7ebf\u63a5\u597d\uff1a\u4e3b\u677f24pin\u3001CPU 8pin\u3001\u663e\u5361\u4f9b\u7535\u2014\u2014Radeon VII\u662f\u53cc8pin\u4f9b\u7535\uff0c\u4e24\u5f20\u5361\u89814\u4e2a8pin\u63a5\u53e3\uff0c\u4e70\u7535\u6e90\u65f6\u786e\u8ba4\u63a5\u53e3\u591f\u7528\u3002<\/p>\n

\u7b2c3\u6b65\uff1a\u63d2\u663e\u5361\u3002<\/strong>\u4e24\u5f20Radeon VII\u63d2\u8fdb\u4e3b\u677f\u7684PCIe x16\u63d2\u69fd\u3002X99\u5e73\u53f0\u4e00\u822c\u6709\u591a\u4e2aPCIe\u63d2\u69fd\uff0c\u4e24\u5f20\u5361\u6700\u597d\u5206\u5f00\u63d2\uff0c\u4e2d\u95f4\u7559\u6563\u70ed\u7a7a\u95f4\u3002<\/p>\n

\u7b2c4\u6b65\uff1a\u5f00\u673a\u9a8c\u8bc1\u3002<\/strong>\u63a5\u4e0a\u663e\u793a\u5668\u548c\u952e\u9f20\uff0c\u5f00\u673a\u8fdbBIOS\uff0c\u80fd\u770b\u5230CPU\u578b\u53f7\u548c\u5185\u5b58\u5bb9\u91cf\u5c31\u8bf4\u660e\u786c\u4ef6\u6ca1\u95ee\u9898\u3002<\/p>\n

\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>BIOS\u91cc\u770b\u5230”Intel Xeon E5-2680 v4″\u548c32GB\u5185\u5b58\uff0c\u786c\u4ef6\u7ec4\u88c5\u5b8c\u6210\u3002<\/p>\n

<\/p>\n

\u56db\u3001\u88c5Ubuntu\u7cfb\u7edf<\/h2>\n

vLLM\u53ea\u652f\u6301Linux\uff0c\u88c5Ubuntu 22.04 LTS\u6700\u7701\u5fc3\u3002\u9700\u8981\u4e00\u4e2a8GB\u4ee5\u4e0a\u7684U\u76d8\u505a\u542f\u52a8\u76d8\uff1a<\/p>\n

\u7b2c1\u6b65\uff1a\u4e0b\u8f7d\u7cfb\u7edf\u955c\u50cf\u3002<\/strong>\u53bb Ubuntu\u5b98\u7f51\uff08https:\/\/ubuntu.com\/download\/server\uff09\u4e0b\u8f7d22.04 LTS Server\u7248ISO\u6587\u4ef6\u3002<\/p>\n

\u7b2c2\u6b65\uff1a\u505a\u542f\u52a8\u76d8\u3002<\/strong>\u7528 Rufus\uff08https:\/\/rufus.ie\uff09\u628aISO\u5199\u5165U\u76d8\uff0cU\u76d8\u6a21\u5f0f\u9009”DD\u955c\u50cf”\u5199\u5165\u3002<\/p>\n

\u7b2c3\u6b65\uff1a\u5b89\u88c5\u7cfb\u7edf\u3002<\/strong>U\u76d8\u63d2\u5230\u670d\u52a1\u5668\u4e0a\u5f00\u673a\uff0c\u8fdb\u542f\u52a8\u83dc\u5355\u9009U\u76d8\u542f\u52a8\uff0c\u4e00\u8def\u6309\u63d0\u793a\u88c5\u7cfb\u7edf\u3002\u5206\u533a\u76f4\u63a5\u7528\u9ed8\u8ba4\u6574\u76d8\u5b89\u88c5\uff0c\u7528\u6237\u540d\u5bc6\u7801\u81ea\u5df1\u8bbe\u4e00\u4e2a\u8bb0\u597d\u3002<\/p>\n

\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>\u88c5\u5b8c\u91cd\u542f\u8fdb\u7cfb\u7edf\uff0c\u547d\u4ee4\u884c\u8f93\u5165\u4e0b\u9762\u547d\u4ee4\uff0c\u80fd\u770b\u5230\u663e\u5361\u4fe1\u606f\u5c31\u5bf9\u4e86\uff1a<\/p>\n

\n# \u67e5\u770b\u663e\u5361\u662f\u5426\u88ab\u7cfb\u7edf\u8bc6\u522b\nlspci | grep -i amd\n<\/pre>\n
\u770b\u5230\u4e24\u884c “AMD … Vega 20” \u5f00\u5934\u7684\u8bbe\u5907\uff0c\u5c31\u662f\u4e24\u5f20Radeon VII\u90fd\u8ba4\u51fa\u6765\u4e86\u3002<\/p>\n
\u4e94\u3001\u88c5Docker\uff08\u8dd1vLLM\u7684\u5bb9\u5668\u73af\u5883\uff09<\/h2>\n
vLLM\u7528Docker\u90e8\u7f72\u6700\u7701\u4e8b\uff0c\u4e0d\u7528\u624b\u52a8\u88c5\u4e00\u5806Python\u4f9d\u8d56\u3002\u88c5Docker\u5b98\u65b9\u811a\u672c\u4e00\u6761\u547d\u4ee4\uff1a<\/p>\n
\n# \u5b98\u65b9\u4e00\u952e\u5b89\u88c5Docker\ncurl -fsSL https:\/\/get.docker.com | sh\n\n# \u628a\u5f53\u524d\u7528\u6237\u52a0\u5165docker\u7ec4\uff08\u7701\u5f97\u6bcf\u6b21sudo\uff09\nsudo usermod -aG docker $USER\n\n# \u91cd\u65b0\u767b\u5f55\u7ec8\u7aef\u8ba9\u914d\u7f6e\u751f\u6548\nexit\n<\/pre>\n
\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>\u91cd\u65b0\u767b\u5f55\u540e\u8f93\u5165 docker --version<\/code>\uff0c\u80fd\u663e\u793a\u7248\u672c\u53f7\u8bf4\u660e\u88c5\u597d\u4e86\u3002<\/p>\n
\u516d\u3001\u62c9\u53d6vLLM\u955c\u50cf\uff08AMD\u663e\u5361\u4e13\u7528\u7248\uff09<\/h2>\n
\u8fd9\u91cc\u662f\u6700\u5173\u952e\u7684\u4e00\u6b65\uff01\u5b98\u65b9vLLM\u4e0d\u652f\u6301AMD GFX906\u67b6\u6784\uff08Radeon VII\u7684\u6838\u5fc3\uff09\uff0c\u5fc5\u987b\u7528\u5927\u795eMLZY\u7ef4\u62a4\u7684vLLM\u4fee\u6539\u7248Docker\u955c\u50cf\uff0c\u5b83\u4e13\u95e8\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n
\u955c\u50cf\u662fv0.5.2.4.5\u7684LZY3\u7248\u672c\uff0c\u4f5c\u8005\u628a\u90e8\u7f72\u6559\u7a0b\u548c\u547d\u4ee4\u90fd\u6574\u7406\u5728\u4e86\u81ea\u5df1\u7684GitHub\u4ed3\u5e93\uff08\u89c6\u9891\u7b80\u4ecb\u91cc\u6709\u94fe\u63a5\uff0c\u6216\u8005GitHub\u641c\u7d22 “MLZY vLLM” \u5c31\u80fd\u627e\u5230\uff09\u3002\u62c9\u53d6\u547d\u4ee4\u683c\u5f0f\u5982\u4e0b\uff1a<\/p>\n
\n# \u4eceDocker Hub\u62c9\u53d6MLZY\u7684vLLM\u4fee\u6539\u7248\u955c\u50cf\uff08\u4ee5\u4f5c\u8005GitHub\u6559\u7a0b\u91cc\u7684\u5b8c\u6574\u955c\u50cf\u540d\u4e3a\u51c6\uff09\ndocker pull mlzy\/vllm:lzy3\n\n# \u62c9\u5b8c\u786e\u8ba4\u955c\u50cf\u5728\u672c\u5730\ndocker images\n<\/pre>\n
\u26a0\ufe0f \u6ce8\u610f\uff1a<\/strong>\u955c\u50cf\u5b8c\u6574\u540d\u79f0\u4ee5\u4f5c\u8005GitHub\u4ed3\u5e93\u8bf4\u660e\u4e3a\u51c6\uff0c\u5982\u679c\u4e0a\u9762\u8fd9\u4e2a\u62c9\u53d6\u5931\u8d25\uff0c\u53bb\u4f5c\u8005GitHub\u770bREADME\u91cc\u5199\u7684\u6700\u65b0\u955c\u50cf\u6807\u7b7e\u3002<\/p>\n
\u4e03\u3001\u4e0b\u8f7d\u6a21\u578b\uff08320\u4ebf\u53c2\u65704-bit\u91cf\u5316\u7248\uff09<\/h2>\n
\u63a8\u8350\u8dd1QwQ-32B\u7684AWQ\u91cf\u5316\u7248\u2014\u2014320\u4ebf\u53c2\u6570\u63a8\u7406\u6a21\u578b\uff0c\u5199\u4ee3\u7801\u3001\u6570\u5b66\u3001\u903b\u8f91\u63a8\u7406\u90fd\u5f88\u5f3a\uff0c4-bit\u91cf\u5316\u540e\u4f53\u79ef\u7ea620G\uff0c\u4e24\u5f2016G\u663e\u5b58\u521a\u597d\u653e\u4e0b\u3002<\/p>\n
\u6a21\u578b\u4e0b\u8f7d\u5730\u5740 \ud83d\udc49<\/strong> HuggingFace: Qwen\/QwQ-32B-AWQ<\/a><\/p>\n
\u56fd\u5185\u4e0b\u8f7d\u5730\u5740 \ud83d\udc49<\/strong> ModelScope: Qwen\/QwQ-32B-AWQ<\/a>\uff08\u56fd\u5185\u76f4\u8fde\uff0c\u901f\u5ea6\u5feb\uff0c\u63a8\u8350\uff09<\/p>\n
\u4e0b\u8f7d\u5230\u670d\u52a1\u5668\u4e0a\u7684 \/models \u76ee\u5f55\uff1a<\/p>\n
\n# \u521b\u5efa\u6a21\u578b\u76ee\u5f55\nmkdir -p \/models\n\n# \u7528ModelScope\u4e0b\u8f7d\uff08\u56fd\u5185\u5feb\uff09\uff0c\u88c5\u4e00\u4e0b\u4e0b\u8f7d\u5de5\u5177\npip install modelscope\n\n# \u4e0b\u8f7d\u6a21\u578b\u5230 \/models\uff08\u4f1a\u81ea\u52a8\u5efaQwQ-32B-AWQ\u6587\u4ef6\u5939\uff09\nmodelscope download --model Qwen\/QwQ-32B-AWQ --local_dir \/models\/QwQ-32B-AWQ\n<\/pre>\n
\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>\u4e0b\u8f7d\u5b8c\u6210\u540e\u68c0\u67e5\u6a21\u578b\u6587\u4ef6\uff1a<\/p>\n
\n# \u786e\u8ba4\u6a21\u578b\u6587\u4ef6\u90fd\u5728\nls -lh \/models\/QwQ-32B-AWQ | head\n\n# \u770b\u5230 .safetensors \u6743\u91cd\u6587\u4ef6\uff0c\u603b\u5927\u5c0f20G\u5de6\u53f3\u5c31\u5bf9\u4e86\n<\/pre>\n
\u26a0\ufe0f \u94fe\u63a5\u5931\u6548\u4e86\u600e\u4e48\u529e\uff1f<\/strong>\u522b\u6025\uff0c\u76f4\u63a5\u5728\u767e\u5ea6\u641c\u7d22”QwQ-32B-AWQ \u4e0b\u8f7d”\uff0c\u6216\u8005\u53bbHuggingFace\/ModelScope\u5b98\u7f51\u641c\u6a21\u578b\u540d\uff0c\u90fd\u80fd\u627e\u5230\u6700\u65b0\u5730\u5740\u3002<\/p>\n
\u516b\u3001\u542f\u52a8vLLM\u670d\u52a1\uff08\u5173\u952e\uff1a\u5f00\u5f20\u91cf\u5e76\u884c\uff09<\/h2>\n
\u542f\u52a8\u547d\u4ee4\u91cc\u6700\u6838\u5fc3\u7684\u53c2\u6570\u662f --tensor-parallel-size 2<\/code>\uff0c\u610f\u601d\u5c31\u662f\u8ba9\u4e24\u5f20\u663e\u5361\u540c\u65f6\u5e72\u6d3b\u3002\u8fd9\u4e2a\u5fc5\u987b\u5199\u5bf9\uff0c\u51991\u7684\u8bdd\u53ea\u6709\u4e00\u5f20\u5361\u5728\u5de5\u4f5c\uff0c\u901f\u5ea6\u51cf\u534a\uff1a<\/p>\n
\n# \u542f\u52a8vLLM\u670d\u52a1\u5668\uff0c\u6302\u8f7d\u6a21\u578b\u76ee\u5f55\uff0c\u5f00\u542f\u5f20\u91cf\u5e76\u884c2\ndocker run -d --name vllm-server \\\n  --gpus all \\\n  -v \/models:\/models \\\n  -p 8000:8000 \\\n  mlzy\/vllm:lzy3 \\\n  vllm serve \/models\/QwQ-32B-AWQ \\\n  --tensor-parallel-size 2 \\\n  --quantization awq \\\n  --max-model-len 8192\n\n# \u770b\u542f\u52a8\u65e5\u5fd7\ndocker logs -f vllm-server\n<\/pre>\n
\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>\u65e5\u5fd7\u91cc\u770b\u5230 “CUDA graphs” \u6216\u8005 “Starting vLLM server” \u5b57\u6837\uff0c\u8bf4\u660e\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210\u3001\u670d\u52a1\u5df2\u542f\u52a8\u3002\u7b2c\u4e00\u6b21\u542f\u52a8\u8981\u52a0\u8f7d20G\u6a21\u578b\uff0c\u7b49\u51e0\u5206\u949f\u6b63\u5e38\u3002<\/p>\n
\u26a0\ufe0f \u6ce8\u610f\uff1a<\/strong>\u5982\u679c\u65e5\u5fd7\u62a5\u663e\u5b58\u4e0d\u8db3\uff08OOM\uff09\uff0c\u628a --max-model-len<\/code> \u8c03\u5c0f\u52304096\u518d\u8bd5\u3002<\/p>\n
\u4e5d\u3001\u6d4b\u8bd5\uff1a8\u4eba\u540c\u65f6\u7528\u90fd\u4e0d\u5361<\/h2>\n
vLLM\u670d\u52a1\u542f\u52a8\u540e\u662fOpenAI\u517c\u5bb9API\uff0c\u4efb\u4f55\u652f\u6301OpenAI\u63a5\u53e3\u7684\u7a0b\u5e8f\u90fd\u80fd\u76f4\u63a5\u8fde\u3002\u5148\u672c\u5730\u6d4b\u4e00\u4e0b\u80fd\u4e0d\u80fd\u6b63\u5e38\u5bf9\u8bdd\uff1a<\/p>\n
\n# \u53d1\u4e00\u4e2a\u804a\u5929\u8bf7\u6c42\u6d4b\u8bd5\ncurl http:\/\/localhost:8000\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\"model\": \"\/models\/QwQ-32B-AWQ\", \"messages\": [{\"role\": \"user\", \"content\": \"\u7528Python\u5199\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\"}], \"max_tokens\": 100}'\n<\/pre>\n
\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>\u8fd4\u56de\u4e00\u6bb5JSON\uff0c\u91cc\u9762”content”\u5b57\u6bb5\u6709\u6a21\u578b\u751f\u6210\u7684\u4ee3\u7801\uff0c\u5c31\u8bf4\u660e\u6a21\u578b\u6b63\u5e38\u5de5\u4f5c\u4e86\u3002<\/p>\n
\u7136\u540e\u6d4b\u8bd5\u5e76\u53d1\u80fd\u529b\u2014\u20148\u4e2a\u7528\u6237\u540c\u65f6\u63d0\u95ee\uff1a<\/p>\n
\n# \u5e76\u53d1\u6d4b\u8bd5\u811a\u672c\uff088\u4e2a\u5e76\u53d1\u8bf7\u6c42\u540c\u65f6\u6253\u8fc7\u53bb\uff09\nfor i in $(seq 1 8); do\n  curl -s http:\/\/localhost:8000\/v1\/chat\/completions \\\n    -H \"Content-Type: application\/json\" \\\n    -d '{\"model\": \"\/models\/QwQ-32B-AWQ\", \"messages\": [{\"role\": \"user\", \"content\": \"\u89e3\u91ca\u4e00\u4e0b\u4ec0\u4e48\u662fTCP\u4e09\u6b21\u63e1\u624b\"}], \"max_tokens\": 200}' &\ndone\nwait\n<\/pre>\n
\ud83d\udc41\ufe0f \u9a8c\u8bc1\uff1a<\/strong>8\u4e2a\u8bf7\u6c42\u5168\u90e8\u5728\u51e0\u79d2\u5185\u8fd4\u56de\uff0c\u603b\u541e\u5410\u91cf\u8d85\u8fc7100 tokens\/s\uff0c\u6bcf\u4e2a\u7a97\u53e3\u90fd\u5728\u6d41\u7545\u8f93\u51fa\u2014\u2014\u8fd9\u5c31\u662fvLLM\u5f20\u91cf\u5e76\u884c+\u8fde\u7eed\u6279\u5904\u7406\u7684\u5a01\u529b\uff0c\u6bd4llama.cpp\u5feb\u4e86\u7ea66\u500d\u3002<\/p>\n
<\/p>\n
\u8fde\u4e0a\u4e4b\u540e\u4f60\u5c31\u80fd\u7528\u4efb\u4f55\u652f\u6301OpenAI API\u7684\u5ba2\u6237\u7aef\uff08\u6bd4\u5982ChatBox\u3001NextChat\u3001\u5404\u7c7bAI\u5de5\u5177\u7bb1\uff09\u586b\u4e0a http:\/\/\u670d\u52a1\u5668IP:8000<\/code> \u8fd9\u4e2a\u5730\u5740\uff0c\u767d\u5ad6\u81ea\u5df1\u7684\u79c1\u6709\u5927\u6a21\u578b\u4e86\u3002<\/p>\n
\u5341\u3001\u5e38\u89c1\u95ee\u9898<\/h2>\n
Q1\uff1a\u6211\u7684\u4e3b\u677f\u53ea\u6709\u4e00\u4e2aPCIe x16\u63d2\u69fd\uff0c\u80fd\u63d2\u4e24\u5f20\u5361\u5417\uff1f<\/h3>\n
\u4e0d\u884c\uff0c\u5f20\u91cf\u5e76\u884c\u5fc5\u987b\u4e24\u5f20\u5361\u90fd\u5728x16\uff08\u6216\u81f3\u5c11x8\uff09\u5e26\u5bbd\u7684\u63d2\u69fd\u4e0a\u3002\u9009\u4e3b\u677f\u65f6\u786e\u8ba4\u6709\u4e24\u4e2a\u5168\u901fPCIe\u63d2\u69fd\uff0c\u534e\u5357X99M Plus\u8fd9\u7c7b\u53cc\u63d2\u69fd\u7684\u677f\u5b50\u624d\u884c\u3002<\/p>\n
Q2\uff1a\u663e\u5b58\u4e0d\u591f\u62a5OOM\u600e\u4e48\u529e\uff1f<\/h3>\n
\u4e24\u4e2a\u529e\u6cd5\uff1a\u4e00\u662f\u628a–max-model-len\u4ece8192\u8c03\u5c0f\u52304096\uff0c\u7701\u663e\u5b58\uff1b\u4e8c\u662f\u786e\u8ba4\u6a21\u578b\u662fAWQ 4-bit\u91cf\u5316\u7248\uff0c\u522b\u4e0b\u62108-bit\u7684\u3002\u4e24\u5f20\u536132G\u663e\u5b58\u8dd132B AWQ\u662f\u521a\u521a\u597d\u7684\u914d\u7f6e\u3002<\/p>\n
Q3\uff1a\u4e3a\u4ec0\u4e48\u6211\u8dd1MoE\u6a21\u578b\u5f88\u6162\/\u8dd1\u4e0d\u8d77\u6765\uff1f<\/h3>\n
\u8fd9\u5957\u65b9\u6848\u5bf9MoE\uff08\u6df7\u5408\u4e13\u5bb6\uff09\u6a21\u578b\u652f\u6301\u6709\u9650\uff0c\u91cf\u5316MoE\u6a21\u578b\u66f4\u662f\u4e0d\u652f\u6301\u3002\u5b83\u5c31\u662f\u7ed9AWQ\u91cf\u5316\u7a20\u5bc6\u6a21\u578b\u51c6\u5907\u7684\uff0c\u522b\u62ff\u5b83\u786c\u8dd1MoE\u3002<\/p>\n
Q4\uff1a\u4e00\u5b9a\u8981\u7528MLZY\u7684\u955c\u50cf\u5417\uff1f\u5b98\u65b9vLLM\u4e0d\u884c\uff1f<\/h3>\n
\u4e0d\u884c\u3002Radeon VII\u7528\u7684GFX906\u67b6\u6784\uff0c\u5b98\u65b9vLLM\u4e0d\u8ba4\u8fd9\u4e2a\u67b6\u6784\uff0c\u8dd1\u4e0d\u8d77\u6765\u3002MLZY\u7684\u4fee\u6539\u7248\u4e13\u95e8\u652f\u6301GFX906\uff0c\u5fc5\u987b\u7528\u5b83\u3002<\/p>\n
Q5\uff1a\u5bb6\u91cc\u7535\u8d39\u625b\u5f97\u4f4f\u5417\uff1f<\/h3>\n
\u6ee1\u8f7d\u529f\u8017\u7ea6600W\uff0c\u8dd1\u8d77\u6765\u786e\u5b9e\u8d39\u7535\u3002\u4f46\u65e5\u5e38\u95ee\u7b54\u662f\u95f4\u6b47\u6027\u7684\uff0c\u4e0d\u7528\u7684\u65f6\u5019\u670d\u52a1\u95f2\u7740\u4e00\u5c0f\u65f6\u5927\u6982\u51e0\u5341\u74e6\u3002\u60f3\u60f3\u7701\u4e0b\u7684\u51e0\u4e07\u5757\u663e\u5361\u94b1\uff0c\u7535\u8d39\u90fd\u662f\u6bdb\u6bdb\u96e8\u3002<\/p>\n
\u603b\u7ed3<\/h2>\n
\u82b12000\u591a\u5757\u62fc\u4e00\u53f0\u80fd\u8dd1320\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u7684\u670d\u52a1\u5668\uff0c\u6838\u5fc3\u5c31\u4e09\u4ef6\u4e8b\uff1a\u4e24\u5f20\u4e8c\u624bRadeon VII\u6361\u6f0f + MLZY\u4fee\u6539\u7248vLLM\u955c\u50cf + \u5f20\u91cf\u5e76\u884c\u53c2\u6570\u5f002<\/strong>\u3002\u8fd9\u5957\u65b9\u6848\u6bd4\u4e70\u65b0\u663e\u5361\u7701\u4e86\u5341\u500d\u7684\u94b1\uff0c\u6bd4llama.cpp\u5feb6\u500d\uff0c8\u4eba\u540c\u65f6\u7528\u90fd\u4e0d\u5361\uff0c\u62ff\u6765\u672c\u5730\u8dd1\u6a21\u578b\u3001\u642d\u79c1\u6709API\u3001\u505a\u5f00\u6e90\u9879\u76ee\u90fd\u662f\u795e\u5668\u3002<\/p>\n
\u7ed9\u65b0\u624b\u7684\u4e00\u53e5\u8bdd\u5efa\u8bae\uff1a\u7167\u7740\u6e05\u5355\u4e70\u914d\u4ef6\u65f6\uff0c\u663e\u5361\u548c\u7535\u6e90\u8fd9\u4e24\u6837\u522b\u8d2a\u4fbf\u5b9c\uff0c\u5176\u5b83\u90fd\u80fd\u5c06\u5c31\u3002\u88c5\u597d\u7cfb\u7edf\u540e\u5148\u8dd1\u901a\u6d4b\u8bd5\u518d\u7814\u7a76\u8fdb\u9636\u73a9\u6cd5\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u5144\u5f1f\u4eec\uff0c\u60f3\u8dd1\u5927\u6a21\u578b\u4f46\u88ab\u663e\u5361\u4ef7\u683c\u529d\u9000\u7684\u770b\u8fc7\u6765\uff01\u4e00\u5f204090\u8981\u4e00\u4e07\u591a\uff0c\u4e00\u5f20A100\u4e8c\u624b\u90fd\u8981\u597d\u51e0\u4e07\uff0c\u666e\u901a\u4eba\u6839\u672c\u73a9\u4e0d\u8d77… <\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"neve_meta_sidebar":"","neve_meta_container":"","neve_meta_enable_content_width":"","neve_meta_content_width":0,"neve_meta_title_alignment":"","neve_meta_author_avatar":"","neve_post_elements_order":"","neve_meta_disable_header":"","neve_meta_disable_footer":"","neve_meta_disable_title":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-215","post","type-post","status-publish","format-standard","hentry","category-default"],"_links":{"self":[{"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/posts\/215","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/comments?post=215"}],"version-history":[{"count":0,"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/posts\/215\/revisions"}],"wp:attachment":[{"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/media?parent=215"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/categories?post=215"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/leetap.com\/blog\/wp-json\/wp\/v2\/tags?post=215"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}