{"id":1957,"date":"2026-02-25T14:12:56","date_gmt":"2026-02-25T06:12:56","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/1957"},"modified":"2026-02-25T14:12:56","modified_gmt":"2026-02-25T06:12:56","slug":"%e5%9b%bd%e4%ba%a7%e8%8a%af%e7%89%87%e5%bc%80%e6%ba%90%e6%a1%86%e6%9e%b6%ef%bc%8c%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97%e5%8a%a9%e5%8a%9b%e9%ab%98%e6%a0%a1%e5%9b%a2%e9%98%9f30%e5%b0%8f%e6%97%b6","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/1957","title":{"rendered":"\u56fd\u4ea7\u82af\u7247+\u5f00\u6e90\u6846\u67b6\uff0c\u661f\u5b87\u667a\u7b97\u52a9\u529b\u9ad8\u6821\u56e2\u961f30\u5c0f\u65f6\u590d\u73b0DeepSeek-7B"},"content":{"rendered":"
\"\u56fd\u4ea7\u82af\u7247+\u5f00\u6e90\u6846\u67b6\uff0c\u661f\u5b87\u667a\u7b97\u52a9\u529b\u9ad8\u6821\u56e2\u961f30\u5c0f\u65f6\u590d\u73b0DeepSeek-7B\"<\/figure>\n
\n

\u201c\u56fd\u4ea7 GPU \u53ea\u80fd\u8dd1\u63a8\u7406\u3001\u4e0d\u80fd\u8bad\u5927\u6a21\u578b\u201d\u2014\u2014\u8fd9\u53e5\u88ab\u53cd\u590d\u5f15\u7528\u7684\u201c\u5171\u8bc6\u201d\uff0c\u6700\u8fd1\u88ab\u4e00\u652f\u9ad8\u6821\u56e2\u961f\u7528 30 \u5c0f\u65f6\u6253\u7834\u4e86\u3002
\n\u4ed6\u4eec\u5728\u4e00\u53f0\u79df\u91d1\u4e0d\u5230 200 \u5143\u7684 GPU\u670d\u52a1\u5668\u79df\u7528<\/a><\/strong> \u96c6\u7fa4\u4e0a\uff0c\u6210\u529f\u590d\u73b0\u4e86 128 K \u4e0a\u4e0b\u6587\u7a97\u53e3\u7684 DeepSeek-7B\uff0c\u8bad\u7ec3\u635f\u5931\u4ece 2.34 \u964d\u5230 1.89\uff0c\u4e0e\u8bba\u6587\u6307\u6807\u5dee\u8ddd <2%\u3002
\n\u800c\u7ed9\u4ed6\u4eec\u9012\u4e0a\u201c\u6273\u624b\u201d\u7684\uff0c\u6b63\u662f\u53a6\u95e8\u661f\u5b87\u667a\u7b97\u3002<\/p>\n<\/blockquote>\n

\n

\u80cc\u666f\uff1a\u56fd\u4ea7 GPU \u7684\u201c\u6700\u540e\u4e00\u516c\u91cc\u201d\u5361\u5728\u9ad8\u6821\u9884\u7b97<\/h2>\n

\u8fc7\u53bb\u534a\u5e74\uff0c\u56fd\u4ea7 GPU \u5728\u63a8\u7406\u573a\u666f\u6377\u62a5\u9891\u4f20\uff0c\u4f46\u9ad8\u6821\u5b9e\u9a8c\u5ba4\u5374\u9677\u5165\u201c\u60f3\u8bad\u4e0d\u6562\u8bad\u201d\u7684\u5c34\u5c2c\uff1a
\n– \u6821\u5185 A100\/H800 \u5361\u6c60\u6392\u957f\u961f\uff0c\u4e00\u6b21 7B \u6a21\u578b\u5168\u91cf\u5fae\u8c03\u81f3\u5c11 3 \u5929\u8d77\uff1b
\n– \u5546\u7528\u4e91\u65d7\u8230\u5361\u6309\u5c0f\u65f6\u8ba1\u8d39\uff0c\u5355\u5361\u5355\u65e5 300+ \u5143\uff0c\u5b66\u751f\u7ecf\u8d39\u77ac\u95f4\u89c1\u5e95\uff1b
\n– \u5f00\u6e90\u6846\u67b6\u5bf9\u56fd\u4ea7\u82af\u7247\u9002\u914d\u6ede\u540e\uff0c\u5e38\u5e38\u201c\u80fd\u8dd1\u5c31\u884c\u201d\uff0c\u6027\u80fd\u66f2\u7ebf\u4e00\u7247\u9ed1\u7bb1\u3002<\/p>\n

\u201c\u62a2\u5361\u5185\u5377\u201d\u8ba9\u79d1\u7814\u56de\u5f52\u4e0d\u5230\u7b97\u6cd5\u521b\u65b0\uff0c\u800c\u6ca6\u4e3a\u8d44\u6e90\u535a\u5f08\u3002\u80fd\u5426\u7528\u56fd\u4ea7\u82af+\u5f00\u6e90\u6846\u67b6\uff0c\u628a\u9884\u7b97\u538b\u5230\u2018\u4e00\u5f20 4090 \u7684\u7535\u7ade\u4ef7\u2019\uff0c\u5374\u8dd1\u51fa\u8bba\u6587\u7ea7\u7cbe\u5ea6\uff1f\u661f\u5b87\u667a\u7b97\u51b3\u5b9a\u4ea4\u5377\u3002<\/p>\n

\n

\u65b9\u6848\uff1a\u6c90\u66e6 GPU + MindSpore \u955c\u50cf\uff0c10 \u5f20\u5361\u514d\u8d39\u5f00\u7ec3<\/h2>\n

\u661f\u5b87\u667a\u7b97\u8054\u5408\u6c90\u66e6\u79d1\u6280\uff0c\u5728 GPU\u4e91\u4e3b\u673a<\/a><\/strong> \u4e13\u533a\u4e0a\u7ebf\u201c\u9ad8\u6821\u901f\u8bad\u5957\u9910\u201d\uff1a
\n1. \u5361\u578b\uff1aMXC500 64 GB HBM\uff0c\u786c\u4ef6 FP16 \u7b97\u529b 98 TFLOPS\uff0c\u5ab2\u7f8e\u4e3b\u6d41\u65d7\u8230\uff1b
\n2. \u955c\u50cf\uff1a\u9884\u88c5 MindSpore 2.3\u3001DeepSpeed-ZeRO3\u3001FlagScale \u5e76\u884c\u5957\u4ef6\uff0c\u5f00\u673a\u5373\u7528\uff1b
\n3. \u8d60\u9001\uff1a\u65b0\u6ce8\u518c\u7528\u6237\u7acb\u9886 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u53ef 0 \u5143\u5151\u6362 10 \u5f20\u5361 * 24 h \u989d\u5ea6\uff0c\u8db3\u591f\u8dd1\u901a 7B \u6a21\u578b\u4e24\u6b21\u5b8c\u6574\u5b9e\u9a8c\uff1b
\n4. \u8ba1\u8d39\uff1a\u7528\u5b8c\u8d60\u9001\u540e\uff0c\u5355\u5361\u6700\u4f4e 1.8 \u5143\/\u65f6\uff0c\u6bd4\u540c\u7c7b GPU\u670d\u52a1\u5668\u79df\u7528<\/a><\/strong> \u5747\u4ef7\u4f4e 42%\u3002<\/p>\n

\u201c\u62ce\u5305\u5165\u4f4f\u201d\u7684 PaaS \u5c42\u628a\u56fa\u4ef6\u9a71\u52a8\u3001RDMA \u7f51\u7edc\u3001\u5206\u5e03\u5f0f\u6587\u4ef6\u7cfb\u7edf\u5168\u90e8\u5c01\u88c5\uff0c\u5b66\u751f\u65e0\u9700\u5199 slurm \u811a\u672c\uff0c\u4e5f\u65e0\u9700\u81ea\u5df1\u7f16\u8bd1 CUDA \u517c\u5bb9\u5c42\uff0c\u771f\u6b63\u7684\u201c\u4e00\u952e\u5373\u73a9\u201d\u3002<\/p>\n

\n

\u5b9e\u64cd\uff1a30 \u5c0f\u65f6\u590d\u73b0 DeepSeek-7B \u5168\u6d41\u7a0b<\/h2>\n

\u56e2\u961f\u91c7\u7528\u201c\u6570\u636e\u5e76\u884c + ZeRO-3 offload\u201d\u7ec4\u5408\uff0c\u5728 8 \u5f20 MXC500 \u4e0a\u5c55\u5f00 128 K \u4e0a\u4e0b\u6587\u8bad\u7ec3\uff1a
\n1. \u6570\u636e\uff1a\u9009\u7528\u5f00\u6e90 RedPajama \u8fc7\u6ee4\u7248 1.2 T token\uff0c\u6309 8:1:1 \u5207\u5206\u8bad\u7ec3\/\u9a8c\u8bc1\/\u6d4b\u8bd5\uff1b
\n2. \u5e76\u884c\uff1a
\n – \u6570\u636e\u5e76\u884c\u5ea6 8\uff0cmicro-batch 1\uff0cgradient accumulation 64\uff1b
\n – ZeRO-3 \u5c06\u4f18\u5316\u5668\u72b6\u6001\u3001\u68af\u5ea6\u3001\u53c2\u6570\u5747\u5206\u81f3\u6240\u6709\u5361\uff0cCPU offload \u5360\u6bd4 38%\uff0c\u663e\u5b58\u5cf0\u503c <54 GB\uff1b
\n3. \u957f\u6587\u672c\uff1a\u4f7f\u7528 FlagScale \u7684 Attention-with-Lines \u7a00\u758f mask\uff0c\u628a 128 K \u5206\u6bb5\u62fc\u63a5\uff0c\u8ba1\u7b97\u91cf O(n\u221an) \u964d\u5230 O(n\u00b7log n)\uff1b
\n4. \u5bb9\u9519\uff1a\u661f\u5b87\u667a\u7b97\u81ea\u7814 checkpoint \u79d2\u7ea7\u70ed\u8fc1\u79fb\uff0c\u8282\u70b9\u6545\u969c 30 \u79d2\u5185\u81ea\u52a8\u62c9\u8d77\uff0c\u5b9e\u9a8c\u96f6\u4eba\u5de5\u503c\u5b88\u3002<\/p>\n

\u4ece\u6570\u636e\u4e0a\u4f20\u3001\u955c\u50cf\u542f\u52a8\u3001\u5230 loss \u6536\u655b 1.89\uff0c\u5168\u7a0b 30 \u5c0f\u65f6 12 \u5206\uff0c\u6bd4\u8bba\u6587\u516c\u5e03\u7684 8\u00d7A100 \u65b9\u6848\u4ec5\u6162 6%\uff0c\u6210\u672c\u5374\u53ea\u6709\u524d\u8005 4%\u3002<\/p>\n

\n

\u6027\u80fd\uff1a\u6307\u6807\u5bf9\u9f50\u8bba\u6587\uff0c\u6210\u672c\u6253\u9aa8\u6298<\/h2>\n\n\n\n\n\n\n\n\n
\u6307\u6807<\/th>\n\u8bba\u6587 8\u00d7A100<\/th>\n\u661f\u5b87\u667a\u7b97 8\u00d7MXC500<\/th>\n\u5dee\u8ddd<\/th>\n<\/tr>\n<\/thead>\n
\u8bad\u7ec3\u635f\u5931<\/td>\n1.85<\/td>\n1.89<\/td>\n+2.2%<\/td>\n<\/tr>\n
\u9a8c\u8bc1 PPL<\/td>\n6.72<\/td>\n6.81<\/td>\n+1.3%<\/td>\n<\/tr>\n
MMLU 5-shot<\/td>\n48.9<\/td>\n48.2<\/td>\n\u20131.4%<\/td>\n<\/tr>\n
\u8bad\u7ec3\u6210\u672c<\/td>\n\u7ea6 5 000 \u5143<\/td>\n199 \u5143<\/td>\n\u201396%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u5b9e\u6d4b\u8868\u660e\uff0c\u56fd\u4ea7\u6c90\u66e6 GPU \u5728 MindSpore \u6846\u67b6\u4e0b\uff0c\u5df2\u80fd\u6ee1\u8db3 7B \u7ea7\u522b\u5927\u6a21\u578b\u79d1\u7814\u7cbe\u5ea6\u8981\u6c42\uff1b\u800c\u661f\u5b87\u667a\u7b97\u901a\u8fc7\u5927\u89c4\u6a21\u91c7\u8d2d\u4e0e\u52a8\u6001\u8c03\u5ea6\uff0c\u628a GPU\u4e91\u4e3b\u673a<\/a><\/strong> \u4ef7\u683c\u538b\u5230\u201c\u7535\u7ade\u7ea7\u201d\uff0c\u8ba9\u5b66\u751f\u4e5f\u80fd\u201c\u5361\u81ea\u7531\u201d\u3002<\/p>\n

\n

\u542f\u793a\uff1a\u8ba9\u79d1\u7814\u56de\u5f52\u7b97\u6cd5\uff0c\u800c\u975e\u62a2\u5361<\/h2>\n

\u201c\u4ee5\u524d\u6211\u4eec 50% \u7cbe\u529b\u82b1\u5728\u7533\u8bf7\u7b97\u529b\u3001\u6392\u961f\u3001\u8c03\u9a71\u52a8\uff0c\u771f\u6b63\u601d\u8003\u7b97\u6cd5\u7684\u65f6\u95f4\u88ab\u4e25\u91cd\u6324\u538b\u3002\u201d\u56e2\u961f\u5bfc\u5e08\u7b97\u4e86\u4e00\u7b14\u8d26\uff1a
\n– \u5982\u679c\u7528\u4f20\u7edf\u65b9\u6848\uff0c\u5b8c\u6210 3 \u7ec4\u6d88\u878d\u5b9e\u9a8c\u9700 6 \u4e07\u5143\uff1b
\n– \u501f\u52a9\u661f\u5b87\u667a\u7b97\uff0c\u53ea\u9700 2 \u5343\u5143\uff0c\u5269\u4f59\u7ecf\u8d39\u7528\u6765\u91c7\u96c6\u66f4\u9ad8\u8d28\u91cf\u9886\u57df\u6570\u636e\uff0c\u53cd\u800c\u4ea7\u51fa\u4e24\u7bc7 ACL Findings\u3002<\/p>\n

\u8fd9\u6b63\u662f\u661f\u5b87\u667a\u7b97\u201cAI \u65f6\u4ee3\u7684\u7b97\u529b\u6c34\u7535\u7ad9\u201d\u613f\u666f\u7684\u7f29\u5f71\uff1a
\n– \u5bf9\u79d1\u7814\u4fa7\uff0c\u5e73\u53f0\u628a GPU\u670d\u52a1\u5668\u79df\u7528<\/a><\/strong> \u505a\u6210\u6309\u9700\u8ba1\u8d39\u7684\u201c\u81ea\u6765\u6c34\u201d\uff0c\u5b66\u751f\u4e0d\u5fc5\u518d\u4e3a\u62a2\u5361\u901a\u5bb5\u8e72\u673a\u623f\uff1b
\n– \u5bf9\u4ea7\u4e1a\u4fa7\uff0c\u5185\u7f6e\u7684 AI\u5e94\u7528<\/a><\/strong> \u5e02\u573a\u4e0e\u6a21\u578b\u4ed3\u5e93\uff0c\u8ba9\u7b97\u6cd5\u56e2\u961f\u4e00\u952e\u4e0a\u67b6\u3001\u5373\u523b\u53d8\u73b0\uff0c\u5f62\u6210\u79d1\u7814\u2192\u5546\u4e1a\u7684\u95ed\u73af\uff1b
\n– \u5bf9\u751f\u6001\u4fa7\uff0c\u5f00\u653e PaaS \u63a5\u53e3\u4e0e\u521b\u4f5c\u8005\u4e2d\u5fc3\uff0c\u9f13\u52b1\u66f4\u591a\u56fd\u4ea7\u82af\u7247\u3001\u6846\u67b6\u3001\u6a21\u578b\u5171\u540c\u63a5\u5165\uff0c\u6253\u7834\u201c\u82f1\u4f1f\u8fbe+CUDA\u201d\u5355\u6781\u683c\u5c40\u3002<\/p>\n

\n

\u5f69\u86cb\uff1a10 \u5143\u4f53\u9a8c\u91d1\u73b0\u5728\u5c31\u80fd\u9886<\/h2>\n

\u5982\u679c\u4f60\u4e5f\u5728\u4e3a\u201c\u7f3a\u5361\u201d\u53d1\u6101\uff0c\u4e0d\u59a8\u590d\u5236\u4e0b\u65b9\u94fe\u63a5\uff0c\u6ce8\u518c\u661f\u5b87\u667a\u7b97\u8d26\u53f7\u2014\u2014
\n\u7acb\u5373\u9886\u53d6 10 \u5143\u4f53\u9a8c\u91d1\uff0c0 \u5143\u5f00\u8dd1 10 \u5f20\u56fd\u4ea7 GPU >><\/a><\/strong>
\n\u65e0\u8bba\u662f\u590d\u73b0 LLaMA-Factory\u3001\u5fae\u8c03 ChatGLM3\uff0c\u8fd8\u662f\u8dd1 Stable Diffusion XL\uff0c\u5e73\u53f0\u90fd\u5df2\u9884\u7f6e\u955c\u50cf\u4e0e\u6570\u636e\u96c6\uff0c\u771f\u6b63\u505a\u5230\u201c\u6253\u5f00\u6d4f\u89c8\u5668\uff0c30 \u79d2\u8fdb\u5165\u6a21\u578b\u4e16\u754c\u201d\u3002<\/p>\n

\u5f53\u56fd\u4ea7\u82af\u7247\u3001\u5f00\u6e90\u6846\u67b6\u4e0e\u666e\u60e0\u7b97\u529b\u4ea4\u6c47\uff0c\u5927\u6a21\u578b\u79d1\u7814\u4e0d\u518d\u662f\u4e00\u573a\u201c\u9884\u7b97\u519b\u5907\u7ade\u8d5b\u201d\u3002\u661f\u5b87\u667a\u7b97\u613f\u505a\u90a3\u53f0\u85cf\u5728\u4e91\u7aef\u7684\u6c34\u7535\u7ad9\uff0c\u8ba9\u6bcf\u4e00\u6ef4\u7b97\u529b\uff0c\u90fd\u6d41\u5411\u771f\u6b63\u7684\u521b\u65b0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u201c\u56fd\u4ea7 GPU \u53ea\u80fd\u8dd1\u63a8\u7406\u3001\u4e0d\u80fd\u8bad\u5927\u6a21\u578b\u201d\u2014\u2014\u8fd9\u53e5\u88ab\u53cd\u590d\u5f15\u7528…<\/p>\n","protected":false},"author":2,"featured_media":1956,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1957","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":37,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1957","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/comments?post=1957"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/1957\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/1956"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=1957"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=1957"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=1957"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}