{"id":3117,"date":"2026-03-10T16:21:29","date_gmt":"2026-03-10T08:21:29","guid":{"rendered":"https:\/\/www.starverse-ai.com\/guide\/archives\/3117"},"modified":"2026-03-10T16:21:29","modified_gmt":"2026-03-10T08:21:29","slug":"%e4%bb%8e-0-%e5%88%b0-1-%e8%ae%ad%e7%bb%83%e5%9e%82%e7%9b%b4%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%9a%e6%98%9f%e5%ae%87%e6%99%ba%e7%ae%97-10-%e4%b8%87%e9%a2%84%e7%ae%97-30-%e5%a4%a9%e5%ae%9e%e6%88%98","status":"publish","type":"post","link":"https:\/\/www.starverse-ai.com\/guide\/archives\/3117","title":{"rendered":"\u4ece 0 \u5230 1 \u8bad\u7ec3\u5782\u76f4\u5927\u6a21\u578b\uff1a\u661f\u5b87\u667a\u7b97 10 \u4e07\u9884\u7b97 30 \u5929\u5b9e\u6218\u590d\u76d8"},"content":{"rendered":"
\"\u4ece<\/figure>\n
\n

\u201c\u5f53\u540c\u884c\u8fd8\u5728\u7ea0\u7ed3\u2018\u5927\u6a21\u578b\u5230\u5e95\u8981\u4e0d\u8981\u81ea\u5df1\u8bad\u2019\u65f6\uff0c\u6211\u4eec\u5df2\u7ecf\u628a 10B \u53c2\u6570\u7684\u6cd5\u5f8b\u5782\u76f4\u6a21\u578b\u8dd1\u8fdb\u4e86\u751f\u4ea7\u73af\u5883\u3002\u201d
\n\u2014\u2014\u67d0\u5f8b\u6240 AI \u5b9e\u9a8c\u5ba4\u8d1f\u8d23\u4eba\u5728\u6700\u8fd1\u4e00\u573a\u95ed\u95e8\u5206\u4eab\u4e0a\u7684\u8fd9\u53e5\u8bdd\uff0c\u8ba9\u201c10 \u4e07\u9884\u7b97\u300130 \u5929\u5468\u671f\u3001\u5355\u5361 80G\u201d\u6210\u4e3a\u5708\u5185\u65b0\u8c08\u8d44\u3002\u8fc7\u53bb\uff0c\u8bad\u7ec3\u767e\u4ebf\u7ea7\u6a21\u578b\u662f\u5927\u5382\u201c\u519b\u5907\u8d5b\u201d\uff0c\u5982\u4eca\u5374\u80fd\u5728\u4e00\u4e2a\u6708\u5185\u88ab\u4e00\u652f 5 \u4eba\u5c0f\u56e2\u961f\u5543\u4e0b\u6765\uff0c\u5e55\u540e\u5173\u952e\u6b63\u662f\u4e91\u5316\u7b97\u529b\u4e0e\u5de5\u7a0b\u4f18\u5316\u5e26\u6765\u7684\u201c\u5e73\u6c11\u5316\u201d\u62d0\u70b9\u3002<\/p>\n<\/blockquote>\n

\u4ece 0 \u5230 1 \u8bad\u7ec3\u5782\u76f4\u5927\u6a21\u578b\uff1a\u661f\u5b87\u667a\u7b97 10 \u4e07\u9884\u7b97 30 \u5929\u5b9e\u6218\u590d\u76d8<\/h2>\n

01 \u9700\u6c42\uff1a10B \u53c2\u6570\u6cd5\u5f8b\u5927\u6a21\u578b\uff0c500 GB \u8bad\u7ec3\u6570\u636e<\/h3>\n

\u9879\u76ee\u542f\u52a8\u524d\uff0c\u56e2\u961f\u7ed9\u81ea\u4e2a\u513f\u5b9a\u4e86\u4e09\u6761\u786c\u6760\uff1a
\n1. \u6a21\u578b\u89c4\u6a21 10B\uff0c\u65e2\u4fdd\u8bc1\u751f\u6210\u8d28\u91cf\uff0c\u53c8\u907f\u5f00\u201c\u5343\u4ebf\u7ea7\u201d\u8d44\u91d1\u9ed1\u6d1e\uff1b
\n2. \u6cd5\u5f8b\u573a\u666f C-eval \u5f97\u5206 \u226565\uff0c\u8d85\u8d8a\u540c\u89c4\u6a21\u901a\u7528\u5e95\u5ea7 5% \u4ee5\u4e0a\uff1b
\n3. \u7aef\u5230\u7aef 30 \u5929\u4ea4\u4ed8\uff0c\u9884\u7b97\u5929\u82b1\u677f 10 \u4e07\u5143\u3002<\/p>\n

\u6570\u636e\u4fa7\uff0c500 GB \u539f\u59cb\u8bed\u6599\u6765\u81ea\u88c1\u5224\u6587\u4e66\u3001\u5408\u540c\u8303\u672c\u3001\u6cd5\u8003\u9898\u5e93\u4e0e\u5b9e\u52a1\u95ee\u7b54\uff0c\u53bb\u91cd\u3001\u8131\u654f\u3001\u5206\u6bb5\u540e\u5f97\u5230 380 GB \u53ef\u7528\u8bed\u6599\uff0c\u518d\u7ecf\u89c4\u5219 + \u6a21\u578b\u4e8c\u6b21\u6e05\u6d17\uff0c\u6700\u7ec8\u4ea7\u51fa 120 \u4ebf token \u9ad8\u8d28\u91cf\u8bad\u7ec3\u96c6\u3002<\/p>\n

02 \u8d44\u6e90\u914d\u7f6e\uff1a16\u00d7A100 80G + 100TB \u5e76\u884c\u6587\u4ef6\u7cfb\u7edf<\/h3>\n

\u201c\u79df\u201d\u8fd8\u662f\u201c\u4e70\u201d\uff1f\u4e00\u5f20 A100 80G \u73b0\u8d27\u5e02\u4ef7 8 \u4e07+\uff0c\u81ea\u5efa\u96c6\u7fa4\u81f3\u5c11\u8981 128 \u4e07\u672c\u91d1\uff0c\u8fd8\u6ca1\u7b97\u673a\u623f\u3001\u7535\u8d39\u548c\u8fd0\u7ef4\u3002\u56e2\u961f\u628a\u76ee\u6807\u9501\u5b9a\u5728GPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u5e73\u53f0\uff0c\u5bf9\u6bd4\u4e86\u5e02\u9762 7 \u5bb6\u5382\u5546\u540e\uff0c\u6700\u7ec8\u62bc\u6ce8\u661f\u5b87\u667a\u7b97\uff1a
\n– 16 \u5f20 A100 80G SXM \u7248\uff0cNVLink \u5168\u4e92\u8054\uff0c\u5355\u673a 600G \u5e26\u5bbd\uff1b
\n– 100TB \u5e76\u884c\u6587\u4ef6\u7cfb\u7edf\uff0cPOSIX \u63a5\u53e3\uff0c1.5 \u4e07 MB\/s \u805a\u5408\u5e26\u5bbd\uff0c\u65e0\u9700\u989d\u5916\u914d\u7f6e\u5bf9\u8c61\u5b58\u50a8\u7f51\u5173\uff1b
\n– \u6309\u5c0f\u65f6\u8ba1\u8d39\uff0c\u968f\u65f6\u5f00\u505c\uff0c\u652f\u6301 Spot \u5b9e\u4f8b\uff0c\u6210\u672c\u518d\u964d 18%\u3002 <\/p>\n

\u6574\u4e2a\u6d41\u7a0b\u4ece\u6ce8\u518c\u5230\u62c9\u8d77\u96c6\u7fa4\u53ea\u82b1\u4e86 12 \u5206\u949f\u2014\u2014\u65b0\u7528\u6237\u6ce8\u518c\u9001\u7684 10 \u5143\u4f53\u9a8c\u91d1\u521a\u597d\u62b5\u6389\u9996\u65e5\u6d4b\u8bd5\u673a\u65f6\uff0c\u771f\u6b63\u201c\u96f6\u95e8\u69db\u201d\u8e29\u5751\u3002<\/p>\n

03 \u8bad\u7ec3\u7b56\u7565\uff1a\u6570\u636e\u5e76\u884c + ZeRO-3\uff0cMFU \u62c9\u5230 58%<\/h3>\n

\u6cd5\u5f8b\u5927\u6a21\u578b\u5bf9\u201c\u4e8b\u5b9e\u4e00\u81f4\u6027\u201d\u6781\u5ea6\u654f\u611f\uff0c\u56e2\u961f\u6cbf\u7528 LLaMA-2 10B \u57fa\u7840\u7ed3\u6784\uff0c\u91c7\u7528\u4ee5\u4e0b\u5de5\u7a0b\u7ec4\u5408\uff1a
\n1. \u6570\u636e\u5e76\u884c<\/strong> 16 \u8def + ZeRO-3 offload\uff0c\u628a\u4f18\u5316\u5668\u72b6\u6001\u3001\u68af\u5ea6\u3001\u53c2\u6570\u5168\u90e8\u5207\u7247\uff0c\u5355\u5361\u663e\u5b58\u5360\u7528\u538b\u5230 62G\uff1b
\n2. FlashAttention-2<\/strong> \u66ff\u6362\u539f\u751f Attention\uff0c\u957f\u6587\u672c 4k\u21928k \u65e0\u9700\u989d\u5916\u663e\u5b58\uff1b
\n3. \u878d\u5408\u7b97\u5b50<\/strong> \u4e0e cuDNN 8.9<\/strong> \u6df1\u5ea6\u7f16\u8bd1\uff0cGPU \u5185\u6838\u5229\u7528\u7387 MFU \u4ece 42% \u63d0\u5347\u5230 58%\uff0c\u76f8\u5f53\u4e8e\u767d\u6361 3.2 \u5f20\u5361\u3002 <\/p>\n

\u8bad\u7ec3 3.2 \u4e07\u4ebf token\uff0c\u603b\u8ba1 22.5 \u5c0f\u65f6\uff0ccheckpoint \u6bcf 2 \u5c0f\u65f6\u5f02\u6b65\u4e0a\u4f20\u81f3\u661f\u5b87\u667a\u7b97\u4e91\u5b58\u50a8\uff0c\u65ad\u70b9\u7eed\u8bad\u96f6\u611f\u77e5\u3002<\/p>\n

04 \u6210\u672c\u62c6\u89e3\uff1a\u673a\u65f6 8.5 \u4e07\uff0c\u6570\u636e\u6e05\u6d17 0.8 \u4e07\uff0c\u5408\u8ba1 9.3 \u4e07<\/h3>\n\n\n\n\n\n\n\n\n
\u9879\u76ee<\/th>\n\u91d1\u989d\uff08\u5143\uff09<\/th>\n\u5907\u6ce8<\/th>\n<\/tr>\n<\/thead>\n
GPU\u670d\u52a1\u5668\u79df\u7528<\/td>\n85 000<\/td>\n16\u00d7A100 80G\uff0c30 \u5929\uff0c\u542b 100TB \u5e76\u884c\u6587\u4ef6\u7cfb\u7edf<\/td>\n<\/tr>\n
\u6570\u636e\u6e05\u6d17<\/td>\n8 000<\/td>\n\u89c4\u5219\u5f15\u64ce + \u5c0f\u6a21\u578b\u8fc7\u6ee4 + \u4eba\u5de5\u62bd\u68c0 5%<\/td>\n<\/tr>\n
\u5176\u4ed6\u6742\u8d39<\/td>\n300<\/td>\n\u516c\u7f51\u6d41\u91cf\u3001\u76d1\u63a7\u65e5\u5fd7<\/td>\n<\/tr>\n
\u603b\u8ba1<\/strong><\/td>\n93 000<\/strong><\/td>\n\u9884\u7b97\u7ed3\u4f59 7 000\uff0c\u76f4\u63a5\u5212\u5165\u4e0b\u4e00\u8f6e SFT \u8fed\u4ee3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u5bf9\u6bd4\u81ea\u5efa\u65b9\u6848\uff0c\u4ec5\u8d44\u672c\u652f\u51fa\u5c31\u8282\u7701 120 \u4e07+\uff0c\u8fd9\u8fd8\u6ca1\u7b97\u7535\u8d39\u4e0e\u8fd0\u7ef4\uff1b\u5bf9\u6bd4\u540c\u7c7bGPU\u4e91\u4e3b\u673a<\/strong>\u6309\u5305\u6708\u8ba1\u8d39\uff0c\u661f\u5b87\u667a\u7b97 Spot \u6a21\u5f0f\u518d\u7701 18%\uff0c\u771f\u6b63\u8ba9\u201c\u8bad\u7ec3\u81ea\u7531\u201d\u843d\u5230 10 \u4e07\u7ea7\u3002<\/p>\n

05 \u7ed3\u679c\uff1aC-eval \u6cd5\u5f8b\u7c7b 68.4\uff0c\u9886\u5148\u540c\u89c4\u6a21\u901a\u7528\u6a21\u578b 9%<\/h3>\n

\u7ecf\u8fc7 5-shot \u8bc4\u6d4b\uff0c\u65b0\u6a21\u578b\u5728 C-eval \u6cd5\u5f8b\u5927\u7c7b\u53d6\u5f97 68.4 \u5206\uff0c\u8f83\u5b98\u65b9\u5f00\u6e90 10B \u901a\u7528\u6a21\u578b\u63d0\u9ad8 5.7 \u5206\uff0c\u76f8\u5bf9\u589e\u76ca 9%\u3002\u5728\u5f8b\u6240\u5185\u90e8\u771f\u5b9e\u573a\u666f\u2014\u2014\u201c\u5408\u540c\u6761\u6b3e\u98ce\u9669\u8bc6\u522b\u201d\u4e0e\u201c\u8d77\u8bc9\u72b6\u751f\u6210\u201d\u76f2\u6d4b\u4e2d\uff0c\u5f8b\u5e08\u91c7\u7eb3\u7387\u5206\u522b\u8fbe\u5230 78% \u4e0e 82%\uff0c\u8d85\u8fc7 GPT-3.5-turbo 6 \u4e2a\u767e\u5206\u70b9\u3002\u66f4\u5173\u952e\u7684\u662f\uff0c\u4ece\u6570\u636e\u51c6\u5907\u5230\u6a21\u578b\u4ea4\u4ed8\u4ec5 30 \u5929\uff0c\u6bd4\u4f20\u7edf\u65b9\u6848\u538b\u7f29 40% \u5468\u671f\uff0c\u8ba9\u4e1a\u52a1\u65b9\u63d0\u524d\u4e00\u4e2a\u5b63\u5ea6\u4e0a\u7ebf AI \u52a9\u7406\u3002<\/p>\n

\u7ecf\u9a8c\u6c89\u6dc0\uff1a\u8ba9\u201c\u8bad\u7ec3\u5927\u6a21\u578b\u201d\u6210\u4e3a\u666e\u901a\u56e2\u961f\u7684\u6807\u51c6\u52a8\u4f5c<\/h2>\n
    \n
  1. \u7b97\u529b\u4e91\u5316\u662f\u524d\u63d0\uff1a\u65e0\u9700\u4e00\u6b21\u6027\u6295\u5165\u767e\u4e07\u786c\u4ef6\uff0cGPU\u670d\u52a1\u5668\u79df\u7528<\/strong>\u628a\u8d44\u672c\u652f\u51fa\u53d8\u6210\u53ef\u9884\u6d4b\u7684\u8fd0\u8425\u652f\u51fa\uff1b <\/li>\n
  2. \u5de5\u7a0b\u4f18\u5316\u662f\u6838\u5fc3\uff1aZeRO-3 + FlashAttention \u8fd9\u7c7b\u8f6f\u4ef6\u7ea2\u5229\uff0c\u80fd\u628a\u663e\u5b58\u548c MFU \u540c\u65f6\u201c\u69a8\u5e72\u201d\uff1b <\/li>\n
  3. \u751f\u6001\u5e73\u53f0\u662f\u52a0\u901f\u5668\uff1a\u661f\u5b87\u667a\u7b97\u5185\u7f6e\u7684\u6a21\u578b\u4ed3\u5e93\u3001\u6570\u636e\u96c6\u3001\u6301\u4e45\u5316\u4e91\u786c\u76d8\uff0c\u8ba9\u6570\u636e\u642c\u8fd0\u4e0e\u7248\u672c\u7ba1\u7406\u4ece\u201c\u5468\u201d\u7f29\u5230\u201c\u5c0f\u65f6\u201d\uff1b <\/li>\n
  4. \u6210\u672c\u53ef\u63a7\u662f\u7ed3\u679c\uff1a10 \u4e07\u5143\u8bad 10B \u6a21\u578b\u4e0d\u662f\u5671\u5934\uff0c\u800c\u662f\u4e91\u65f6\u4ee3\u7684\u65b0\u57fa\u51c6\u7ebf\u3002 <\/li>\n<\/ol>\n

    \u5982\u679c\u4f60\u4e5f\u60f3\u590d\u73b0\u8fd9\u5957\u6d41\u7a0b\uff0c\u661f\u5b87\u667a\u7b97\u5df2\u628a\u955c\u50cf\u3001\u811a\u672c\u4e0e checkpoint \u4e0a\u67b6\u5230\u516c\u5171\u8d44\u6e90\u5e93\uff0c\u65b0\u7528\u6237\u6ce8\u518c\u5373\u9886 10 \u5143\u4f53\u9a8c\u91d1\uff0c\u8db3\u591f\u8dd1\u901a 2\u00d7A100 8 \u5c0f\u65f6\u70ed\u8eab\u5b9e\u9a8c\u3002\u6253\u5f00 https:\/\/www.starverse-ai.com<\/a>\uff0c\u641c\u7d22\u201c\u6cd5\u5f8b 10B\u201d\u5373\u53ef\u4e00\u952e\u514b\u9686\u73af\u5883\u3002 <\/p>\n

    \n

    \u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u7b97\u529b\u4e0d\u518d\u662f\u62a4\u57ce\u6cb3\uff0c\u5de5\u7a0b\u6548\u7387\u548c\u5546\u4e1a\u5316\u901f\u5ea6\u624d\u662f\u3002\u628a\u6602\u8d35\u7684 GPU \u7559\u7ed9\u4e91\u5382\u5546\uff0c\u628a\u5b9d\u8d35\u7684\u521b\u610f\u7559\u7ed9\u81ea\u5df1\u2014\u2014\u8fd9\u6216\u8bb8\u5c31\u662f\u201c\u5e73\u6c11\u5316 AI\u201d\u6700\u6027\u611f\u7684\u7b54\u6848\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"

    \u201c\u5f53\u540c\u884c\u8fd8\u5728\u7ea0\u7ed3\u2018\u5927\u6a21\u578b\u5230\u5e95\u8981\u4e0d\u8981\u81ea\u5df1\u8bad\u2019\u65f6\uff0c\u6211\u4eec\u5df2\u7ecf\u628a 1…<\/p>\n","protected":false},"author":2,"featured_media":3116,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3117","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-zixun"],"views":73,"_links":{"self":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3117","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=3117"}],"version-history":[{"count":0,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/posts\/3117\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media\/3116"}],"wp:attachment":[{"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/media?parent=3117"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/categories?post=3117"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.starverse-ai.com\/guide\/wp-json\/wp\/v2\/tags?post=3117"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}