{"id":26774,"date":"2025-07-21T08:42:32","date_gmt":"2025-07-21T00:42:32","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=26774"},"modified":"2025-07-21T08:43:05","modified_gmt":"2025-07-21T00:43:05","slug":"%e5%85%a5%e5%b1%80-ai-infra%ef%bc%9a%e7%a8%8b%e5%ba%8f%e5%91%98%e5%bf%85%e9%a1%bb%e4%ba%86%e8%a7%a3%e7%9a%84-ai-%e7%b3%bb%e7%bb%9f%e8%ae%be%e8%ae%a1%e4%b8%8e%e6%8c%91%e6%88%98%e7%9f%a5%e8%af%86","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/07\/21\/26774\/","title":{"rendered":"\u5165\u5c40 AI Infra\uff1a\u7a0b\u5e8f\u5458\u5fc5\u987b\u4e86\u89e3\u7684 AI \u7cfb\u7edf\u8bbe\u8ba1\u4e0e\u6311\u6218\u77e5\u8bc6"},"content":{"rendered":"<div><img data-dominant-color=\"24467b\" data-has-transparency=\"false\" style=\"--dominant-color: #24467b;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-26779\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-300x167-1-1.jpg\" width=\"300\" height=\"167\" alt=\"\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-300x167-1-1.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-300x167-1-1-150x84.jpg 150w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/div>\n<div><\/div>\n<div class=\"article-desc\">\u672c\u6587\u5c06\u5206\u4eab\u4f20\u7edf\u540e\u53f0\u5de5\u7a0b\u5e08\u79ef\u7d2f\u7684\u6280\u672f\u6808\u548c\u65b9\u6cd5\u8bba\uff0c\u5982\u4f55\u5ef6\u7eed\u5e76\u8fc1\u79fb\u5230 AI \u7cfb\u7edf\uff0c\u5e76\u7cfb\u7edf\u6027\u62c6\u89e3 AI Infra \u7684\u786c\u4ef6\u3001\u8f6f\u4ef6\u3001\u8bad\u7ec3\u548c\u63a8\u7406\u6311\u6218\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-01a18e2f=\"\" data-element=\"root\">\n<p>\u672c\u6587\u5c06\u5206\u4eab\u4f20\u7edf\u540e\u53f0\u5de5\u7a0b\u5e08\u79ef\u7d2f\u7684\u6280\u672f\u6808\u548c\u65b9\u6cd5\u8bba\uff0c\u5982\u4f55\u5ef6\u7eed\u5e76\u8fc1\u79fb\u5230 AI \u7cfb\u7edf\uff0c\u5e76\u7cfb\u7edf\u6027\u62c6\u89e3 AI Infra \u7684\u786c\u4ef6\u3001\u8f6f\u4ef6\u3001\u8bad\u7ec3\u548c\u63a8\u7406\u6311\u6218\u3002<\/p>\n<p>&nbsp;<\/p>\n<p>AI Infra \u548c\u4f20\u7edf Infra \u6709\u4ec0\u4e48\u533a\u522b\uff1f\u7a0b\u5e8f\u5458\u79ef\u7d2f\u7684\u6280\u672f\u6808\u548c\u65b9\u6cd5\u8bba\uff0c\u5982\u4f55\u590d\u7528\u5230 AI \u7cfb\u7edf\u67b6\u6784\u8bbe\u8ba1\u4e0a\uff1f<\/p>\n<p>\u968f\u7740\u5927\u6a21\u578b\u6280\u672f\u7684\u7206\u53d1\uff0cAI Infra \u5df2\u6210\u4e3a\u57fa\u7840\u8bbe\u65bd\u9886\u57df\u7684\u6838\u5fc3\u6218\u573a\u3002\u8fc7\u53bb1\u5e74\u591a\u7684\u65f6\u95f4\uff0c\u6211\u4eecQQ\u57fa\u7840\u67b6\u6784\u7b97\u6cd5\u5de5\u7a0b\u56e2\u961f\u843d\u5730\u4e86\u591a\u4e2a\u5927\u6a21\u578b\u5e94\u7528\uff0c\u5305\u62ec\u8bed\u97f3\u5408\u6210\u5927\u6a21\u578b\u3001\u5185\u5bb9\u7406\u89e3\u591a\u6a21\u6001\u5927\u6a21\u578b\u3001\u751f\u6210\u5f0f\u63a8\u8350\u5927\u6a21\u578b\uff0c\u8dd1\u901a\u5927\u6a21\u578b\u8bad\u7ec3\u5230\u63a8\u7406\u7684\u5168\u94fe\u8def\u3002\u8e29\u4e86\u5f88\u591a\u5751\uff0c\u4e5f\u79ef\u7d2f\u4e86\u4e0d\u5c11\u7ecf\u9a8c\u3002\u672c\u6587\u5c06\u5206\u4eab\u4f20\u7edf\u540e\u53f0\u5de5\u7a0b\u5e08\u79ef\u7d2f\u7684\u6280\u672f\u6808\u548c\u65b9\u6cd5\u8bba\uff0c\u5982\u4f55\u5ef6\u7eed\u5e76\u8fc1\u79fb\u5230 AI \u7cfb\u7edf\uff0c\u5e76\u7cfb\u7edf\u6027\u62c6\u89e3 AI Infra \u7684\u786c\u4ef6\u3001\u8f6f\u4ef6\u3001\u8bad\u7ec3\u548c\u63a8\u7406\u6311\u6218\u3002<\/p>\n<p><img data-dominant-color=\"24467b\" data-has-transparency=\"false\" style=\"--dominant-color: #24467b;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-26780\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1.jpg\" width=\"1248\" height=\"696\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1.jpg 1248w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1-300x167.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1-1024x571.jpg 1024w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1-150x84.jpg 150w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/07\/7345a0039eee797866b20071e86ceb710777a1-1-1-768x428.jpg 768w\" sizes=\"auto, (max-width: 1248px) 100vw, 1248px\" \/><\/p>\n<h3>\u4e00\u3001\u786c\u4ef6\u6f14\u8fdb<\/h3>\n<p>\u7ecf\u6d4e\u57fa\u7840\u51b3\u5b9a\u4e0a\u5c42\u5efa\u7b51\u3002\u8f6f\u4ef6\u5c42\u9762\u7684\u67b6\u6784\u8bbe\u8ba1\uff0c\u65e0\u6cd5\u8131\u79bb\u786c\u4ef6\u7ea6\u675f\u3002\u4e86\u89e3\u73b0\u4ee3 AI \u786c\u4ef6\u7279\u6027\u975e\u5e38\u6709\u5fc5\u8981\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/e638e3e92012205a2f764084768a4c6b7e2fb1.webp\" data-type=\"block\" \/><\/p>\n<p>\u4e00\u53f0\u9ad8\u6027\u80fd\u7684GPU\u670d\u52a1\u5668\u53ef\u4ee5\u6362\u4e00\u5957\u6df1\u5733\u623f\u5b50<\/p>\n<h4>1. \u4eceCPU\u4e3a\u4e2d\u5fc3\u5230GPU\u4e3a\u4e2d\u5fc3<\/h4>\n<p>\u4f20\u7edf\u57fa\u7840\u8bbe\u65bd\u4ee5 CPU \u4e3a\u6838\u5fc3\uff0c\u901a\u8fc7\u591a\u7ebf\u7a0b\u548c\u5fae\u670d\u52a1\u6784\u5efa\u5206\u5e03\u5f0f\u7cfb\u7edf\uff0c\u5904\u7406\u9ad8\u5e76\u53d1\u8bf7\u6c42\uff08\u5982 Web \u670d\u52a1\uff09\u3002\u8fd9\u4e9b\u90fd\u6709\u6210\u719f\u7684\u65b9\u6cd5\u8bba\u4e86\uff08\u5982&#8221;\u6d77\u91cf\u670d\u52a1\u4e4b\u9053&#8221;\uff09\u3002\u4e3b\u8981\u5de5\u4f5c\u662f\u903b\u8f91\u4e8b\u52a1\u7684\u5904\u7406\uff0c\u74f6\u9888\u5728\u7f51\u7edc I\/O \u548c CPU \u6838\u5fc3\u6570\u91cf\u3002\u53d1\u5c55\u5230\u4eca\u5929\uff0c\u786c\u4ef6\u5df2\u7ecf\u5f88\u5c11\u662f\u5236\u7ea6 CPU \u7cfb\u7edf\u8bbe\u8ba1\u7684\u74f6\u9888\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/62f825e606bdf373b4a210ae665ad96359c4b8.webp\" data-type=\"block\" \/><\/p>\n<p>\u800c AI Infra \u4ee5 GPU \u4e3a\u6838\u5fc3\uff0c\u5176\u8bbe\u8ba1\u76ee\u6807\u4ece\u903b\u8f91\u4e8b\u52a1\u5904\u7406\u8f6c\u5411\u9ad8\u541e\u5410\u6d6e\u70b9\u8ba1\u7b97\u3002\u6b64\u65f6CPU \u591a\u7ebf\u7a0b\u88ab GPU \u5e76\u884c\u8ba1\u7b97\u66ff\u4ee3\uff0c\u5185\u5b58\u88ab\u663e\u5b58\u66ff\u4ee3\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0cH20\u5355\u536196GB\u663e\u5b58\uff0c\u53ef\u4ee5\u63d0\u4f9b44TFlops\u7684\u5355\u7cbe\u5ea6\u6d6e\u70b9\u8fd0\u7b97\uff0c\u7b97\u529b\u548c\u8bbf\u5b58\u5e26\u5bbd\u662f\u4e3b\u6d41CPU\u6570\u5341\u500d\u751a\u81f3\u6570\u767e\u500d\u3002\u6bcf\u53f0\u673a\u5668\u5b89\u88c58\u5361=768GB\u663e\u5b58\uff0c\u53e6\u5916\u8fd8\u6709 CPU 192\u6838384\u7ebf\u7a0b + 2.3TB \u5185\u5b58\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/a51dd9b620ba527e7a26456658b6eadfb98ff8.webp\" data-type=\"block\" \/><\/p>\n<p>\u4e3a\u4ec0\u4e48 GPU \u4f1a\u6210\u4e3a\u6838\u5fc3\uff1f\u662f\u56e0\u4e3a LLM \u5927\u6a21\u578b\u6bcf\u6b21\u751f\u6210\u4e00\u4e2a token\uff0c\u90fd\u9700\u8981\u8bfb\u53d6\u5168\u91cf\u7684\u6a21\u578b\u53c2\u6570\u3002\u4f20\u7edf\u7684 CPU + \u5185\u5b58\u7684\u7b97\u529b\u548c\u5e26\u5bbd\u65e0\u6cd5\u6ee1\u8db3\u5982\u6b64\u6050\u6016\u7684\u8ba1\u7b97\u5bc6\u5ea6\uff0c\u8ba1\u7b97\u548c\u901a\u4fe1\u90fd\u5fc5\u987b\u8f6c\u79fb\uff08offload\uff09\u5230 GPU \u5185\u5b8c\u6210\u3002CPU \u6210\u4e3a\u6570\u636e\u642c\u8fd0\u5de5\u548c\u201c\u8f85\u52a9\u5904\u7406\u5668\u201d\u3002<\/p>\n<p>\u4e3a\u4e86\u66f4\u76f4\u89c2\u5730\u7406\u89e3\u8fd9\u4e2a\u8ba1\u7b97\u5bc6\u5ea6\uff0c\u6211\u4eec\u505a\u4e00\u4e2a\u7b80\u5355\u7684\u8ba1\u7b97\u3002\u4e0d\u8003\u8651\u8ba1\u7b97\u7684\u5ef6\u65f6\uff0cLLM \u5927\u6a21\u578b\u751f\u6210\u4e00\u4e2a token \u7684\u8017\u65f6\u516c\u5f0f\u8ba1\u7b97\u4e3a\uff1a<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/b9abcb155791fafe486491414be4f2b7a4e1c9.webp\" data-type=\"block\" \/><\/p>\n<p>\u4ee5 DeepSeek-R1-671B-A37B-FP8 \u6a21\u578b\u4e3a\u4f8b\uff0c\u8ba1\u7b97\u4e00\u4e2a token \u8017\u65f6\uff0cH20 \u4e3a 37B \u00d7 1byte \u00f7 4000GB\/s =\u00a09ms\uff0c\u5982\u679c\u662f CPU \u5219\u4e3a 37B \u00d7 1byte \u00f7 64GB\/s =\u00a0578ms\u3002<\/p>\n<h4>2. \u4ece&#8221;\u53bbIOE&#8221;\u5230&#8221;AI\u5927\u578b\u673a&#8221;<\/h4>\n<p>\u663e\u800c\u6613\u89c1\uff0c\u6211\u4eec\u7684\u73b0\u5728\u8eab\u5904\u65b0\u7684\u4e00\u8f6e\u70c8\u706b\u70f9\u6cb9\u7684\u786c\u4ef6\u9769\u547d\u7684\u5386\u53f2\u8fdb\u7a0b\u4e2d\uff0c\u5404\u79cd\u4e13\u7528\u786c\u4ef6\u3001\u4e13\u7528\u7f51\u7edc\u5c42\u51fa\u4e0d\u7a77\u3002DeepSeek-R1 \u548c QWen3-235B \u5343\u4ebf\u7ea7\u53c2\u6570\u8bad\u7ec3\u9700\u5343\u5361 GPU \u96c6\u7fa4\u534f\u540c\uff0c\u901a\u8fc7\u4e13\u7528\u7f51\u7edc\u4e92\u8054\u6784\u5efa\u201cAI\u8d85\u7b97\u201d\uff0c\u5176\u8bbe\u8ba1\u903b\u8f91\u4e0e\u4ee5\u524d\u7684 IBM \u5927\u578b\u673a\u60ca\u4eba\u76f8\u4f3c\u2014\u2014\u4ee5\u786c\u4ef6\u96c6\u4e2d\u5316\u6362\u53d6\u6781\u81f4\u6027\u80fd\u4e0e\u53ef\u9760\u6027\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/d6f43fa45ea26e3bd508256b2efa8f30d6cebc.webp\" data-type=\"block\" \/><\/p>\n<p>IBM\u5927\u578b\u673a<\/p>\n<p>\u4f20\u7edf Infra \u7684\u5206\u5e03\u5f0f\u7406\u5ff5\u8c8c\u4f3c\u5728 AI \u65f6\u4ee3\u5931\u6548\u4e86\u3002\u4f20\u7edf Infra \u8ffd\u6c42\u6a2a\u5411\u6269\u5c55\uff0c\u800c AI Infra \u5448\u73b0 \u201cAI \u5927\u578b\u673a\u201d\u7279\u6027\uff0c\u662f\u56e0\u4e3a\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u7684\u53ef\u4ee5\u5bb9\u5fcd\u6beb\u79d2\u7ea7\u5ef6\u8fdf\uff0c\u4f46 AI \u96c6\u7fa4\u4e0d\u884c\uff0cGPU \u7684\u7b97\u529b\u662f CPU \u7684\u6570\u767e\u500d\uff0c\u5fae\u79d2\u7ea7\u7684\u5ef6\u65f6\u7b49\u5f85\u4e5f\u4f1a\u9020\u6210\u5f88\u5927\u7684\u7b97\u529b\u635f\u8017\uff0c\u9700\u8981\u786c\u4ef6\u7684\u9ad8\u5ea6\u96c6\u6210\u3002\u5728\u53ef\u9884\u89c1\u76841-3\u5e74\u7684\u672a\u6765\uff0c\u8fd9\u6837\u7684\u4e13\u7528\u786c\u4ef6+\u7f51\u7edc\u7684\u96c6\u4e2d\u5f0f\u67b6\u6784\u5f88\u96be\u53d1\u751f\u6bd4\u8f83\u5927\u7684\u6539\u53d8\u3002<\/p>\n<p>\u56de\u987e\u5386\u53f2\uff0c\u6211\u4eec\u603b\u662f\u5728\u5bfb\u6c42\u79d1\u6280\u5e73\u6743\u3002\u524d\u4eba\u63a8\u52a8\u201c\u53bbIOE\u201d\uff08IBM\u5c0f\u578b\u673a\u3001Oracle\u6570\u636e\u5e93\u3001EMC\u5b58\u50a8\uff09\uff0c\u7528\u5206\u5e03\u5f0f\u5ec9\u4ef7x86 pc\u673a\u66ff\u4ee3\u96c6\u4e2d\u5f0f\u9ad8\u7aef\u786c\u4ef6\uff0c\u672c\u8d28\u4e0a\u662f\u5229\u7528\u8f6f\u4ef6\u521b\u65b0\u91cd\u6784\u4e00\u4e2a\u9ad8\u53ef\u7528+\u4f4e\u6210\u672c\u7684\u4e92\u8054\u7f51\u57fa\u7840\u8bbe\u65bd\u3002&#8221;AI\u5927\u578b\u673a&#8221;\u662f\u6280\u672f\u53d1\u5c55\u5fc5\u7531\u4e4b\u8def\uff0c\u4f46\u4e0d\u662f\u7ec8\u6781\u5f62\u6001\u3002\u957f\u671f\uff085\u5e74\uff09\u6765\u770b\uff0c\u5fc5\u7136\u4f1a\u51fa\u73b0 &#8220;AI \u53bb NVIDIA \u5316&#8221;\uff0c\u91cd\u6f14\u201c\u53bb IOE\u201d\u7684\u5386\u53f2\u3002<\/p>\n<h3>\u4e8c\u3001\u8f6f\u4ef6\u6f14\u8fdb<\/h3>\n<p>\u8bf4\u5b8c\u786c\u4ef6\u4f53\u7cfb\u7684\u9769\u547d\uff0c\u63a5\u4e0b\u6765\u518d\u5173\u6ce8\u4e0b\u8f6f\u4ef6\u5c42\u9762\u7684\u53d8\u5316\u3002<\/p>\n<p>\u76f8\u6bd4\u4f20\u7edf\u540e\u53f0\u5e94\u7528\u7684\u589e\u5220\u67e5\u6539\uff0cAI \u5e94\u7528\u7684\u65b0\u8303\u5f0f\u662f\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u3002\u6a21\u578b\u8bad\u7ec3\u662f\u6307\u901a\u8fc7\u6d77\u91cf\u6570\u636e\u62df\u5408\u51fa\u4e00\u4e2a\u590d\u6742\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\uff0c\u63a8\u7406\u5c31\u662f\u5229\u7528\u8bad\u7ec3\u597d\u7684\u795e\u7ecf\u7f51\u7edc\u6a21\u578b\u8fdb\u884c\u8fd0\u7b97\uff0c\u8f93\u5165\u7684\u65b0\u6570\u636e\u6765\u83b7\u5f97\u65b0\u7684\u7ed3\u8bba\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u8bad\u7ec3\u5c31\u662f\u6839\u636e &lt;\u5e74\u9f84, \u8eab\u9ad8&gt; \u7684\u5206\u5e03\u4f7f\u7528\u6700\u5c0f\u4e8c\u4e58\u6cd5\u62df\u5408\u6a21\u578b y = ax + b\uff0c\u63a8\u7406\u5c31\u662f\u5229\u7528\u8fd9\u4e2a\u6a21\u578b y = ax + b\uff0c\u8f93\u5165\u4e00\u4e2a\u65b0\u7684\u5e74\u9f84\uff0c\u9884\u6d4b\u8eab\u9ad8\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/5170d8b6013256950d0438dcb5b3b88d7941b1.webp\" data-type=\"block\" \/><\/p>\n<h4>1. \u6df1\u5ea6\u5b66\u4e60\u6846\u67b6<\/h4>\n<p>\u5de5\u6b32\u5584\u5176\u4e8b\uff0c\u5fc5\u5148\u5229\u5176\u5668\u3002\u4f20\u7edf\u540e\u53f0\u5e94\u7528\u4f9d\u8d56 tRPC \u6216 Spring \u7b49\u5fae\u670d\u52a1\u6846\u67b6\uff0c\u5e2e\u52a9\u6211\u4eec\u5c4f\u853d\u8d1f\u8f7d\u5747\u8861\u3001\u7f51\u7edc\u901a\u4fe1\u7b49\u5e95\u5c42\u7ec6\u8282\uff0c\u6211\u4eec\u53ef\u4ee5\u628a\u7cbe\u529b\u653e\u5728\u4e1a\u52a1\u5b9e\u73b0\u4e0a\u3002<\/p>\n<p>\u4e0e\u4e4b\u76f8\u4f3c\uff0cAI \u5e94\u7528\u5219\u4f9d\u8d56\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u3002\u5982\u679c\u6ca1\u6709\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u6211\u4eec\u5c31\u53ef\u80fd\u9677\u5165\u5728\u832b\u832b\u7684\u6570\u5b66\u6df1\u6e0a\u4e2d\uff0c\u6323\u624e\u4e8e\u75db\u82e6\u7684 GPU \u7f16\u7a0b\u6ce5\u6f6d\u91cc\u3002\u6709\u4e86\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u6211\u4eec\u624d\u53ef\u4ee5\u628a\u6240\u6709\u7cbe\u529b\u82b1\u5728\u8bbe\u8ba1\u6a21\u578b\u548c\u521b\u65b0\u672c\u8eab\u4e0a\uff0c\u800c\u4e0d\u7528\u5173\u6ce8\u5e95\u5c42\u7684\u5b9e\u73b0\u7ec6\u8282\uff0c\u6781\u5927\u964d\u4f4e\u4e86 AI \u5e94\u7528\u7684\u95e8\u69db\u3002<\/p>\n<p>\u5927\u5bb6\u53ef\u80fd\u542c\u8bf4\u8fc7\u4e0d\u540c\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u2014\u2014Tensorflow\uff0cPyTorch\u3002\u73b0\u5728\u662f2025\u5e74\uff0c\u4e0d\u7528\u7ea0\u7ed3\u9009\u54ea\u4e2a\uff0c\u56e0\u4e3a PyTorch \u5c31\u662f AI \u6a21\u578b\u8bad\u7ec3\u3001\u63a8\u7406\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u4e8b\u5b9e\u6807\u51c6\u3002\u5f00\u6e90\u6a21\u578b\u548c\u4ee3\u7801\u90fd\u662f PyTorch \u4e00\u8fb9\u5012\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/a736a640271f799a9db40938effa8a0fd93f2c.webp\" data-type=\"block\" \/><\/p>\n<p>\u5f97\u76ca\u4e8e\u52a8\u6001\u8ba1\u7b97\u56fe\u3001\u81ea\u52a8\u5fae\u5206\u548c\u4e30\u5bcc\u7684 Tensor\u00a0\u64cd\u4f5c\u7b97\u5b50\uff0cPyTorch \u80fd\u5e2e\u52a9\u6211\u4eec\u5feb\u901f\u5b9e\u73b0\u6a21\u578b\u8bbe\u8ba1\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u53ea\u9700\u8981\u63cf\u8ff0\u6a21\u578b\u7ed3\u6784+\u5f85\u5b66\u4e60\u7684\u7f51\u7edc\u53c2\u6570\uff0c\u4e0d\u9700\u8981\u5173\u5fc3\u6570\u5b66\u8ba1\u7b97\u548c GPU \u7f16\u7a0b\u7684\u7ec6\u8282\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/e8fc1c303f1e8f800264897121ea578aa3db42.webp\" data-type=\"block\" \/><\/p>\n<h4>2. GPU \u7f16\u7a0b<\/h4>\n<p>\u7edd\u5927\u90e8\u5206\u7684 AI \u5e94\u7528\uff0c\u7684\u786e\u4e0d\u9700\u8981\u6211\u4eec\u624b\u5199\u6570\u5b66\u8ba1\u7b97\u7684 GPU \u4ee3\u7801\u3002\u4f46\u4e3a\u4e86\u6ee1\u8db3\u6a21\u578b\u521b\u65b0\u7684\u9700\u6c42\uff0c\u6709\u5fc5\u8981\u5b66\u4e60 GPU \u7f16\u7a0b\u3002\u4f8b\u5982 Meta \u53d1\u5e03\u7684 HSTU \u751f\u6210\u5f0f\u63a8\u8350\u6a21\u578b\uff0c\u6838\u5fc3\u7684 hstu_attn \u8ba1\u7b97\uff0c\u5982\u679c\u76f4\u63a5\u7528 PyTorch \u6846\u67b6\u7b97\u5b50\u7ec4\u5408\u5b9e\u73b0\uff0c\u5219\u65f6\u95f4\u590d\u6742\u5ea6\u4e3a O(M * N\u00b2) \uff0c\u5176\u4e2d M \u548c N \u662f\u4e00\u4e2a\u6570\u91cf\u7ea7\uff0c\u76f8\u5f53\u4e8eO(N\u00b3) \u3002\u4f46\u662f\u901a\u8fc7\u81ea\u5b9a\u4e49\u5185\u6838\uff0c\u53ef\u4ee5\u4f18\u5316\u5230 O(N\u00b2)\u3002<\/p>\n<p>\u5728 GPU \u6838\u5fc3\u4e0a\u8fd0\u884c\u7684\u4ee3\u7801\u7247\u6bb5\u79f0\u4e3a\u5185\u6838\uff08kernel\uff09\u3002\u7f16\u5199\u9ad8\u6027\u80fd\u7684 CUDA \u5185\u6838\u9700\u8981\u4e30\u5bcc\u7684\u7ecf\u9a8c\uff0c\u5e76\u4e14\u5b66\u4e60\u66f2\u7ebf\u9661\u5ced\u3002\u56e0\u4e3a\u6211\u4eec\u4e60\u60ef\u4e8e\u4f20\u7edf CPU \u7f16\u7a0b\u5904\u7406\u4e32\u884c\u7684\u8ba1\u7b97\u4efb\u52a1\uff0c\u901a\u8fc7\u591a\u7ebf\u7a0b\u63d0\u9ad8\u5e76\u53d1\u5ea6\u3002\u800c GPU \u91c7\u7528 SIMT \u67b6\u6784\uff0c\u6709\u5927\u91cf\u8ba1\u7b97\u5355\u5143\uff08CUDA Cores\uff09\u548c\u6570\u4e07\u4e2a\u7ebf\u7a0b\uff0c\u4f46\u662f\u88ab\u5206\u7ec4\u540e\u7684\u7ebf\u7a0b\u540c\u4e00\u65f6\u523b\u53ea\u80fd\u6267\u884c\u76f8\u540c\u7684\u6307\u4ee4\u3002\u8fd9\u4e0e\u4f20\u7edfCPU\u7684\u4e32\u884c\u601d\u7ef4\u3001\u4e0d\u540c\u7ebf\u7a0b\u5904\u7406\u4e0d\u540c\u4efb\u52a1\uff0c\u5b58\u5728\u6839\u672c\u6027\u51b2\u7a81\uff0c\u5bfc\u81f4 GPU \u7f16\u7a0b\u5b66\u4e60\u96be\u5ea6\u5927\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/06d370d4319f59df9d63406408acbbd8027dd0.webp\" data-type=\"block\" \/><\/p>\n<p>\u73b0\u5b9e\u751f\u6d3b\u4e2d\u7684SIMT\u67b6\u6784<\/p>\n<p>\u73b0\u5728\u63a8\u8350\u4f7f\u7528 Triton \u7f16\u7a0b\u8bed\u8a00\u5b8c\u6210 GPU kernel \u7684\u5f00\u53d1\uff0c\u5b83\u63d0\u4f9b\u7c7b\u4f3c Python \u7684\u8bed\u6cd5\uff0c\u65e0\u9700\u6df1\u5165\u7406\u89e3 GPU \u786c\u4ef6\u7ec6\u8282\uff08\u5982\u7ebf\u7a0b\u8c03\u5ea6\u3001\u5171\u4eab\u5185\u5b58\u7ba1\u7406\uff09\uff0c\u800c\u4e14\u548c PyTorch \u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u751f\u6001\u7ed3\u5408\u66f4\u597d\u3002\u63a8\u8350\u8fd9\u4e2a\u00a0Triton-Puzzles-Lite\u00a0\u9879\u76ee\u7528\u4f5c Triton \u7684\u5165\u95e8\u5b66\u4e60\u3002<\/p>\n<h4>3. Python \u7f16\u7a0b<\/h4>\n<p>\u6b63\u5982\u5ba2\u6237\u7aef\u5f00\u53d1\u79bb\u4e0d\u5f00Kotlin\/Objective-C\uff0cAI Infra \u7684\u7f16\u7a0b\u7b2c\u4e00\u516c\u6c11\u5c31\u662f Python\u3002PyTorch \u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u7684\u8bbe\u8ba1\u54f2\u5b66\u5f3a\u8c03\u00a0Python \u4f18\u5148\u00a0\u3002<\/p>\n<p>\u4ee5\u524d\u5927\u90e8\u5206\u6a21\u578b\u8fd8\u53ef\u4ee5\u8f7b\u677e\u5bfc\u51fa ONNX\u3001TorchScript \u7b49\u7528 C++ \u90e8\u7f72\uff0c\u73b0\u5728\u968f\u7740\u5bf9\u6a21\u578b\u7684\u7ec6\u7c92\u5ea6\u4f18\u5316\u548c\u63a7\u5236\u8d8a\u6765\u8d8a\u591a\uff0c\u6bd4\u5982 KV Cache\u3001MoE\/\u6a21\u578b\u5e76\u884c\u3001\u590d\u6742\u7684if\/for\u63a7\u5236\u6d41\u3001\u81ea\u5b9a\u4e49 Triton \u7b97\u5b50\u7b49\uff0c\u6a21\u578b\u8d8a\u6765\u8d8a\u96be\u4ee5\u8131\u79bb Python \u7684\u63a7\u5236\u90e8\u7f72\u3002\u7b14\u8005\u4e5f\u4ece\u201cC++ Boy\u201d\u53d8\u6210\u201cPython Boy\u201d\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/183b6ba1491ee21bc6078645f39c08925fe877.webp\" data-type=\"block\" \/><\/p>\n<p>\u7b14\u8005git\u63d0\u4ea4\u7684\u8bed\u8a00\u7edf\u8ba1\u53d8\u5316<\/p>\n<h3>\u4e09\u3001\u6a21\u578b\u8bad\u7ec3\u7684\u6311\u6218<\/h3>\n<p>\u6211\u4eec\u4e00\u76f4\u8ffd\u6c42\u66f4\u5927\u7684\u6a21\u578b\uff0cDeepSeek-R1 \u6709\u6570\u5343\u4ebf\u53c2\u6570\uff0c\u4f7f\u7528\u4e86\u6570\u5341\u4e07\u4ebf token \u7684\u8bad\u7ec3\u6570\u636e\uff0c\u6d89\u53ca\u7b97\u529b\u3001\u5b58\u50a8\u3001\u901a\u4fe1\u7b49\u591a\u7ef4\u5ea6\u7684\u5de5\u7a0b\u6311\u6218\u3002\u6709\u4e86 PyTorch \u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u53ea\u662f AI \u5e94\u7528\u843d\u5730\u7684\u4e07\u91cc\u957f\u5f81\u7b2c\u4e00\u6b65\u3002\u63a5\u4e0b\u6765\u6211\u4eec\u5c06\u8ba8\u8bba\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u4e4b\u4e0a\u7684\u6a21\u578b\u8bad\u7ec3\u7684\u6311\u6218\u3002<\/p>\n<h4>1. \u5b58\u5f97\u4e0b<\/h4>\n<p>DeepSeek-R1 \u6a21\u578b\u5927\u5c0f=670GB\uff0c\u800c\u4e00\u53f0 GPU \u670d\u52a1\u5668\u67098\u5f20H20\u5361\uff0c\u63d0\u4f9b768GB\u663e\u5b58\uff0c\u8db3\u591f\u5b58\u4e0b\u4e00\u4e2a\u5b8c\u6574\u7684 DeepSeek \u6a21\u578b\u3002\u90a3\u6574\u4e2a\u884c\u4e1a\u4e3a\u4ec0\u4e48\u8fd8\u6295\u5165\u5927\u91cf\u7684\u4eba\u529b\u7269\u529b\uff0c\u9876\u7740\u901a\u4fe1\u5ef6\u65f6\u9020\u6210\u7684\u7b97\u529b\u635f\u8017\uff0c\u4e5f\u8981\u5efa\u8bbe\u5206\u5e03\u5f0f GPU \u96c6\u7fa4\uff1f\u6838\u5fc3\u539f\u56e0\u662f\u5355\u53f0 GPU \u670d\u52a1\u5668\u201c\u5b58\u4e0d\u4e0b\u201d\u3002<\/p>\n<p><strong>(1) \u663e\u5b58\u523a\u5ba2\uff1a\u4e2d\u95f4\u6fc0\u6d3b<\/strong><\/p>\n<p>\u5982\u4e0b\u56fe\u6240\u793a\u7684\u6a21\u578b\uff0cx1\/x2\/x3\/x4 \u8fd9\u4e9b\u4e2d\u95f4\u53d8\u91cf\u5c31\u662f&#8221;\u4e2d\u95f4\u6fc0\u6d3b&#8221;\u3002\u5b83\u4eec\u662f\u795e\u7ecf\u7f51\u7edc\u524d\u5411\u4f20\u64ad\uff08Forward\uff09\u7684\u201c\u5806\u6808\u5e27\uff08Stack Frame\uff09\u201d\u2014\u2014\u8bb0\u5f55\u6bcf\u4e00\u5c42\u5904\u7406\u540e\u7684\u6570\u636e\u5feb\u7167\uff0c\u786e\u4fdd\u53cd\u5411\u4f20\u64ad\uff08Backward\uff09\u53ef\u56de\u6eaf\u68af\u5ea6\uff0c\u6839\u636e\u9884\u6d4b\u8bef\u5dee\u8c03\u6574\u6a21\u578b\u6743\u91cd\uff0c\u6700\u5c0f\u5316\u635f\u5931\u51fd\u6570\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/c9700f752176be84a6d345841ccd9843ca82ba.webp\" data-type=\"block\" \/><\/p>\n<p>\u8fd9\u4e9b\u4e2d\u95f4\u6fc0\u6d3b\u4e3a\u4ec0\u4e48\u4f1a\u6210\u4e3a&#8221;\u663e\u5b58\u523a\u5ba2&#8221;\uff1f\u662f\u56e0\u4e3a\u4e2d\u95f4\u6fc0\u6d3b\u7684\u7a7a\u95f4\u590d\u6742\u5ea6\u662f\u548c\u8f93\u5165\u6570\u636e\u957f\u5ea6\u6b63\u76f8\u5173\u7684\uff0c\u7279\u522b\u7684\uff0c\u5bf9\u4e8e LLM \u6765\u8bf4\u662fO\uff08N\u00b2\uff09\u6b63\u6bd4\u4e8e\u8f93\u5165\u6570\u636e\u957f\u5ea6\u7684\u5e73\u65b9\uff0c\u8fd9\u662f\u4e00\u4e2a\u6307\u6570\u7206\u70b8\u5f0f\u589e\u957f\u7684\u6570\u5b57\u3002\u7c7b\u4f3c\u51fd\u6570\u9012\u5f52\u4e0d\u65ad\u589e\u957f\u7684\u201c\u5806\u6808\u5e27\u201d\u5bfc\u81f4\u7684\u5185\u5b58\u6ea2\u51fa\uff0c\u6211\u4eec\u9047\u5230\u4e86\u00a0AI Infra \u7684 OOM\uff08Out of Memory\uff09\u6311\u6218\u3002<\/p>\n<p>\u501f\u52a9 PyTorch \u7684 profiler \u5de5\u5177\uff0c\u6211\u4eec\u53ef\u4ee5\u76f4\u89c2\u5730\u770b\u5230\u8fd9\u4e2aOOM\u3002\u4e0b\u56fe\u662f\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e0d\u540c\u9636\u6bb5\u7684\u663e\u5b58\u5206\u914d\uff0c\u5305\u62ec\u6a21\u578b\u53c2\u6570\uff08Parameter\uff09\u3001\u4f18\u5316\u5668\u72b6\u6001\uff08Optimizer state\uff09\u3001\u4e2d\u95f4\u6fc0\u6d3b\uff08Activation\uff09\u3001\u68af\u5ea6\uff08Gradient\uff09\u3002\u5728\u524d\u5411\u4f20\u64ad\u7ed3\u675f\u540e\u51fa\u73b0\u4e00\u4e2a\u663e\u5b58\u5360\u7528\uff08\u4e2d\u95f4\u6fc0\u6d3b\uff09\u7684\u5c16\u5cf0\uff0c\u8fdc\u5927\u4e8e\u6a21\u578b\u53c2\u6570\u672c\u8eab\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/47270c87703bafd3d7e333c190978caee07636.webp\" data-type=\"block\" \/><\/p>\n<p><strong>(2) \u6a21\u578b\u5e76\u884c<\/strong><\/p>\n<p>\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u4f7f\u7528\u5206\u7247\uff08Sharding\uff09\u7b56\u7565\u89e3\u51b3\u5355\u673a\u5b58\u4e0d\u4e0b\u7684\u95ee\u9898\u3002\u4e0e\u4e4b\u76f8\u4f3c\uff0cAI Infra \u63d0\u51fa\u201c\u6a21\u578b\u5e76\u884c\u201d\uff0c\u5c31\u662f\u5c06\u5355\u4e2a\u5927\u6a21\u578b\u62c6\u5206\u4e3a\u591a\u4e2a\u5b50\u6a21\u5757\uff0c\u5e76\u5206\u5e03\u5230\u4e0d\u540c GPU \u4e0a\u534f\u540c\u5de5\u4f5c\uff0c\u901a\u8fc7\u901a\u4fe1\u6765\u5171\u4eab\u6570\u636e\u3002\u6709\u4e0d\u540c\u7684\u201c\u62c6\u5206\u6a21\u578b\u201d\u7b56\u7565\uff0c\u4f8b\u5982\u6309\u6a21\u578b\u6a21\u5757\u5212\u5206\uff0c\u6309\u5f20\u91cf\uff08Tensor\uff09\u5212\u5206\u7684\uff0c\u4e5f\u53ef\u4ee5\u5c06\u591a\u79cd\u62c6\u5206\u65b9\u6cd5\u7ed3\u5408\u8d77\u6765\u4e00\u8d77\u4f7f\u7528\u3002PyTorch \u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u548c\u5f00\u6e90\u65b9\u6848 Megatron \u90fd\u80fd\u5e2e\u52a9\u6211\u4eec\u9ad8\u6548\u5730\u5b9e\u73b0\u6a21\u578b\u5e76\u884c\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/5345f0e8238400b87b1850775c27fafa725cda.webp\" data-type=\"block\" \/><\/p>\n<p>\u4e0d\u540c\u7684\u6a21\u578b\u5e76\u884c\u7b56\u7565<\/p>\n<h4>2. \u7b97\u5f97\u5feb<\/h4>\n<p>\u5efa\u8bbe\u5206\u5e03\u5f0f GPU \u96c6\u7fa4\u7684\u539f\u56e0\uff0c\u4e00\u4e2a\u662f\u56e0\u4e3a\u201c\u5355\u673a\u5b58\u4e0d\u4e0b\u201d\uff0c\u53e6\u5916\u4e00\u4e2a\u662f\u63d0\u5347\u8bad\u7ec3\u901f\u5ea6\u3002\u4f46\u7b80\u5355\u7684\u673a\u5668\u5806\u53e0\uff0c\u7b97\u529b\u4e0d\u4e00\u5b9a\u6709\u7ebf\u6027\u7684\u589e\u957f\u3002\u56e0\u4e3a\u5206\u5e03\u5f0f\u8bad\u7ec3\u5e76\u4e0d\u662f\u7b80\u5355\u5730\u628a\u539f\u6765\u4e00\u4e2a GPU \u505a\u7684\u4e8b\u60c5\u5206\u7ed9\u591a\u4e2a GPU \u5404\u81ea\u505a\u3002\u9700\u8981\u534f\u8c03\u591a\u4e2a GPU \u673a\u5668\u8ba1\u7b97\u4efb\u52a1\u5206\u914d\uff0cGPU \u673a\u5668\u4e4b\u95f4\u7684\u6570\u636e\u4f20\u8f93\u4f1a\u5f15\u5165\u7f51\u7edcIO\u548c\u901a\u4fe1\u5f00\u9500\uff0c\u964d\u4f4e\u8bad\u7ec3\u901f\u5ea6\u3002<\/p>\n<p><strong>(1) \u901a\u4fe1\u8ba1\u7b97\u91cd\u53e0<\/strong><\/p>\n<p>\u5982\u4e0b\u56fe\u6240\u793a\u7684\u5e38\u89c4\u8bad\u7ec3\u65f6\u5e8f\u662f\u4e32\u8054\u5f0f\u7684\uff0c\u5b58\u5728\u8bb8\u591a\u7f51\u7edc IO\uff0cGPU \u5229\u7528\u7387\u4f4e\uff0c\u8bad\u7ec3\u901f\u5ea6\u6162\u3002\u6211\u4eec\u5e0c\u671b GPU \u5927\u90e8\u5206\u65f6\u95f4\u90fd\u5728\u8ba1\u7b97\uff0c\u800c\u4e0d\u662f\u82b1\u5728\u6570\u636e\u4f20\u8f93\u6216\u7b49\u5f85\u5176\u4ed6 GPU \u7684\u5de5\u4f5c\u4e0a\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/f66b0da41028f6b4dc98338d221d94a3158dbe.webp\" data-type=\"block\" \/><\/p>\n<p>\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u6211\u4eec\u901a\u8fc7\u591a\u7ebf\u7a0b\u6216\u5f02\u6b65 IO \u907f\u514d\u963b\u585e CPU \u4e3b\u7ebf\u7a0b\uff0c\u4e0e\u4e4b\u76f8\u4f3c\uff0cAI Infra \u63d0\u51fa\u901a\u4fe1\u8ba1\u7b97\u91cd\u53e0\u7684\u65b9\u6cd5\u8bba\u3002GPU \u7f16\u7a0b\u6a21\u578b\u4e2d\u6709\u6d41\uff08stream\uff09\u7684\u6982\u5ff5\uff0c\u4e00\u4e2a\u6d41\u8868\u793a\u4e00\u4e2a GPU \u64cd\u4f5c\u961f\u5217\uff0c\u8be5\u961f\u5217\u4e2d\u7684\u64cd\u4f5c\u5c06\u4ee5\u6dfb\u52a0\u5230\u6d41\u4e2d\u7684\u5148\u540e\u987a\u5e8f\u800c\u4f9d\u6b21\u6267\u884c\u3002\u4e0d\u540c\u6d41\u4e4b\u95f4\u53ef\u4ee5\u5e76\u884c\u6267\u884c\u3002\u90a3\u4e48\u901a\u8fc7\u4ee4\u8ba1\u7b97\u548c\u901a\u4fe1\u64cd\u4f5c\u52a0\u5165\u4e0d\u540c\u7684\u6d41\u4e2d\uff0c\u53ef\u4ee5\u505a\u5230\u4e8c\u8005\u7684\u6267\u884c\u5728\u65f6\u95f4\u4e0a\u91cd\u53e0\u3002\u4f8b\u5982 TorchRec \u7684\u00a0\u8bad\u7ec3\u6d41\u6c34\u7ebf\u00a0\u80fd\u5e2e\u52a9\u6211\u4eec\u5b9e\u73b0\u9ad8\u6548\u7684\u901a\u4fe1\u8ba1\u7b97\u91cd\u53e0\u3002<\/p>\n<h3>\u56db\u3001\u6a21\u578b\u63a8\u7406\u7684\u6311\u6218<\/h3>\n<p>AI \u6a21\u578b\u8bad\u7ec3\u6210\u672c\u5f88\u9ad8\uff0c\u4f18\u79c0\u5982 DeepSeek \u4e5f\u8981\u70e7\u6389500\u4e07\u7f8e\u91d1\uff0c\u4f46\u518d\u8d35\u4e5f\u53ea\u662f\u4e00\u6b21\u6027\u7684\u3002\u800c\u6a21\u578b\u63a8\u7406\u7684\u6210\u672c\u66f4\u9ad8\uff0c\u56e0\u4e3a\u7528\u6237\u8d8a\u591a\uff0cAI \u6a21\u578b\u63a8\u7406\u6b21\u6570\u8d8a\u591a\uff0c\u603b\u6210\u672c\u8d8a\u9ad8\u3002\u6a21\u578b\u63a8\u7406\u9762\u5bf9\u7684\u6311\u6218\u548c\u4f20\u7edf Infra \u975e\u5e38\u76f8\u4f3c\uff0c\u4e3b\u8981\u662f2\u4e2a\u6311\u6218\uff1a\u9ad8\u541e\u5410\uff08\u964d\u672c\uff09\uff0c\u4f4e\u5ef6\u65f6\uff08\u589e\u6548\uff09\u3002<\/p>\n<h4>1. \u964d\u4f4e\u5ef6\u65f6<\/h4>\n<p>\u73b0\u5728\u7684 AI \u6a21\u578b\u8d8a\u6765\u8d8a\u591a\u5730\u76f4\u9762\u7ec8\u7aef\u7528\u6237\uff0c\u9700\u8981\u548c\u7528\u6237\u8fdb\u884c\u5b9e\u65f6\u7684\u4ea4\u4e92\uff0c\u4f8b\u5982\u6587\u672c\u5bf9\u8bdd\u548c\u8bed\u97f3\u5408\u6210\u3002\u6a21\u578b\u63a8\u7406\u8017\u65f6\u8fc7\u9ad8\uff0c\u4f1a\u76f4\u63a5\u9020\u6210\u7528\u6237\u4f53\u9a8c\u53d7\u635f\uff0c\u7528\u6237\u6d41\u5931\u4e0e\u8f6c\u5316\u7387\u4e0b\u964d\u3002<\/p>\n<p>\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u6211\u4eec\u4f7f\u7528\u94fe\u63a5\u590d\u7528\u3001\u7f13\u5b58\u3001\u67d4\u6027\u7b49\u6280\u672f\u964d\u4f4e\u7cfb\u7edf\u54cd\u5e94\u65f6\u95f4\u3002AI Infra \u4e5f\u6709\u76f8\u4f3c\u7684\u505a\u6cd5\u3002<\/p>\n<p><strong>(1) CUDA Graph<\/strong><\/p>\n<p>\u5728 GPU \u7f16\u7a0b\u6a21\u578b\u4e2d\uff0cCPU \u548c GPU \u662f\u5f02\u6784\u7684\uff0cCPU \u901a\u8fc7 API\uff08\u4f8b\u5982 CUDA API\uff09 \u5411 GPU \u63d0\u4ea4\u4efb\u52a1\uff0c\u7136\u540e\u5f02\u6b65\u7b49\u5f85 GPU \u7684\u8ba1\u7b97\u7ed3\u679c\u8fd4\u56de\u3002GPU \u6536\u5230\u4efb\u52a1\u540e\uff0c\u4f1a\u6267\u884c\u5185\u6838\u542f\u52a8\u3001\u5185\u5b58\u62f7\u8d1d\u3001\u8ba1\u7b97\u7b49\u64cd\u4f5c\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\uff0c\u6d89\u53ca\u5230 CPU \u4e0e GPU \u4e4b\u95f4\u7684\u901a\u4fe1\u3001\u9a71\u52a8\u7a0b\u5e8f\u7684\u5904\u7406\u4ee5\u53ca GPU \u4efb\u52a1\u7684\u8c03\u5ea6\u7b49\u73af\u8282\uff0c\u4f1a\u4ea7\u751f\u4e00\u5b9a\u7684\u5ef6\u8fdf\u3002\u6a21\u578b\u63a8\u7406\u9700\u8981\u6267\u884c\u5927\u91cf\u91cd\u590d\u7684 GPU \u64cd\u4f5c\uff0c\u6bcf\u4e2a\u7684 GPU \u64cd\u4f5c\u90fd\u8981\u91cd\u590d\u6267\u884c\u4e0a\u8bc9\u73af\u8282\uff0c\u8fd9\u4e9b\u975e\u6838\u5fc3\u7684 GPU \u5f00\u9500\u4f1a\u6210\u500d\u6570\u5730\u653e\u5927\uff0c\u5f71\u54cd\u6700\u7ec8\u54cd\u5e94\u65f6\u95f4\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/34b661189f368c78c43571736ad8e818ba9f30.webp\" data-type=\"block\" \/><\/p>\n<p>CPU \u548c GPU \u901a\u4fe1<\/p>\n<p>\u5728\u4f20\u7edf\u540e\u53f0\u670d\u52a1\uff0c\u6211\u4eec\u4f7f\u7528 Redis \u7684 Lua \u811a\u672c\u5c01\u88c5\u591a\u4e2a Redis \u64cd\u4f5c\u548c\u8ba1\u7b97\u903b\u8f91\uff0c\u4e00\u6b21\u63d0\u4ea4\uff0c\u51cf\u5c11\u7f51\u7edc\u5f00\u9500\u3002\u4e0e\u4e4b\u76f8\u4f3c\uff0cAI Infra \u5229\u7528 CUDA Graph \u6280\u672f\u5c06\u591a\u4e2a GPU \u64cd\u4f5c\u8f6c\u5316\u4e3a\u4e00\u4e2a\u6709\u5411\u65e0\u73af\u56fe\uff08DAG\uff09\uff0c\u7136\u540e\u4e00\u6b21\u6027\u63d0\u4ea4\u6574\u4e2a DAG \u63d0\u4ea4\u5230 GPU \u6267\u884c\uff0c\u7531GPU\u81ea\u8eab\u6765\u7ba1\u7406\u8fd9\u4e9b\u64cd\u4f5c\u7684\u4f9d\u8d56\u5173\u7cfb\u548c\u6267\u884c\u987a\u5e8f\uff0c\u4ece\u800c\u51cf\u5c11 CPU \u4e0e GPU \u4e4b\u95f4\u7684\u4ea4\u4e92\u5f00\u9500\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/81c4c9c203fcd1dcb441562b37bf798643ea45.webp\" data-type=\"block\" \/><\/p>\n<p>\u591a\u4e2a GPU \u5185\u6838\u542f\u52a8\u8f6c\u5316\u4e3a CUDA Graph<\/p>\n<p><strong>(2) KV Cache\uff1a\u7a7a\u95f4\u6362\u65f6\u95f4<\/strong><\/p>\n<p>LLM \u5927\u6a21\u578b\u63a8\u7406\u5b58\u5728\u5927\u91cf\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\uff0c\u4e14\u9ad8\u5ea6\u4f9d\u8d56\u4e0a\u4e0b\u6587\u4fe1\u606f\u3002\u6bcf\u6b21\u63a8\u7406\u90fd\u9700\u8981\u5c06\u4e4b\u524d\u751f\u6210\u8fc7\u7684\u8bcd\u91cd\u65b0\u8f93\u5165\u6a21\u578b\u8fdb\u884c\u8ba1\u7b97\u3002\u8fd9\u79cd\u8ba1\u7b97\u65b9\u5f0f\u4f7f\u5f97\u590d\u6742\u5ea6\u8fbe\u5230\u4e86 O\uff08N\u00b2\uff09\uff0c\u5176\u4e2d\u5fc5\u7136\u5b58\u5728\u5927\u91cf\u7684\u91cd\u590d\u8ba1\u7b97\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u7ed9\u5b9a\u201c\u5929\u6c14\u201d\uff0c\u6a21\u578b\u4f1a\u9010\u4e2a\u9884\u6d4b\u5269\u4e0b\u7684\u5b57\uff0c\u5047\u8bbe\u63a5\u4e0b\u6765\u9884\u6d4b\u7684\u4e24\u4e2a\u5b57\u4e3a\u201c\u771f\u597d\u201c\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/5139dc5106fdfcbc47b840f7d021b92b8d11cb.webp\" data-type=\"block\" \/><\/p>\n<p>\u5c06\u201d\u771f\u201c\u62fc\u63a5\u5230\u201d\u5929\u6c14\u201c\u7684\u540e\u9762\uff0c\u5373\u65b0\u7684\u8f93\u5165\u4e3a\u201d\u5929\u6c14\u771f\u201c\uff0c\u518d\u9884\u6d4b\u201d\u597d\u201c<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/s2.51cto.com\/oss\/202507\/20\/d54205c12a1012d213220137e4f2c71f1e3387.webp\" data-type=\"block\" \/><\/p>\n<p>\u89c2\u5bdf\u5230\uff0c\u7ecf\u8fc7\u591a\u6b21\u9884\u6d4b\u540e\uff0cX @ W_K\u00a0\u548c\u00a0X @ W_V\u00a0\u7684\u7ed3\u679c\u4e0a\u534a\u90e8\u5206\u90fd\u662f\u76f8\u540c\u7684\u3002\u8fd9\u662f\u7531\u4e8e LLM \u6a21\u578b\u7ed3\u6784\u7684\u7279\u6b8a\u8bbe\u8ba1\u5bfc\u81f4\u7684\u3002\u8fd9\u4e9b\u91cd\u590d\u8ba1\u7b97\u7684\u7ed3\u679c\u53ef\u4ee5\u7f13\u5b58\uff08\u5373 KV Cache\uff09\u4e0b\u6765\uff0c\u7a7a\u95f4\u6362\u65f6\u95f4\uff0c\u51cf\u5c11\u8ba1\u7b97\u91cf\u3002\u51e0\u4e4e\u6240\u6709\u7684 LLM \u63a8\u7406\u6846\u67b6\u90fd\u652f\u6301\u4e86 KV Cache\uff0c\u4f8b\u5982vLLM\u00a0\u3002<\/p>\n<p><strong>(3) \u6d41\u5f0f\u54cd\u5e94<\/strong><\/p>\n<p>\u6709\u65f6\u5019\u6a21\u578b\u63a8\u7406\u5ef6\u65f6\u5b9e\u5728\u907f\u514d\u4e0d\u4e86\uff0c\u53ef\u4ee5\u4ece\u5de5\u7a0b\u4ea4\u4e92\u4e0a\u60f3\u529e\u6cd5\u3002\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u7684 RPC \u901a\u4fe1\u662f\u4e00\u95ee\u4e00\u7b54\u65b9\u5f0f\uff0c\u8fd9\u79cd\u65b9\u5f0f\u4e0d\u592a\u9002\u5408\u8bed\u97f3\u5408\u6210\u6216\u8005\u6587\u672c\u5bf9\u8bdd\u7684\u573a\u666f\u3002\u56e0\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u9700\u8981\u51e0\u79d2-\u51e0\u5341\u79d2\uff0c\u5982\u679c\u7b49\u5f85\u6a21\u578b\u63a8\u7406\u7ed3\u675f\u624d\u5c55\u793a\u7ed3\u679c\uff0c\u7528\u6237\u4f1a\u7b49\u5f85\u8f83\u957f\u7684\u65f6\u95f4\uff0c\u4f53\u9a8c\u5f88\u5dee\u3002<\/p>\n<p>\u6d41\u5f0f\u54cd\u5e94\u5c31\u662f\u5f53\u6a21\u578b\u63a8\u7406\u8ba1\u7b97\u5f97\u5230\u7b2c\u4e00\u4e2atoken\u6216\u8005\u7b2c\u4e00\u4e2a\u97f3\u9891\u5e27\u7684\u65f6\u5019\uff0c\u7acb\u9a6c\u5c55\u793a\u6216\u8005\u64ad\u653e\u7ed9\u7528\u6237\uff0c\u540c\u65f6\u540e\u7eed\u7684\u6a21\u578b\u63a8\u7406\u7ed3\u679c\u5728\u5df2\u7ecf\u5efa\u7acb\u7684 TCP \u6d41\u4e0a\u7ee7\u7eed\u987a\u5e8f\u4f20\u8f93\u3002\u5de5\u7a0b\u4e0a\u4ece\u5173\u6ce8\u6a21\u578b\u63a8\u7406\u7684\u6574\u4f53\u8017\u65f6\uff0c\u6539\u4e3a\u5173\u6ce8\u9996token\u6216\u9996\u4e2a\u97f3\u9891\u5e27\u7684\u8017\u65f6\u3002\u51e0\u4e4e\u6240\u6709\u7684 LLM \u63a8\u7406\u6846\u67b6\u90fd\u652f\u6301\u4e86\u6d41\u5f0f\u54cd\u5e94\u3002<\/p>\n<h4>2. \u63d0\u9ad8\u541e\u5410\u91cf<\/h4>\n<p>\u63d0\u9ad8\u541e\u5410\u91cf\u662f\u7a0b\u5e8f\u5458\u5728\u4f20\u7edf Infra \u9886\u57df\u5b5c\u5b5c\u4e0d\u5026\u7684\u8ffd\u6c42\uff0c\u56e0\u4e3a\u66f4\u9ad8\u7684\u541e\u5410\u91cf\u610f\u5473\u7740\u66f4\u4f4e\u7684\u673a\u5668\u6210\u672c\u3002\u5b9e\u73b0 AI \u5e94\u7528\u7684\u9ad8\u541e\u5410\u672c\u8d28\u4e0a\u5c31\u662f\u63d0\u9ad8\u6602\u8d35\u7684 GPU \u7684\u5229\u7528\u7387\uff0c\u8ba9 GPU \u5355\u4f4d\u65f6\u95f4\u80fd\u5b8c\u6210\u66f4\u591a\u7684\u4efb\u52a1\u3002<\/p>\n<p>\u5c3d\u7ba1\u6a21\u578b\u63a8\u7406\u9700\u8981\u6267\u884c\u4e07\u4ebf\u6b21\u6d6e\u70b9\u8fd0\u7b97\uff0c\u4f46\u00a0GPU \u6709\u5927\u91cf\u7684\u8ba1\u7b97\u5355\u5143\uff08CUDA Cores\uff09\uff0c\u5355\u4e2a\u8bf7\u6c42\u7684\u6a21\u578b\u63a8\u7406\u5f88\u96be\u4ee4 GPU \u5229\u7528\u7387\u8fbe\u5230\u9971\u548c\u3002\u63d0\u9ad8 GPU \u5229\u7528\u7387\u67092\u4e2a\u65b9\u6cd5\uff1a\u4f20\u7edf\u6279\u5904\u7406\u548c\u8fde\u7eed\u6279\u5904\u7406\u3002\u8fd9\u91cc\u7684\u201c\u4f20\u7edf\u6279\u5904\u7406\u201d\u662f\u76f8\u5bf9\u4e8e\u201c\u8fde\u7eed\u6279\u5904\u7406\u201d\u8fd9\u6837\u7684\u65b0\u578b\u6279\u5904\u7406\u65b9\u5f0f\u800c\u8a00\u7684\u3002<\/p>\n<p><strong>(1) \u4f20\u7edf\u6279\u5904\u7406<\/strong><\/p>\n<p>\u5176\u5b9e\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u4e5f\u5927\u91cf\u4f7f\u7528\u4e86\u6279\u5904\u7406\uff0c\u4f8b\u5982 Redis \u7684 MGet \u547d\u4ee4\uff0c\u5355\u6b21\u8bf7\u6c42\u5c31\u5b8c\u6210\u6240\u6709 key \u7684\u83b7\u53d6\uff0c\u5c06 N \u6b21\u7f51\u7edc\u5f80\u8fd4\uff08RTT\uff09\u538b\u7f29\u4e3a1\u6b21\u3002\u4e0e\u4e4b\u76f8\u4f3c\uff0c\u6a21\u578b\u63a8\u7406\u7684\u6279\u5904\u7406\u5c31\u662f\u5c06\u591a\u4e2a\u8f93\u5165\u6837\u672c\u6253\u5305\uff08batch\uff09\uff0c\u5c06\u539f\u672c\u4e32\u884c\u7684 N \u6b21\u8f7b\u91cf\u7684\u63a8\u7406\u8ba1\u7b97\uff0c\u5408\u5e76\u4e3a 1 \u6b21\u91cd\u91cf\u7684\u8ba1\u7b97\uff0c\u5b9e\u73b0\u5355\u4f4d\u65f6\u95f4\u5185\u5904\u7406\u66f4\u591a\u7684\u8bf7\u6c42\uff0c\u63d0\u9ad8\u4e86 GPU \u5229\u7528\u7387\u3002<\/p>\n<p>\u201c\u6253\u5305\u8f93\u5165\u6837\u672c\u201d\u662f\u4e00\u4e2a\u5171\u6027\u9700\u6c42\uff0c\u5927\u90e8\u5206\u63a8\u7406\u6846\u67b6\u90fd\u63d0\u4f9b\u8be5\u529f\u80fd\uff0c\u4f8b\u5982 Triton Inference Server \u7684\u00a0Batcher\u00a0\u3002<\/p>\n<p><strong>(2) \u8fde\u7eed\u6279\u5904\u7406<\/strong><\/p>\n<p>\u4f20\u7edf\u6279\u5904\u7406\u7c7b\u4f3c \u201c\u56fa\u5b9a\u73ed\u6b21\u7684\u516c\u4ea4\u8f66\u201d\uff1a\u4e58\u5ba2\uff08\u8bf7\u6c42\uff09\u5fc5\u987b\u7b49\u5f85\u53d1\u8f66\u65f6\u95f4\uff08\u7ec4\u5efa\u4e00\u4e2abatch\uff09\uff0c\u53d1\u8f66\u540e\u6240\u6709\u4e58\u5ba2\u540c\u6b65\u524d\u8fdb\u3002\u5373\u4f7f\u6709\u4e58\u5ba2\u63d0\u524d\u4e0b\u8f66\uff08\u77ed\u8bf7\u6c42\u5b8c\u6210\uff09\uff0c\u8f66\u8f86\u4ecd\u9700\u7b49\u5f85\u6240\u6709\u4e58\u5ba2\u5230\u8fbe\u7ec8\u70b9\uff08\u957f\u8bf7\u6c42\u5b8c\u6210\uff09\u624d\u80fd\u8fd4\u7a0b\u63a5\u65b0\u4e58\u5ba2\u3002\u4f20\u7edf\u6279\u5904\u7406\u5b58\u5728\u7740\u8d44\u6e90\u6d6a\u8d39\uff1aGPU \u8981\u7b49\u5f85\u957f\u8bf7\u6c42\u5904\u7406\u5b8c\uff0c\u4e0d\u80fd\u5904\u7406\u65b0\u7684\u8bf7\u6c42\u800c\u7a7a\u95f2\u3002<\/p>\n<p>\u8fd9\u4e2a\u95ee\u9898\u5728 LLM \u5e94\u7528\u9886\u57df\u663e\u5f97\u7279\u522b\u7a81\u51fa\uff0c\u56e0\u4e3a\u4e0d\u540c\u7528\u6237\u8bf7\u6c42 Prompt\uff0c\u6a21\u578b\u7684\u56de\u7b54\u7ed3\u679c\u957f\u5ea6\u5dee\u5f02\u5de8\u5927\uff0c\u5982\u679c\u4f7f\u7528\u4f20\u7edf\u6279\u5904\u7406\uff0cGPU \u7a7a\u95f2\u7387\u5f88\u9ad8\u3002\u8fd9\u4e2a\u672c\u8d28\u4e0a\u662f\u4e2a\u4efb\u52a1\u8c03\u5ea6\u95ee\u9898\uff0c\u4f20\u7edf\u540e\u53f0\u670d\u52a1\u6211\u4eec\u4f7f\u7528\u5de5\u4f5c\u7a83\u53d6\u7b97\u6cd5\uff08work stealing\uff09\u89e3\u51b3\u7ebf\u7a0b\u7a7a\u95f2\u95ee\u9898\uff0c\u4e0e\u4e4b\u76f8\u4f3c\uff0cAI Infra \u63d0\u51fa\u201c\u8fde\u7eed\u6279\u5904\u7406\u201d\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n<p>\u8fde\u7eed\u6279\u5904\u7406\u7c7b\u4f3c\u201c\u968f\u65f6\u968f\u5730\u62fc\u8f66\u7684\u987a\u98ce\u8f66\u201d\uff0c\u6bcf\u8f86\u8f66\uff08GPU\uff09\u5728\u884c\u7a0b\u4e2d\u53ef\u968f\u65f6\u4e0a\/\u4e0b\u5ba2\u3002\u65b0\u4e58\u5ba2\uff08\u8bf7\u6c42\uff09\u76f4\u63a5\u52a0\u5165\u5f53\u524d\u8f66\u8f86\u7684\u7a7a\u4f4d\uff08\u7a7a\u95f2\u8ba1\u7b97\u5355\u5143\uff09\uff0c\u5df2\u5b8c\u6210\u7684\u4e58\u5ba2\u7acb\u5373\u4e0b\u8f66\uff08\u91ca\u653e\u8d44\u6e90\uff09\u3002\u51e0\u4e4e\u6240\u6709\u7684 LLM \u63a8\u7406\u6846\u67b6\u90fd\u652f\u6301\u4e86\u8fde\u7eed\u6279\u5904\u7406\u80fd\u529b\uff0c\u4f8b\u5982 vLLM \u7684\u00a0Continuous Batching<\/p>\n<h3>\u4e94\u3001\u7ed3\u8bed<\/h3>\n<p>AI Infra \u9762\u5bf9\u7684\u5de5\u7a0b\u6311\u6218\uff0c\u4f8b\u5982\u8ba1\u7b97\u3001\u5b58\u50a8\u3001\u901a\u4fe1\uff0c\u5927\u90e8\u5206\u662f\u65b0\u65f6\u4ee3\u7684\u8001\u95ee\u9898\uff0c\u6211\u4eec\u5728\u4f20\u7edf Infra \u9886\u57df\u90fd\u80fd\u627e\u5230\u5bf9\u5e94\u7684\u573a\u666f\u548c\u89e3\u51b3\u601d\u8def\u3002\u5dee\u5f02\u53ea\u5728\u4e8e\u6218\u573a\u4ece CPU \u8f6c\u79fb\u5230 GPU\uff0c\u4f20\u7edf\u540e\u53f0\u5de5\u7a0b\u5e08\u79ef\u7d2f\u7684\u65b9\u6cd5\u8bba\uff0c\u4f9d\u7136\u53ef\u4ee5\u65e0\u7f1d\u8854\u63a5\u5230 AI Infra\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_26774\" class=\"pvc_stats total_only  \" data-element-id=\"26774\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u5c06\u5206\u4eab\u4f20\u7edf\u540e\u53f0\u5de5\u7a0b\u5e08\u79ef\u7d2f\u7684\u6280\u672f\u6808\u548c\u65b9\u6cd5\u8bba\uff0c\u5982\u4f55\u5ef6\u7eed\u5e76\u8fc1\u79fb\u5230 AI \u7cfb\u7edf\uff0c\u5e76\u7cfb\u7edf\u6027\u62c6\u89e3 AI Infra \u7684 [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_26774\" class=\"pvc_stats total_only  \" data-element-id=\"26774\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":26776,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,20,80],"tags":[1819],"class_list":["post-26774","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-20","category-80","tag--ai-infra-ai-"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26774","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=26774"}],"version-history":[{"count":2,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26774\/revisions"}],"predecessor-version":[{"id":26781,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26774\/revisions\/26781"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/26776"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=26774"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=26774"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=26774"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}