{"id":40112,"date":"2025-04-13T23:02:49","date_gmt":"2025-04-13T15:02:49","guid":{"rendered":"https:\/\/17aitech.com\/?p=40112"},"modified":"2025-04-13T23:02:49","modified_gmt":"2025-04-13T15:02:49","slug":"%e4%b8%8d%e7%94%a8%e8%8b%b1%e4%bc%9f%e8%be%begpu%ef%bc%81%e5%8d%8e%e4%b8%ba%e7%9b%98%e5%8f%a4ultra%e6%9d%a5%e4%ba%86%ef%bc%9a%e6%98%87%e8%85%be%e5%8e%9f%e7%94%9f%e3%80%81135b%e7%a8%a0%e5%af%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=40112","title":{"rendered":"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-13-4\" target=\"_blank\">\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b<\/a><\/p>\n<p>\u7ec8\u4e8e\uff0c\u534e\u4e3a\u76d8\u53e4\u5927\u6a21\u578b\u7cfb\u5217\u4e0a\u65b0\u4e86\uff0c\u800c\u4e14\u662f\u6607\u817e\u539f\u751f\u7684\u901a\u7528\u5343\u4ebf\u7ea7\u8bed\u8a00\u5927\u6a21\u578b\u3002<\/p>\n<p>\u6211\u4eec\u77e5\u9053\uff0c\u5982\u4eca\u5404\u5927\u79d1\u6280\u516c\u53f8\u7eb7\u7eb7\u53d1\u5e03\u767e\u4ebf\u3001\u5343\u4ebf\u7ea7\u6a21\u578b\u3002\u4f46\u8fd9\u4e9b\u5927\u90e8\u5206\u6a21\u578b\u8bad\u7ec3\u4e3b\u8981\u4f9d\u8d56\u82f1\u4f1f\u8fbe\u7684 GPU\u3002<\/p>\n<p>\u800c\u73b0\u5728\u7684\u60c5\u5f62\u4e0b\uff0c\u56fd\u5185\u7814\u7a76\u56e2\u961f\u5f88\u96be\u83b7\u5f97\u8db3\u591f\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u8fd9\u4e5f\u5236\u7ea6\u4e86\u56fd\u5185\u5927\u6a21\u578b\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png\"><\/a><\/section>\n<p>\u6211\u4eec\u770b\u5230\u534e\u4e3a\u76d8\u53e4\u53d1\u5e03\u7684\u8fd9\u7bc7\u65b0\u7814\u7a76\uff0c\u8bc1\u660e\u4e86\u57fa\u4e8e\u5168\u56fd\u4ea7\u7684\u6607\u817e\u4e5f\u53ef\u4ee5\u5b9e\u73b0\u9886\u5148\u7684\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u7814\u7a76\u4e0e\u5f00\u53d1\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-84d1e02126b79970abf8125d21cb2a73.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-84d1e02126b79970abf8125d21cb2a73.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u6280\u672f\u62a5\u544a\u6807\u9898\uff1aPangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUs<\/p>\n<\/li>\n<li>\n<p>\u6280\u672f\u62a5\u544a\u5730\u5740\uff1ahttps:\/\/github.com\/pangu-tech\/pangu-ultra\/blob\/main\/pangu-ultra-report.pdf<\/p>\n<\/li>\n<\/ul>\n<p>\u7814\u7a76\u79f0\u534e\u4e3a\u76d8\u53e4\u56e2\u961f\u6210\u529f\u5f00\u53d1\u51fa\u57fa\u4e8e\u6607\u817e\u7b97\u529b\u8bad\u7ec3\u7684\u5343\u4ebf\u7ea7\u901a\u7528\u8bed\u8a00\u5927\u6a21\u578b Pangu Ultra\u3002\u5728\u6548\u679c\u4e0a\uff0cPangu Ultra \u5728\u591a\u4e2a\u9886\u57df\u548c\u8bc4\u6d4b\u4e0a\u8d85\u8d8a\u4e4b\u524d Llama 405B \u548c Mistral Large 2 \u7b49\u7a20\u5bc6\u6a21\u578b\uff0c\u5e76\u53ef\u4ee5\u4e0e DeepSeek-R1 \u7b49\u66f4\u5927\u89c4\u6a21\u7684\u7a00\u758f\u6a21\u578b\u4e00\u8f83\u9ad8\u4e0b\u3002<\/p>\n<p>Pangu Ultra \u662f\u4e00\u4e2a\u62e5\u6709 94 \u5c42\u67b6\u6784\u3001\u603b\u53c2\u6570\u91cf\u8fbe 135B \u7684\u8d85\u5927\u7a20\u5bc6\u6a21\u578b\u3002\u9488\u5bf9\u8d85\u6df1\u5343\u4ebf\u7ea7\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u95ee\u9898\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u65b0\u7684\u7a33\u5b9a\u6027\u67b6\u6784\u548c\u521d\u59cb\u5316\u65b9\u6cd5\uff0c\u6210\u529f\u5b9e\u73b0\u4e86\u5728 13.2T \u9ad8\u8d28\u91cf\u6570\u636e\u4e0a\u7684\u5168\u6d41\u7a0b\u65e0 loss \u7a81\u523a\u957f\u7a33\u8bad\u7ec3\u3002\u540c\u65f6\uff0c\u5728\u7cfb\u7edf\u5b9e\u73b0\u5c42\u9762\uff0c\u56e2\u961f\u901a\u8fc7\u4e00\u7cfb\u5217\u7cfb\u7edf\u4f18\u5316\u7b56\u7565\uff0c\u5728 8192 \u5f20\u6607\u817e NPU \u6784\u5efa\u7684\u5927\u89c4\u6a21\u96c6\u7fa4\u4e0a\u5c06\u7b97\u529b\u5229\u7528\u7387\uff08MFU\uff09\u63d0\u5347\u81f3 50%\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\uff0c\u8ba9\u6211\u4eec\u4ece\u6a21\u578b\u67b6\u6784\u3001\u6a21\u578b\u8bad\u7ec3\u7b49\u65b9\u9762\uff0c\u8be6\u7ec6\u4e86\u89e3\u4e0b Pangu Ultra \u7684\u6280\u672f\u7ec6\u8282\u3002<\/p>\n<p><strong>\u6a21\u578b\u67b6\u6784<\/strong><\/p>\n<p><strong>\u57fa\u7840\u67b6\u6784\u4fe1\u606f<\/strong>\uff1aPangu Ultra \u5305\u542b 1350 \u4ebf\u53c2\u6570\u3001\u91c7\u7528\u4e86 94 \u5c42\u7684 Transformer \u7ed3\u6784\u3002\u5176\u4e2d FFN \u91c7\u7528 SwiGLU \u6fc0\u6d3b\u3002\u6ce8\u610f\u529b\u5c42\u91c7\u7528 GQA \u964d\u4f4e KV \u7f13\u5b58\u5360\u7528\u3002<\/p>\n<p>Pangu Ultra \u9488\u5bf9\u5927\u89c4\u6a21\u6781\u6df1\u6a21\u578b\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u95ee\u9898\u63d0\u51fa\u4e86 Depth-scaled sandwich-norm \u548c TinyInit \u521d\u59cb\u5316\u4e24\u9879\u6280\u672f\u3002<\/p>\n<p><strong>Depth-scaled sandwich-norm<\/strong>\uff1a\u4e0e\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\u4fa7\u91cd\u5728\u5bbd\u5ea6\u4e0a\u6269\u5c55\u4e0d\u540c\uff0c\u5927\u89c4\u6a21\u7a20\u5bc6\u6a21\u578b\u901a\u5e38\u91c7\u7528\u66f4\u6df1\u7684\u5c42\u6570\u3002\u7136\u800c\uff0c\u6df1\u5ea6\u589e\u52a0\u4f1a\u52a0\u5267\u8bad\u7ec3\u7a33\u5b9a\u6027\u7684\u6311\u6218\u3002\u8003\u8651\u5230\u9884\u8bad\u7ec3\u7684\u5de8\u5927\u6210\u672c\uff0c\u4fdd\u8bc1\u5927\u6a21\u578b\u7684\u7a33\u5b9a\u8bad\u7ec3\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>Pre-LN \u5728\u57fa\u4e8e Transformer \u7684\u5927\u8bed\u8a00\u6a21\u578b\u67b6\u6784\u4e2d\u88ab\u5e7f\u6cdb\u5e94\u7528\uff0c\u4f46\u91c7\u7528 Pre-LN \u7684\u6a21\u578b\u4e2d\uff0c\u5404\u5b50\u5c42\u8f93\u51fa\u5c3a\u5ea6\u7684\u6ce2\u52a8\u5bb9\u6613\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002\u4e3a\u89e3\u51b3\u6b64\u95ee\u9898\uff0cSandwich-Norm \u5728\u6b8b\u5dee\u8fde\u63a5\u524d\u5bf9\u6bcf\u4e2a\u5b50\u5c42\u8f93\u51fa\u989d\u5916\u65bd\u52a0 layer norm\u3002\u867d\u7136 Sandwich-Norm \u80fd\u4fdd\u6301\u5355\u4e2a\u5b50\u5c42\u8f93\u51fa\u7684\u5c3a\u5ea6\u7a33\u5b9a\u6027\uff0c\u4f46\u8de8\u8d8a\u591a\u5c42\u7684\u6b8b\u5dee\u8fde\u63a5\u4ecd\u4f1a\u5bfc\u81f4\u8f93\u51fa\u8303\u6570\u9010\u6e10\u7d2f\u79ef\uff0c\u8fdb\u800c\u5f15\u53d1\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002<\/p>\n<p>\u4e3a\u6b64\uff0cPangu Ultra \u63d0\u51fa Depth-Scaled Sandwich-Norm\uff08DSSN\uff09\u3002\u8be5\u6280\u672f\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u5b50\u5c42\u8f93\u51fa\u7684 layer norm \u7684 gamma \u53c2\u6570\u6765\u8c03\u63a7\u5404\u5c42\u8f93\u51fa\u5c3a\u5ea6\uff0c\u901a\u8fc7\u628a gamma \u53c2\u6570\u521d\u59cb\u5316\u4e3a\u4e0e\u7f51\u7edc\u6df1\u5ea6\u7684\u5e73\u65b9\u6839\u5012\u6570\u6210\u6bd4\u4f8b\u3002\u4e0b\u56fe\u5c55\u793a\u4e86 Depth-Scaled Sandwich-Norm \u4e0e Pre-LN \u67b6\u6784\u7684\u5dee\u5f02\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\u6240\u63d0\u51fa\u65b9\u6cd5\u76f8\u6bd4\u57fa\u51c6\u65b9\u6cd5\u5728\u7a33\u5b9a\u6027\u548c\u6536\u655b\u6027\u90fd\u5177\u6709\u8f83\u5927\u4f18\u52bf\uff08\u89c1\u6a21\u578b\u7ed3\u679c\u4e0e\u5206\u6790\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-af8e37ae3abd65de9c6f22247be771e0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-af8e37ae3abd65de9c6f22247be771e0.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 Pre-LN \u4e0e DSSN \u67b6\u6784\u5bf9\u6bd4\u3002\u76f8\u6bd4 Sandiwich norm\uff0cDSSN \u5bf9 layer norm \u7cfb\u6570\u505a\u4e86\u7279\u6b8a\u7684 re-scale<\/sup><\/p>\n<p><strong>TinyInit<\/strong>: \u6a21\u578b\u521d\u59cb\u5316\u5bf9\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6027\u80fd\u5177\u6709\u5173\u952e\u5f71\u54cd\u3002\u57fa\u4e8e Transformer \u7684\u5927\u8bed\u8a00\u6a21\u578b\u666e\u904d\u91c7\u7528 small init \u7b56\u7565\uff0c\u5373\u6240\u6709\u6743\u91cd\u521d\u59cb\u5316\u4e3a\u6807\u51c6\u5dee\u4e3a<span data-pm-slice=\"0 0 []\"><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-15dc2cc16154f6b956d3fa9fcfba7636.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-15dc2cc16154f6b956d3fa9fcfba7636.png\"><\/a>\u7684\u6b63\u6001\u5206\u5e03\u3002\u53e6\u5916\u6709\u5de5\u4f5c\u5c06\u6b8b\u5dee\u5c42\u6743\u91cd\u5728\u521d\u59cb\u5316\u65f6\u7f29\u653e<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-13ef79d8ecad40a2d73babb5aa0bb039.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-13ef79d8ecad40a2d73babb5aa0bb039.png\"><\/a>\u500d<\/span><span data-pm-slice=\"0 0 []\">\u3002<\/span><\/p>\n<p>Pangu \u63d0\u51fa\u4e86\u540c\u65f6\u6839\u636e\u6a21\u578b\u6df1\u5ea6\u548c\u5bbd\u5ea6\u7f29\u653e\u7684\u53c2\u6570\u521d\u59cb\u5316\u7b56\u7565 TinyInit\uff0c\u6240\u6709\u6743\u91cd\u521d\u59cb\u5316\u6807\u51c6\u5dee<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b8fa1250ae6ca7faf9330adb15067a80.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b8fa1250ae6ca7faf9330adb15067a80.png\"><\/a>\uff0c\u80fd\u52a0\u901f loss \u6536\u655b\u5e76\u63d0\u5347\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u3002TinyInit \u80fd\u4f7f\u6a21\u578b\u5404\u90e8\u5206\u7684\u53c2\u6570\u5c3a\u5ea6\u66f4\u52a0\u4e00\u81f4\uff0c\u4ece\u800c\u6709\u5229\u4e8e\u4f18\u5316\u548c\u6536\u655b\uff08\u89c1\u6a21\u578b\u7ed3\u679c\u4e0e\u5206\u6790\uff09\u3002<\/p>\n<p><strong>Pangu Tokenizer<\/strong>\uff1a\u5728 Pangu Ultra \u7684\u4e2d\uff0c\u4f5c\u8005\u4f18\u5316\u4e86\u5206\u8bcd\u5668\uff08Tokenizer\uff09\u3002\u4f20\u7edf\u65b9\u6cd5\u5728\u6784\u5efa\u8bcd\u6c47\u8868\u65f6\uff0c\u5e38\u56e0\u6570\u636e\u5206\u5e03\u5bfc\u81f4\u901a\u7528\u6587\u672c\u5360\u6bd4\u8fc7\u9ad8\uff0c\u800c\u4ee3\u7801\u3001\u6570\u5b66\u7b49\u4e13\u4e1a\u9886\u57df\u7684\u8bcd\u5143\u4ee3\u8868\u6027\u4e0d\u8db3\u3002\u4e3a\u89e3\u51b3\u6b64\u95ee\u9898\uff0cPangu Ultra \u91c7\u7528\u4e86\u300c\u9886\u57df\u611f\u77e5\u300d\u7684\u5206\u8bcd\u8bcd\u8868\u7b56\u7565\uff1a<\/p>\n<ul>\n<li>\n<p>\u5bf9\u901a\u7528\u4e2d\u6587\u3001\u901a\u7528\u82f1\u6587\u3001\u4ee3\u7801\u3001\u6570\u5b66\u7b49\u591a\u4e2a\u5173\u952e\u9886\u57df\u7684\u6570\u636e\u72ec\u7acb\u8fdb\u884c\u8bcd\u9891\u5206\u6790\uff0c\u751f\u6210\u9886\u57df\u4e13\u5c5e\u8bcd\u6c47\u8868\u3002<\/p>\n<\/li>\n<li>\n<p>\u968f\u540e\uff0c\u5c06\u8fd9\u4e9b\u8bcd\u6c47\u8868\u5408\u5e76\u4e0e\u53bb\u91cd\uff0c\u5f62\u6210\u4e86\u4e00\u4e2a\u5305\u542b 153376 \u8bcd\u5143\u7684\u5206\u8bcd\u8bcd\u8868\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u79cd\u65b9\u6cd5\u63d0\u5347\u4e86\u8bcd\u6c47\u8868\u5728\u4e0d\u540c\u9886\u57df\u95f4\u7684\u5747\u8861\u6027\uff0c\u786e\u4fdd\u4e86\u6a21\u578b\u5728\u5904\u7406\u6587\u672c\u3001\u4ee3\u7801\u3001\u6570\u5b66\u7b49\u591a\u6837\u5316\u4efb\u52a1\u65f6\uff0c\u90fd\u80fd\u83b7\u5f97\u66f4\u7cbe\u51c6\u3001\u9ad8\u6548\u7684\u7406\u89e3\u57fa\u7840\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u826f\u597d\u7684\u6574\u4f53\u538b\u7f29\u6548\u7387\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-5ec44b09a53090441cb85b25024d30ab.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-5ec44b09a53090441cb85b25024d30ab.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u5404\u9886\u57df\u8bcd\u5728 Pangu \u8bcd\u8868\u4e2d\u7684\u5206\u5e03<\/sup><\/p>\n<p><strong>\u6a21\u578b\u8bad\u7ec3<\/strong><\/p>\n<p><strong>\u9884\u8bad\u7ec3\u7b56\u7565\u4e0e\u6570\u636e<\/strong>\uff1aPangu Ultra \u7684\u9884\u8bad\u7ec3\u8fc7\u7a0b\u7684\u4e09\u4e2a\u9636\u6bb5\uff1a12T tokens \u7684\u901a\u7528\u80fd\u529b\u8bad\u7ec3\u30010.8T tokens \u7684\u63a8\u7406\u80fd\u529b\u589e\u8bad\u4ee5\u53ca 0.4T tokens \u7684\u9000\u706b\u9636\u6bb5\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b691ef8cee8880f363edeb9d7e01b0f0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-b691ef8cee8880f363edeb9d7e01b0f0.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Pangu Ultra \u9884\u8bad\u7ec3\u5404\u4e2a\u9636\u6bb5\u7684\u6570\u636e\u914d\u6bd4<\/sup><\/p>\n<p>Pangu Ultra \u6a21\u578b\u7684\u9884\u8bad\u7ec3\u8bed\u6599\u8986\u76d6\u4e86\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u7684 13.2T tokens\uff0c\u5982\u4e0a\u8868\u6240\u793a\uff0c\u6574\u4e2a\u9884\u8bad\u7ec3\u8fc7\u7a0b\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5\uff1a\u901a\u7528\u9636\u6bb5\uff08General\uff09\u3001\u63a8\u7406\u9636\u6bb5\uff08Reasoning\uff09\u548c\u9000\u706b\u9636\u6bb5\uff08Annealing\uff09\u3002<\/p>\n<p>\u901a\u7528\u9636\u6bb5\uff0812T\uff09\u65e8\u5728\u5e2e\u52a9\u6a21\u578b\u79ef\u7d2f\u77e5\u8bc6\u548c\u8bed\u8a00\u8868\u8fbe\u80fd\u529b\uff0c\u63a8\u7406\u9636\u6bb5\uff080.8T\uff09\u65e8\u5728\u5f3a\u5316\u6a21\u578b\u63a8\u7406\u6280\u80fd\uff0c\u800c\u9000\u706b\u9636\u6bb5\uff080.4T\uff09\u5219\u8fdb\u4e00\u6b65\u6559\u4f1a\u6a21\u578b\u5982\u4f55\u8fd0\u7528\u5df2\u5b66\u5230\u7684\u77e5\u8bc6\u548c\u63a8\u7406\u6280\u80fd\u3002<\/p>\n<p>\u56e2\u961f\u4e3a\u8bad\u7ec3\u8bed\u6599\u6253\u4e86\u8d28\u91cf\u4e0e\u96be\u5ea6\u7684\u6807\u7b7e\u5206\u6570\uff0c\u5e76\u5728\u4e0a\u8ff0\u4e09\u4e2a\u9636\u6bb5\u4e2d\u91c7\u7528\u8bfe\u7a0b\u5f0f\u7684\u6570\u636e\u91c7\u6837\u7b56\u7565\uff0c\u5373\u5b8c\u6210\u4e00\u4e2a\u4ece\u6613\u5230\u96be\u7684\u5b66\u4e60\u8fc7\u7a0b\uff1a<\/p>\n<ul>\n<li>\n<p>\u901a\u7528\u9636\u6bb5\u8986\u76d6\u4e86\u5927\u91cf\u7684\u4e66\u7c4d\u3001\u7f51\u9875\u3001\u767e\u79d1\u3001\u591a\u8bed\u8a00\u4ee5\u53ca\u5404\u884c\u5404\u4e1a\u7684\u8bed\u6599\uff0c\u5b9e\u9645\u8bad\u7ec3\u4e2d\u5206\u6210\u4e86\u4e24\u4e2a\u5b50\u9636\u6bb5\uff0c\u8bad\u7ec3\u91cf\u5206\u522b\u4e3a 7.4T \u548c 4.6T tokens\uff1b<\/p>\n<\/li>\n<li>\n<p>\u63a8\u7406\u9636\u6bb5\u91cd\u70b9\u63d0\u9ad8\u6cdb\u6570\u5b66\u3001\u7406\u79d1\u4ee5\u53ca\u4ee3\u7801\u6570\u636e\u7684\u6bd4\u91cd\uff0c\u5f3a\u63a8\u7406\u6570\u636e\u5360\u6bd4\u8d85\u8fc7 60%\uff1b<\/p>\n<\/li>\n<li>\n<p>\u9000\u706b\u9636\u6bb5\u5c06\u6307\u4ee4\u7c7b\u6570\u636e\u5360\u6bd4\u63d0\u9ad8\u5230 20%\uff0c\u65e8\u5728\u6fc0\u53d1\u6a21\u578b\u66f4\u597d\u7684\u5e94\u7528\u77e5\u8bc6\u548c\u63a8\u7406\u6280\u80fd\u3002\u56e2\u961f\u8bbe\u8ba1\u4e86\u5927\u91cf\u7684\u6307\u4ee4\u7c7b\u95ee\u7b54\u5bf9\uff0c\u5e76\u4e14\u5305\u542b\u957f\u3001\u77ed\u601d\u7ef4\u94fe\uff0c\u8fd9\u4e9b\u63a8\u7406\u8def\u5f84\u7ecf\u8fc7\u7cbe\u5fc3\u4f18\u5316\uff0c\u4ee5\u786e\u4fdd\u5185\u5bb9\u6e05\u6670\u4e14\u903b\u8f91\u8fde\u8d2f\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u6570\u636e\u8d28\u91cf\u8bc4\u4f30\u662f\u63d0\u5347\u6a21\u578b\u8bad\u7ec3\u6548\u679c\u7684\u6838\u5fc3\u73af\u8282\uff0cPangu Ultra \u7ed3\u5408\u89c4\u5219\u7b5b\u9009\u4e0e AI \u8bc4\u4f30\u4f18\u5316\u6570\u636e\u8d28\u91cf\u3002\u56e2\u961f\u8bbe\u8ba1\u4eba\u5de5\u6807\u6ce8\u6570\u636e\uff0c\u5fae\u8c03\u76d8\u53e4\u7cfb\u5217\u6a21\u578b\u4f5c\u4e3a\u8d28\u91cf\u8bc4\u4f30\u5668\uff0c\u4ece\u6570\u636e\u5e72\u51c0\u5ea6\u3001\u6d41\u7545\u6027\u3001\u6559\u80b2\u4ef7\u503c\u548c\u4fe1\u606f\u5bc6\u5ea6\u56db\u4e2a\u7ef4\u5ea6\uff0c\u5bf9\u8d85\u8fc7 10T \u8bed\u6599\u8fdb\u884c\u6253\u5206\uff0c\u5e76\u4e14\u8d4b\u4e88\u9ad8\u8d28\u91cf\u8bed\u6599\u66f4\u9ad8\u7684\u91c7\u6837\u6982\u7387\u3002<\/p>\n<p>\u56e2\u961f\u57fa\u4e8e Pangu 26 \u4ebf\u53c2\u6570\u7684\u4ee3\u7406\u6a21\u578b\u8fdb\u884c\u4e86\u5927\u91cf\u7684\u6d88\u878d\u5b9e\u9a8c\uff0c\u7ed3\u679c\u8868\u660e\u57fa\u4e8e\u4f4e\u8d28\u91cf\u6570\u636e\u8bad\u7ec3\u7684\u6a21\u578b\u9700\u8981\u591a 1.6 \u500d\u8bad\u7ec3\u91cf\u624d\u80fd\u8fbe\u6210\u57fa\u4e8e\u9ad8\u8d28\u91cf\u6570\u636e\u7684\u540c\u7b49\u6548\u679c\uff0c\u8fdb\u4e00\u6b65\u5370\u8bc1\u4e86\u6570\u636e\u8d28\u91cf\u5bf9\u4e8e\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u7684\u5173\u952e\u4ef7\u503c\u3002<\/p>\n<p><strong>\u957f\u5e8f\u5217\u6269\u5c55<\/strong>\uff1aPangu Ultra \u6a21\u578b\u901a\u8fc7\u4e24\u9636\u6bb5\u957f\u5ea6\u6269\u5c55\u8bad\u7ec3\u5c06\u53ef\u652f\u6301\u7684\u8f93\u5165 token \u6570\u589e\u52a0\u5230 128K\uff0c\u53ef\u4ee5\u8f93\u5165\u7ea6 10 \u4e07\u82f1\u8bed\u5355\u8bcd\u6216\u8005 17 \u4e07\u4e2d\u6587\u6c49\u5b57\u3002\u9488\u5bf9\u957f\u5e8f\u5217\u8bad\u7ec3\u4e2d RoPE \u57fa\u7840\u9891\u7387\u8fd9\u4e00\u91cd\u8981\u53c2\u6570\uff0c\u5728\u8bad\u7ec3\u524d\u9996\u5148\u5728\u76ee\u6807\u6269\u5c55\u957f\u5ea6\u7684\u7279\u5b9a\u9a8c\u8bc1\u96c6\u4e0a\u5bf9\u4e0d\u540c\u7684\u53c2\u6570\u8fdb\u884c\u8bc4\u6d4b\u6765\u641c\u7d22\u51fa\u6700\u4f18\u7684\u53c2\u6570\uff0c\u4fdd\u8bc1\u4e86\u957f\u5ea6\u6269\u5c55\u7684\u6548\u679c\u3002<\/p>\n<p><strong>\u540e\u8bad\u7ec3\u7b56\u7565\u4e0e\u6570\u636e<\/strong>\uff1a\u5728\u6a21\u578b\u540e\u8bad\u7ec3\u9636\u6bb5\uff0c\u901a\u8fc7\u4e24\u9636\u6bb5\u4f18\u5316\u5b9e\u73b0\u80fd\u529b\u8dc3\u5347\uff1a\u9996\u5148\u91c7\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u5efa\u7acb\u57fa\u7840\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\uff0c\u968f\u540e\u5f15\u5165\u57fa\u4e8e\u7ed3\u679c\u7684\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6846\u67b6\uff0c\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u590d\u6742\u63a8\u7406\u3001\u4ef7\u503c\u5bf9\u9f50\u548c\u6307\u4ee4\u6267\u884c\u80fd\u529b\u3002\u4e3a\u5145\u5206\u53d1\u6325\u6607\u817e\u7b97\u529b\u4f18\u52bf\uff0c\u7814\u7a76\u56e2\u961f\u7279\u522b\u8bbe\u8ba1\u4e86\u5177\u6709\u5ef6\u8fdf\u5bb9\u5fcd\u7279\u6027\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\uff0c\u914d\u5408\u878d\u5408\u786e\u5b9a\u6027\u4fe1\u53f7\u4e0e\u6a21\u578b\u8bc4\u4f30\u7684\u6df7\u5408\u5956\u52b1\u7cfb\u7edf\uff0c\u5728\u6570\u5b66\u63a8\u5bfc\u3001\u4ee3\u7801\u751f\u6210\u548c\u901a\u7528\u95ee\u9898\u89e3\u51b3\u4e09\u5927\u9886\u57df\u6784\u5efa\u7cbe\u51c6\u53cd\u9988\u673a\u5236\uff0c\u786e\u4fdd\u5927\u89c4\u6a21\u7b56\u7565\u4f18\u5316\u7684\u6548\u7387\u4e0e\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u6a21\u578b\u5728 AIME 2024\u3001MATH-500\u3001GPQA Diamond \u548c LiveCodeBench \u7b49\u6838\u5fc3\u63a8\u7406\u57fa\u51c6\u4e0a\u53d6\u5f97\u4e86\u7406\u60f3\u7684\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u540e\u8bad\u7ec3\u67b6\u6784\u7684\u6709\u6548\u6027\uff0c\u8fd9\u65e2\u5f97\u76ca\u4e8e\u9884\u8bad\u7ec3\u9636\u6bb5 0.8T \u89c4\u6a21\u7684\u4e13\u9879\u63a8\u7406\u6570\u636e\u79ef\u7d2f\uff0c\u4e5f\u6e90\u4e8e\u5f3a\u5316\u5b66\u4e60\u9636\u6bb5\u5bf9\u6a21\u578b\u6f5c\u529b\u7684\u6df1\u5ea6\u6fc0\u53d1\u3002\u540c\u65f6\u6a21\u578b\u4ecd\u4fdd\u6301\u5f3a\u5927\u7684\u901a\u7528\u8bed\u8a00\u7406\u89e3\u80fd\u529b\uff08MMLU-pro \u548c ArenaHard\uff09\uff0c\u5f70\u663e\u4e86\u6280\u672f\u8def\u5f84\u7684\u5747\u8861\u6027\uff0c\u7ed3\u679c\u89c1\u6a21\u578b\u7ed3\u679c\u4e0e\u5206\u6790\u90e8\u5206\u3002<\/p>\n<p><strong>\u7cfb\u7edf\u4f18\u5316<\/strong><\/p>\n<p>Pangu Ultra 135B \u7684\u8bad\u7ec3\u73af\u5883\u662f\u4e00\u4e2a\u914d\u5907\u4e86 8192 \u4e2a\u6607\u817e NPU \u7684\u5927\u89c4\u6a21\u8ba1\u7b97\u96c6\u7fa4\u3002\u56e2\u961f\u901a\u8fc7\u6df7\u5408\u5e76\u884c\u7b56\u7565\u3001\u7ec6\u7c92\u5ea6\u8d1f\u8f7d\u5747\u8861\u8c03\u4f18\u3001\u9ad8\u6548\u878d\u5408\u7b97\u5b50\u3001\u5b50\u5e8f\u5217\u5207\u5206\u4ee5\u53ca\u6570\u636e\u7f13\u5b58\u5171\u4eab\u7b49\u6280\u672f\u624b\u6bb5\uff0c\u5728 8192 \u5361\u89c4\u6a21\u7684\u96c6\u7fa4\u4e2d\u5b9e\u73b0\u4e86\u8d85\u8fc7 50% \u7684 MFU\uff08Model FLOPs Utilization\uff09\u3002<\/p>\n<p><strong>\u5e76\u884c\u7b56\u7565<\/strong>\uff1a\u4e3a\u4e86\u6269\u5c55 Pangu Ultra \u7684\u8bad\u7ec3\u89c4\u6a21\u5e76\u63d0\u5347\u96c6\u7fa4\u7ebf\u6027\u5ea6\uff0c\u56e2\u961f\u91c7\u7528\u4e86\u6df7\u5408\u5e76\u884c\u7b56\u7565\u3002\u5728 8192 \u5361\u89c4\u6a21\u7684\u96c6\u7fa4\u4e2d\uff0c\u4f7f\u7528\u4e86 128 \u8def\u6570\u636e\u5e76\u884c\uff08Data Parallelism\uff09\u30018 \u8def\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallelism\uff09\u548c 8 \u8def\u6d41\u6c34\u7ebf\u5e76\u884c\uff08Pipeline Parallelism\uff09\uff0c\u540c\u65f6\u7ed3\u5408\u4e86 ZeRO \u548c\u5e8f\u5217\u5e76\u884c\uff08Sequence Parallelism\uff09\uff0c\u4ee5\u964d\u4f4e\u6a21\u578b\u53c2\u6570\u3001\u4f18\u5316\u5668\u72b6\u6001\u548c\u6fc0\u6d3b\u503c\u7684\u663e\u5b58\u5360\u7528\u3002\u7531\u4e8e batch-size \u7684\u9650\u5236\uff0c\u5927\u89c4\u6a21\u96c6\u7fa4\u8bad\u7ec3\u4e2d\u6bcf\u4e2a\u6570\u636e\u5e76\u884c\uff08DP\uff09\u7ec4\u7684\u6279\u6b21\u8f83\u5c0f\uff0c\u5bfc\u81f4\u8f83\u9ad8\u7684\u6d41\u6c34\u7ebf\u7a7a\u6ce1\u7387\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u56e2\u961f\u5f15\u5165\u4e86 6 \u8def\u865a\u62df\u6d41\u6c34\u7ebf\uff08Virtual Pipeline\uff09\u8c03\u5ea6\u7b97\u6cd5\uff0c\u5c06\u8bad\u7ec3\u7a7a\u6ce1\u7387\u4ece 30.45% \u964d\u4f4e\u81f3 6.8%\u3002\u901a\u8fc7\u4e00\u7cfb\u5217\u7cbe\u7ec6\u7684\u8d1f\u8f7d\u5747\u8861\u4f18\u5316\uff0c\u5728 BF16 \u8bad\u7ec3\u7cbe\u5ea6\u4e0b\u5b9e\u73b0\u4e86 43% \u7684 MFU\u3002<\/p>\n<p><strong>\u7cfb\u7edf\u4f18\u5316<\/strong>\uff1a\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\u5927\u89c4\u6a21\u96c6\u7fa4\u7684\u8bad\u7ec3\u6548\u7387\uff0c\u56e2\u961f\u4ece\u591a\u4e2a\u65b9\u9762\u8fdb\u884c\u4e86\u7cfb\u7edf\u4f18\u5316\uff0c\u5c06 8192 \u5361\u8bad\u7ec3\u7684 MFU \u4ece 43% \u63d0\u5347\u81f3 52%\u3002\u5173\u952e\u4f18\u5316\u6280\u672f\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p><strong>MC2\uff08Merged Compute and Communication\uff09\u901a\u7b97\u878d\u5408<\/strong><\/p>\n<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u5c06\u8bad\u7ec3\u4e2d\u7684\u77e9\u9635\u4e58\u6cd5\uff08MatMul\uff09\u8ba1\u7b97\u4e0e\u5f20\u91cf\u5e76\u884c\uff08TP\uff09\u5207\u5206\u5f15\u5165\u7684\u901a\u4fe1\u64cd\u4f5c\u7ec6\u7c92\u5ea6\u62c6\u5206\uff0c\u5e76\u5bf9\u8ba1\u7b97\u548c\u901a\u4fe1\u64cd\u4f5c\u8fdb\u884c\u6df1\u5ea6\u6d41\u6c34\u7ebf\u7f16\u6392\uff0c\u5b9e\u73b0\u4e86\u901a\u4fe1\u4e0e\u77e9\u9635\u4e58\u6cd5\u7684\u9ad8\u6548\u91cd\u53e0\uff0c\u663e\u8457\u63d0\u5347\u4e86\u8d44\u6e90\u5229\u7528\u7387\u548c\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n<ul>\n<li>\n<p><strong>NPU Fusion Attention\uff08NFA\uff09<\/strong><\/p>\n<\/li>\n<\/ul>\n<p>\u9488\u5bf9\u6607\u817e NPU \u4f18\u5316\u7684\u81ea\u6ce8\u610f\u529b\uff08Self-Attention\uff09\u878d\u5408\u7b97\u5b50\uff0c\u652f\u6301 Attention Mask \u538b\u7f29\uff0c\u907f\u514d\u4e86\u663e\u5f0f\u6784\u9020 Attention Mask \u5e26\u6765\u7684\u8ba1\u7b97\u548c\u663e\u5b58\u5f00\u9500\u3002\u5728 Pangu Ultra \u8bad\u7ec3\u4e2d\uff0c\u6839\u636e\u6bcf\u4e2a\u6837\u672c\u7684\u7ed3\u675f\u6807\u8bb0\uff08EOD\uff09\u8ba1\u7b97\u51fa\u5b9e\u9645\u5e8f\u5217\u957f\u5ea6\uff08actual_seq_len\uff09\uff0c\u5e76\u5c06\u5176\u4f20\u5165 NFA\u3002NFA \u5185\u90e8\u4f7f\u7528\u4e00\u4e2a 2048\u00d72048 \u7684\u4e0b\u4e09\u89d2\u77e9\u9635\u4f5c\u4e3a\u7d20\u6750\u5e93\uff0c\u6839\u636e actual_seq_len \u52a8\u6001\u6784\u9020\u8bad\u7ec3\u65f6\u7684 Attention Mask\uff0c\u4ece\u800c\u5728\u91cd\u7f6e Attention Mask \u7684\u573a\u666f\u4e0b\u5b9e\u73b0\u9ad8\u6548\u7684\u81ea\u6ce8\u610f\u529b\u8ba1\u7b97\u3002<\/p>\n<ul>\n<li>\n<p><strong>\u5176\u4ed6\u878d\u5408\u7b97\u5b50<\/strong><\/p>\n<\/li>\n<\/ul>\n<p>\u9664\u4e86 MC2 \u548c NFA\uff0c\u8bad\u7ec3\u8fd8\u91c7\u7528\u4e86 RMSNorm\u3001SwiGLU\u3001RoPE \u878d\u5408\u7b97\u5b50\u3001\u68af\u5ea6\u7d2f\u52a0\u878d\u5408\u4ee5\u53ca PP send\/recv \u878d\u5408\u7b49\u6280\u672f\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u4e86\u7cfb\u7edf\u6027\u80fd\u3002<\/p>\n<ul>\n<li>\n<p><strong>\u5b50\u5e8f\u5217\u5207\u5206<\/strong><\/p>\n<\/li>\n<\/ul>\n<p>\u4e0a\u4e0b\u6587\u5e76\u884c\uff08Context Parallelism\uff0cCP\uff09\u662f\u957f\u5e8f\u5217\u8bad\u7ec3\u4e2d\u5e38\u7528\u7684\u4f18\u5316\u65b9\u6cd5\u3002\u4e3a\u4e86\u5b9e\u73b0 CP \u5207\u5206\u4e0b\u7684\u8d1f\u8f7d\u5747\u8861\uff0cMegatron-LM \u5c06\u5e8f\u5217\u5207\u5206\u4e3a 2\u00d7CP \u4e2a\u5b50\u5e8f\u5217\uff0c\u6bcf\u4e2a\u8bbe\u5907\u8d1f\u8d23\u8ba1\u7b97\u4e0a\u4e0b\u4e24\u4e2a chunk \u7684\u6570\u636e\uff08\u89c1\u56fe 1.(b)\uff09\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u5207\u5206\u65b9\u5f0f\u5728\u91cd\u7f6e Attention Mask \u7684\u573a\u666f\u4e0b\u4ecd\u4f1a\u5bfc\u81f4\u8d1f\u8f7d\u4e0d\u5747\uff08\u89c1\u56fe 1.(c)\uff09\u3002Pangu Ultra \u91c7\u7528\u4e86\u6539\u8fdb\u7684\u5b50\u5e8f\u5217\u5207\u5206\u5e76\u884c\u65b9\u5f0f\uff0c\u9488\u5bf9\u6bcf\u4e2a\u6837\u672c\u4e2d\u7684\u5b50\u5e8f\u5217\u91c7\u7528\u8d1f\u8f7d\u5747\u8861\u7684\u5207\u5206\u7b56\u7565\uff0c\u6bcf\u4e2a\u8282\u70b9\u8d1f\u8d23\u8ba1\u7b97\u5b50\u5e8f\u5217\u4e2d\u7684\u4e24\u4e2a chunks\uff08\u89c1\u56fe 1.(d)\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-bb97133a56c6c3042ef6fcda7e0e45a2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-bb97133a56c6c3042ef6fcda7e0e45a2.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 1. \u5b50\u5e8f\u5217\u5207\u5206\u7684\u5e8f\u5217\u5e76\u884c\u65b9\u5f0f<\/sup><\/p>\n<ul>\n<li>\n<p><strong>\u663e\u5b58\u4f18\u5316<\/strong><\/p>\n<\/li>\n<\/ul>\n<p>\u5141\u8bb8\u540c\u4e00\u8ba1\u7b97\u8bbe\u5907\u4e0a\u7684\u4e0d\u540c vpp stage \u4e4b\u95f4\u5171\u4eab attention mask\/actual_seq_len\uff0cRoPE sin\/cos\uff0c position embedding \u7b49\u6570\u636e\u3002\u907f\u514d\u91cd\u590d\u7684\u8ba1\u7b97\u548c\u663e\u5b58\u5f00\u9500\u3002<\/p>\n<p><strong>\u6a21\u578b\u7ed3\u679c\u4e0e\u5206\u6790<\/strong><\/p>\n<p>Pangu Ultra \u5b9e\u73b0\u4e86\u6607\u817e\u8fd1\u4e07\u5361\u5927\u96c6\u7fa4\u4e0a\u7ea6 13T \u6570\u636e\u7684\u957f\u7a33\u8bad\u7ec3\uff0cDSSN \u548c TinyInit \u4fdd\u969c\u8bad\u7ec3\u5168\u7a0b\u6ca1\u6709\u51fa\u73b0\u4efb\u4f55 loss \u7a81\u523a\uff0c\u5982\u4e0b\u56fe\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-0175ca7fde88a4d9c02a6d3911fb5662.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-0175ca7fde88a4d9c02a6d3911fb5662.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Pangu Ultra \u9884\u8bad\u7ec3 Loss\uff0c\u5168\u6d41\u7a0b\u65e0 loss \u7a81\u523a<\/sup><\/p>\n<p>Pangu Ultra \u7684\u9884\u8bad\u7ec3\u57fa\u5ea7\u6a21\u578b\u6d4b\u8bc4\u7ed3\u679c\u5982\u4e0b\u3002\u5bf9\u6bd4\u7a20\u5bc6\u67b6\u6784\u7684\u4ee3\u8868 Qwen2.5-72B \u548c Llama 405B \u4ee5\u53ca MoE \u67b6\u6784\u7684\u4ee3\u8868 DeepSeek V3\u3002Pangu Ultra \u5728\u5927\u591a\u6570 benchmark \u4e0a\u53d6\u5f97\u4e86\u6700\u597d\u7684\u6548\u679c\uff0c\u548c\u540c\u4e3a\u7a20\u5bc6\u67b6\u6784\u7684\u6a21\u578b\u5bf9\u6bd4\u4f18\u52bf\u66f4\u52a0\u660e\u663e\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-ddd00b45d14e5d5b93d84288281c1918.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-ddd00b45d14e5d5b93d84288281c1918.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0Pangu Ultra Base \u6d4b\u8bc4\u7ed3\u679c\uff0c\u7c97\u4f53\u8868\u660e\u6700\u597d\u7ed3\u679c\uff0c\u4e0b\u5212\u7ebf\u8868\u660e\u8be5\u6a21\u578b\u5728 dense \u4e2d\u6700\u597d<\/sup><\/p>\n<p>\u7ecf\u8fc7\u540e\u8bad\u7ec3\u4e4b\u540e\uff0cPangu Ultra \u5728\u4e3b\u8981\u7684 Reasoning benchmark \u4e0a\u7684\u8868\u73b0\u5982\u4e0b\u6240\u793a\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-399609235d8a8138e81df78dcbed4d0b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-399609235d8a8138e81df78dcbed4d0b.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 Pangu Ultra \u5728 Reasoning Benchmarks \u4e0a\u7684\u8868\u73b0<\/sup><\/p>\n<p>Pangu Ultra \u5728 AIME24\uff0cMATH-500\uff0cGPQA\uff0cMMLU-Pro \u7b49\u6307\u6807\u4e0a\u8d85\u8d8a DeepSeek R1\u3002\u5173\u4e8e\u76d8\u53e4\u540e\u8bad\u7ec3\u4f7f\u7528\u7684\u76f8\u5173\u6280\u672f\u5c06\u5728\u4e4b\u540e\u7684\u62a5\u544a\u4e2d\u53d1\u5e03\u3002<\/p>\n<p>\u9488\u5bf9\u8bad\u7ec3\u7a33\u5b9a\u6027\uff0c\u56e2\u961f\u53d1\u73b0 DSSN \u548c\u5e38\u89c1\u7684 Pre-Norm\uff08Pre-LN\uff09\u67b6\u6784\u5bf9\u6bd4\u80fd\u591f\u5b8c\u5168\u675c\u7edd\u8bad\u7ec3\u4e2d\u7684 loss \u7a81\u523a\u73b0\u8c61\u3002\u5728 gradient norm \u4e0a\uff0c\u4f7f\u7528 DSSN \u7684\u6a21\u578b\u4e5f\u66f4\u52a0\u5e73\u7a33\uff0c\u7a81\u523a\u66f4\u5c11\u3002\u7ecf\u8fc7\u6d4b\u8bc4\uff0cDSSN \u67b6\u6784\u7684\u6a21\u578b\u6548\u679c\u4e5f\u8d85\u51fa Pre-LN \u67b6\u6784\uff0c\u8bf4\u660e\u907f\u514d\u8bad\u7ec3\u7a81\u523a\u7684\u91cd\u8981\u6027\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-319c7cd669104e7e8d6f89d8992a8f74.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-319c7cd669104e7e8d6f89d8992a8f74.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0DSSN \u4e0e Pre-LN \u7684\u8bad\u7ec3\u5bf9\u6bd4\u00a0<\/sup><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fb21ea0c88aeea036fac5994b8041dfe.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fb21ea0c88aeea036fac5994b8041dfe.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 DSSN \u67b6\u6784\u548c Pre-LN \u67b6\u6784\u7684\u6d4b\u8bc4\u6548\u679c\u5bf9\u6bd4<\/sup><\/p>\n<p>\u4f7f\u7528 Sandwich-Norm \u67b6\u6784\u65f6\uff0cRMSNorm \u7684 affine \u53c2\u6570 gamma \u521d\u59cb\u5316\u975e\u5e38\u91cd\u8981\uff0c\u8be5\u7814\u7a76\u63d0\u51fa\u7684 DSSN \u65b9\u6848\u4e0e\u666e\u901a Sandwich-Norm \u67b6\u6784\u5bf9\u6bd4\u8bad\u7ec3 loss \u4e5f\u66f4\u52a0\u5e73\u7a33\uff0c\u4e14\u6536\u655b\u66f4\u5feb\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-ddad49fb5b06275cd31e92caad51686a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-ddad49fb5b06275cd31e92caad51686a.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 DSSN \u5bf9\u6bd4\u666e\u901a Sandwich-Norm<\/sup><\/p>\n<p>\u5173\u4e8e TinyInit\uff0c\u56e2\u961f\u5728 135B \u7684\u6a21\u578b\u89c4\u6a21\u4e0a\u8bad\u7ec3\u4e86\u7ea6 100B tokens\uff0c\u548c\u7ecf\u5178\u57fa\u7ebf\u521d\u59cb\u5316\u65b9\u6848\u76f8\u6bd4\u53d6\u5f97\u4e86\u8f83\u4e3a\u660e\u663e\u7684\u4f18\u52bf\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7bd1291832edf0f8aba840d4e3f52ea7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7bd1291832edf0f8aba840d4e3f52ea7.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0TinyInit \u5bf9\u6bd4\u666e\u901a\u521d\u59cb\u5316\u7684\u6a21\u578b\u6d4b\u8bc4\u6548\u679c<\/sup><\/p>\n<p>\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u9605\u8bfb\u8bba\u6587\u539f\u6587\uff0c\u4e86\u89e3\u66f4\u591a\u7814\u7a76\u5185\u5bb9\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-13-4\" target=\"_blank\">\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68,81,82],"class_list":["post-40112","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-81","tag-82"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=40112\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=40112\",\"url\":\"https:\/\/17aitech.com\/?p=40112\",\"name\":\"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=40112#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=40112#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png\",\"datePublished\":\"2025-04-13T15:02:49+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=40112#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=40112\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=40112#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=40112#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=40112","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=40112","url":"https:\/\/17aitech.com\/?p=40112","name":"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=40112#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=40112#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png","datePublished":"2025-04-13T15:02:49+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=40112#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=40112"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=40112#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-a17449e84bcf6380d5273a19edb2ba55.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=40112#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u4e0d\u7528\u82f1\u4f1f\u8fbeGPU\uff01\u534e\u4e3a\u76d8\u53e4Ultra\u6765\u4e86\uff1a\u6607\u817e\u539f\u751f\u3001135B\u7a20\u5bc6\u901a\u7528\u5927\u6a21\u578b"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40112","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=40112"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40112\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=40112"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=40112"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=40112"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}