{"id":3781,"date":"2025-05-09T16:53:00","date_gmt":"2025-05-09T08:53:00","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=3781"},"modified":"2025-05-09T16:54:32","modified_gmt":"2025-05-09T08:54:32","slug":"bye%ef%bc%8c%e8%8b%b1%e4%bc%9f%e8%be%be%ef%bc%81%e5%8d%8e%e4%b8%banpu%ef%bc%8c%e8%b7%91%e5%87%ba%e4%ba%86%e5%87%86%e4%b8%87%e4%ba%bf%e5%8f%82%e6%95%b0%e5%a4%a7%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2025\/05\/09\/bye%ef%bc%8c%e8%8b%b1%e4%bc%9f%e8%be%be%ef%bc%81%e5%8d%8e%e4%b8%banpu%ef%bc%8c%e8%b7%91%e5%87%ba%e4%ba%86%e5%87%86%e4%b8%87%e4%ba%bf%e5%8f%82%e6%95%b0%e5%a4%a7%e6%a8%a1%e5%9e%8b\/","title":{"rendered":"Bye\uff0c\u82f1\u4f1f\u8fbe\uff01\u534e\u4e3aNPU\uff0c\u8dd1\u51fa\u4e86\u51c6\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b"},"content":{"rendered":"\t\t<div data-elementor-type=\"wp-post\" data-elementor-id=\"3781\" class=\"elementor elementor-3781\">\n\t\t\t\t<div class=\"elementor-element elementor-element-a5e473b e-flex e-con-boxed e-con e-parent\" data-id=\"a5e473b\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-82e856b elementor-widget elementor-widget-text-editor\" data-id=\"82e856b\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<p>\u73b0\u5728\uff0c\u8dd1\u51c6<strong>\u4e07\u4ebf\u53c2\u6570<\/strong>\u7684\u5927\u6a21\u578b\uff0c\u53ef\u4ee5\u5f7b\u5e95\u8ddf\u82f1\u4f1f\u8fbeSay Goodbye\u4e86\u3002<\/p><p>\u5b8c\u6210\u6b64\u4e3e\u7684\uff0c\u6b63\u662f<strong>\u534e\u4e3a\uff01<\/strong><\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/e9b838f7a4dc077482a22b9df1feaa931746767408975.webp\" data-imgfileid=\"100309044\" data-ratio=\"1.0083333333333333\" data-s=\"300,640\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/e9b838f7a4dc077482a22b9df1feaa931746767408975.webp\" data-type=\"jpeg\" data-w=\"1080\" \/><\/section><p>\u6280\u672f\u62a5\u544a\uff1aarxiv.org\/abs\/2505.04519<\/p><p>\u8981\u77e5\u9053\uff0c\u5728\u6b64\u4e4b\u524d\uff0c\u8bad\u7ec3\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u8fd9\u4e8b\uff0c\u662f\u6709\u8bf8\u591a\u201c\u62e6\u8def\u864e\u201d\u5728\u8eab\u4e0a\u7684\u3002<\/p><p>\u4f8b\u5982\u8d1f\u8f7d\u5747\u8861\u96be\u3001\u901a\u4fe1\u5f00\u9500\u5927\u3001\u8bad\u7ec3\u6548\u7387\u4f4e\u7b49\u7b49\u3002<\/p><p>\u800c\u534e\u4e3a\u76d8\u53e4\u56e2\u961f\uff08\u5305\u542b\u8bfa\u4e9a\u65b9\u821f\u5b9e\u9a8c\u5ba4\u3001\u534e\u4e3a\u4e91\u7b49\uff09\u57fa\u4e8e<strong>\u6607\u817e\u56fd\u4ea7\u7b97\u529b\u5e73\u53f0<\/strong>\uff0c\u4e00\u4e3e\u653b\u7834\u4e86\u4e0a\u8ff0\u6240\u6709\u7684\u6311\u6218\u2014\u2014<\/p><p>6000+\u5757\u6607\u817eNPU\u96c6\u7fa4\u4e0a\u5b8c\u6210\u4e86<strong>7180\u4ebf\uff08718B\uff09\u53c2\u6570MoE\u6a21\u578b<\/strong>\u7684\u957f\u671f\u7a33\u5b9a\u8bad\u7ec3\uff0c\u5e76\u901a\u8fc7\u591a\u9879\u7a81\u7834\u6027\u7cfb\u7edf\u4f18\u5316\u6280\u672f\u5b9e\u73b0\u4e86\u663e\u8457\u6027\u80fd\u63d0\u5347\u3002<\/p><p>\u8fd9\u4e9b\u521b\u65b0\u5927\u5e45\u63d0\u9ad8\u4e86\u8bad\u7ec3\u6548\u7387\uff0c\u652f\u6491\u4e86\u884c\u4e1a\u9876\u5c16\u6c34\u5e73\u6a21\u578b\u7684\u5f00\u53d1\uff01<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/366477d9c93e7fafda0167ab1aa429a11746767408975.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/366477d9c93e7fafda0167ab1aa429a11746767408975.webp\" data-ratio=\"0.5609375\" data-type=\"gif\" data-w=\"640\" data-imgfileid=\"100308968\" \/><\/section><p>\u4e0d\u5f97\u4e0d\u8bf4\uff0c\u201c\u56fd\u4ea7\u201d\u4e8c\u5b57\u5728\u5927\u6a21\u578b\u786c\u4ef6\u4e0a\u7684\u542b\u91d1\u91cf\u8fd8\u5728\u6301\u7eed\u4e0a\u5347\u3002<\/p><h2>\u7eaf\u56fd\u4ea7NPU\uff0c\u4e1d\u6ed1\u8dd1\u901a\u51c6\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b<\/h2><p>\u5728\u62c6\u89e3\u534e\u4e3a\u4e00\u7cfb\u5217\u201c\u9ed1\u79d1\u6280\u201d\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u9700\u8981\u66f4\u6df1\u5165\u5730\u4e86\u89e3\u4e00\u4e0b\u8bad\u7ec3\u8d85\u5927\u53c2\u6570MoE\u6a21\u578b\u80cc\u540e\u7684\u56f0\u96be\u3002<\/p><p>\u603b\u4f53\u6765\u770b\uff0c\u5728\u8fd9\u6761\u8def\u4e0a\u6709\u201c\u56db\u5927\u91d1\u521a\u201d\u5728\u4e25\u9635\u628a\u5b88\u3002<\/p><p>\u9996\u5148\u5c31\u662f<strong>\u67b6\u6784\u53c2\u6570\u4f18\u5316\u96be\u9898<\/strong>\uff0c\u9700\u5728\u4f17\u591a\u53c2\u6570\u7ec4\u5408\u4e2d\u63a2\u7d22\u6700\u4f18\u914d\u7f6e\uff0c\u8bbe\u8ba1\u9002\u914d\u6607\u817eNPU\u7684\u5927\u89c4\u6a21MoE\u67b6\u6784\uff0c\u5b9e\u73b0\u8ba1\u7b97\u8d44\u6e90\u7684\u9ad8\u6548\u5229\u7528\u3002<\/p><p>\u5176\u6b21\u662f<strong>\u52a8\u6001\u8d1f\u8f7d\u5747\u8861\u6311\u6218<\/strong>\uff0c\u8def\u7531\u673a\u5236\u9700\u8981\u667a\u80fd\u5206\u914d\u4efb\u52a1\uff0c\u907f\u514d\u4e13\u5bb6\u8d44\u6e90\u5206\u914d\u4e0d\u5747\uff1b\u8fd9\u79cd\u4e0d\u5e73\u8861\u4e0d\u4ec5\u4f1a\u56e0\u201c\u6728\u6876\u6548\u5e94\u201d\u964d\u4f4e\u8bad\u7ec3\u6548\u7387\uff0c\u66f4\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u6536\u655b\u5f02\u5e38\uff0c\u5f71\u54cd\u6700\u7ec8\u6027\u80fd\u8868\u73b0\u3002<\/p><p>\u8fd8\u6709<strong>\u5206\u5e03\u5f0f\u901a\u4fe1\u7684\u74f6\u9888<\/strong>\uff0c\u5728\u8fd1\u4e07\u4ebf\u53c2\u6570\u89c4\u6a21\u4e0b\uff0ctoken\u5728\u4e0d\u540c\u8ba1\u7b97\u8282\u70b9\u95f4\u7684\u4e13\u5bb6\u6d41\u8f6c\u4f1a\u4ea7\u751f\u5de8\u5927\u901a\u4fe1\u5f00\u9500\uff0c\u201c\u901a\u4fe1\u5899\u201d\u95ee\u9898\u6210\u4e3a\u5236\u7ea6\u8bad\u7ec3\u6548\u7387\u7684\u5173\u952e\u56e0\u7d20\u3002<\/p><p>\u6700\u540e\u5c31\u662f<strong>\u786c\u4ef6\u9002\u914d\u590d\u6742\u5ea6<\/strong>\uff0c\u5b9e\u73b0MoE\u7b97\u6cd5\u4e0e\u6607\u817eNPU\u7b49\u4e13\u7528AI\u52a0\u901f\u5668\u7684\u6df1\u5ea6\u534f\u540c\uff0c\u9700\u8981\u6253\u901a\u7b97\u6cd5\u8bbe\u8ba1\u3001\u8f6f\u4ef6\u6846\u67b6\u548c\u786c\u4ef6\u7279\u6027\u7684\u5168\u6808\u4f18\u5316\uff0c\u5145\u5206\u91ca\u653e\u786c\u4ef6\u8ba1\u7b97\u6f5c\u529b\u3002<\/p><p>\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0c\u534e\u4e3a\u7684\u8fd9\u4efd\u6280\u672f\u62a5\u544a\u5206\u522b\u4ece\u6a21\u578b\u67b6\u6784\u3001MoE\u8bad\u7ec3\u5206\u6790\u3001\u7cfb\u7edf\u4f18\u5316\u7b49\u65b9\u9762\uff0c\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u5176\u5982\u4f55\u89c1\u62db\u62c6\u62db\u3002<\/p><p>\u9996\u5148\u5c31\u662f<strong>MoE\u7ed3\u6784\u9009\u578b\u4e0e\u6607\u817e\u4eb2\u548c\u7ed3\u6784\u4f18\u5316<\/strong>\u3002<\/p><p>\u56e2\u961f\u5148\u8fdb\u884c\u5148\u5bfc\u5b9e\u9a8c\uff0c\u786e\u5b9a\u4e86\u7ec6\u7c92\u5ea6\u4e13\u5bb6\u52a0\u4e0a\u5171\u4eab\u4e13\u5bb6\u8fd9\u6837\u7684\u8303\u5f0f\u3002\u968f\u540e\u5728\u6a21\u578b\u9009\u578b\u7684\u65f6\u5019\uff0c\u8003\u8651\u4e86\u591a\u4e2a\u65b9\u9762\u7684\u56e0\u7d20\u3002<\/p><p>\u5728\u8ba1\u7b97\u4e0e\u8bbf\u5b58\u4eb2\u548c\u65b9\u9762\uff0c\u901a\u8fc7\u589e\u5927\u6a21\u578b\u91cc\u7684hidden size\uff08\u9690\u85cf\u5c42\u5927\u5c0f\uff09\uff0c\u540c\u65f6\u964d\u4f4e\u6fc0\u6d3b\u53c2\u6570\u91cf\uff0c\u8fd9\u6837\u4e0d\u4ec5\u80fd\u63d0\u5347\u6a21\u578b\u7684\u8ba1\u7b97\u91cf\uff0c\u8fd8\u53ef\u4ee5\u964d\u4f4e\u8bbf\u5b58\u91cf\uff0c\u63d0\u9ad8\u4e86\u6a21\u578b\u8bad\u7ec3\u65f6\u5bf9\u7b97\u529b\u7684\u5229\u7528\u7387\uff0c\u4ee5\u53ca\u63a8\u7406\u65f6\u7684\u541e\u5410\u91cf\u3002<\/p><p>\u5728\u591a\u7ef4\u5e76\u884c\u4eb2\u548c\u65b9\u9762\uff0c\u91c7\u7528\u6570\u91cf\u4e3a2\u7684\u6307\u6570\u7ea7\u7684\u4e13\u5bb6\u6570\u91cf\uff0c\u8fbe\u6210\u4e86TP8\u00d7EP4\u8d85\u878d\u5408\u5e76\u884c\u7684\u65b9\u5f0f\u3002<\/p><p>\u8fd0\u7528TP-extend-EP\u6280\u672f\uff0c\u907f\u514d\u56e0 TP \u5207\u5206\u7ec6\u7c92\u5ea6\u4e13\u5bb6\u9020\u6210MatMul\uff08\u77e9\u9635\u4e58\u6cd5\uff09\u7b49\u7b97\u5b50\u7684\u6548\u7387\u4e0b\u964d\uff0c\u540c\u65f6\u4f7f\u7528\u5206\u7ec4 AllToAll \u901a\u4fe1\u6280\u672f\u6765\u51cf\u5c11 EP \u901a\u4fe1\u6240\u4ea7\u751f\u7684\u5f00\u9500\u3002<\/p><p>\u5728 DaVinci \u67b6\u6784\u4eb2\u548c\u65b9\u9762\uff0c\u5c06\u5f20\u91cf\u6309\u7167256\u8fdb\u884c\u5bf9\u9f50\u5904\u7406\uff0c\u4f7f\u5176\u80fd\u5b8c\u7f8e\u5339\u914d16\u00d716\u77e9\u9635\u8ba1\u7b97\u5355\u5143\uff0c\u5145\u5206\u91ca\u653e\u6607\u817eNPU\u7684\u7b97\u529b\u3002<\/p><p>\u5728\u6d41\u6c34\u7ebf\u7f16\u6392\u4eb2\u548c\u65b9\u9762\uff0c\u91c7\u7528PP\uff08\u6d41\u6c34\u7ebf\u5e76\u884c\uff09\u3001VPP\uff08\u53ef\u53d8\u6d41\u6c34\u7ebf\u5e76\u884c\uff09\u3001\u7a7a\u5c42\u7b49\u6280\u672f\uff0c\u5b9e\u73b0PP\u548cVPP\u7684\u8d1f\u8f7d\u5747\u8861\uff0c\u51cf\u5c11\u8ba1\u7b97\u8d44\u6e90\u95f2\u7f6e\uff08\u7a7a\u6ce1\uff09\u7684\u60c5\u51b5\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/d060d6486f750a4ea228e7839ef32b021746767408975.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/d060d6486f750a4ea228e7839ef32b021746767408975.webp\" data-ratio=\"0.5824074074074074\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308967\" \/><\/section><p>\u5728<strong>\u6a21\u578b\u7ed3\u6784\u4eff\u771f<\/strong>\u65b9\u9762\uff0c\u56e2\u961f\u6839\u636e\u786c\u4ef6\u7684\u9002\u914d\u7279\u70b9\uff0c\u5bf9\u6a21\u578b\u53c2\u6570\u7684\u9009\u62e9\u8303\u56f4\u8fdb\u884c\u4e86\u5927\u5e45\u8c03\u6574\uff0c\u628a\u539f\u672c\u5e9e\u5927\u7684\u53c2\u6570\u641c\u7d22\u7a7a\u95f4\u7f29\u5c0f\u5230\u4e8610000\u4e2a\u5de6\u53f3\u3002<\/p><p>\u4e3a\u4e86\u80fd\u66f4\u51c6\u786e\u5730\u77e5\u9053\u4e0d\u540c\u6a21\u578b\u7684\u6027\u80fd\u6781\u9650\uff0c\u56e2\u961f\u5f00\u53d1\u4e86\u4e00\u5957\u4e13\u95e8\u7684\u5efa\u6a21\u4eff\u771f\u5de5\u5177\u3002\u8fd9\u4e2a\u5de5\u5177\u5f88\u5389\u5bb3\uff0c\u5b83\u628a\u6a21\u578b\u7ed3\u6784\u3001\u8fd0\u884c\u65f6\u91c7\u7528\u7684\u7b56\u7565\uff0c\u8fd8\u6709\u786c\u4ef6\u7cfb\u7edf\uff0c\u90fd\u62c6\u5206\u6210\u4e86\u4e00\u4e2a\u4e2a\u5c0f\u7684\u53c2\u6570\u3002<\/p><p>\u901a\u8fc7\u5bf9\u7b97\u5b50\u3001Block\u3001Layer\u8fd9\u4e9b\u5c42\u7ea7\u7684\u8ba1\u7b97\u3001\u6570\u636e\u4f20\u8f93\u548c\u8bfb\u53d6\u64cd\u4f5c\u8fdb\u884c\u6a21\u62df\uff0c\u5c31\u80fd\u7b97\u51fa\u6a21\u578b\u4ece\u5934\u5230\u5c3e\u7684\u6574\u4f53\u6027\u80fd\u3002\u7ecf\u8fc7\u548c\u5b9e\u9645\u6d4b\u8bd5\u6570\u636e\u5bf9\u6bd4\uff0c\u53d1\u73b0\u8fd9\u4e2a\u4eff\u771f\u5de5\u5177\u7684\u51c6\u786e\u7387\u80fd\u8fbe\u523085%\u4ee5\u4e0a\u3002<\/p><p>\u56e2\u961f\u7528\u8fd9\u4e2a\u5efa\u6a21\u4eff\u771f\u5de5\u5177\uff0c\u628a\u6240\u6709\u7b26\u5408\u786c\u4ef6\u9002\u914d\u8981\u6c42\u7684\u53c2\u6570\u7ec4\u5408\u90fd\u6d4b\u8bd5\u4e86\u4e00\u904d\uff0c\u4ed4\u7ec6\u8bc4\u4f30\u5b83\u4eec\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u7684\u6570\u636e\u5904\u7406\u901f\u5ea6\uff0c\u6700\u540e\u627e\u5230\u4e86\u6027\u80fd\u76f8\u5bf9\u66f4\u597d\u7684\u6a21\u578b\u7ed3\u6784\uff0c\u5177\u4f53\u60c5\u51b5\u53ef\u4ee5\u770b\u4e0b\u9762\u7684\u56fe\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b47764b5933570e14b4ed08dafaa0c721746767408975.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b47764b5933570e14b4ed08dafaa0c721746767408975.webp\" data-ratio=\"0.44722222222222224\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308966\" \/><\/section><p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u518d\u770b\u4e0b<strong>MoE\u8bad\u7ec3\u7684\u5206\u6790<\/strong>\u3002<\/p><p>\u5728\u8bad\u7ec3MoE\u6a21\u578b\u7684\u65f6\u5019\uff0c\u548c\u666e\u901a\u7684\u7a20\u5bc6\u6a21\u578b\u76f8\u6bd4\uff0c\u6709\u4e2a\u7279\u522b\u8ba9\u4eba\u5934\u75bc\u7684\u95ee\u9898\uff0c\u5c31\u662f\u8d1f\u8f7d\u4e0d\u5747\u8861\u3002<\/p><p>\u6253\u4e2a\u6bd4\u65b9\uff0c\u5c31\u50cf\u4e00\u7fa4\u4eba\u5e72\u6d3b\uff0c\u6709\u7684\u4eba\u5fd9\u5f97\u4e0d\u53ef\u5f00\u4ea4\uff0c\u6709\u7684\u4eba\u5374\u95f2\u7740\u6ca1\u4e8b\u5e72\uff0c\u8fd9\u6837\u6548\u7387\u80af\u5b9a\u9ad8\u4e0d\u4e86\u3002<\/p><p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u79d1\u7814\u754c\u4ece\u7b97\u6cd5\u89d2\u5ea6\u60f3\u4e86\u5f88\u591a\u529e\u6cd5\uff0c\u63d0\u51fa\u4e86\u5404\u79cd\u5404\u6837\u7684\u8f85\u52a9\u635f\u5931\u51fd\u6570\uff0c\u8fd9\u4e9b\u51fd\u6570\u5173\u6ce8\u7684\u5747\u8861\u8303\u56f4\u4e0d\u592a\u4e00\u6837\u3002<\/p><p>\u6bd4\u5982\uff0c\u65e9\u671f\u6709\u4e13\u95e8\u9488\u5bf9\u5e8f\u5217\u7ea7\u522b\u7684\u5747\u8861\u8f85\u52a9\u635f\u5931\uff0c\u8fd8\u6709\u901a\u4e49\u5343\u95ee\u63d0\u51fa\u7684DP &#8211; Group\uff08\u4e5f\u5c31\u662f\u5168\u5c40\u6279\u6b21\u5927\u5c0f\uff09\u5747\u8861\u8f85\u52a9\u635f\u5931\u3002<\/p><p>\u8fd9\u4e9b\u8f85\u52a9\u635f\u5931\u51fd\u6570\uff0c\u5c31\u50cf\u662f\u7ed9MoE\u6a21\u578b\u91cc\u7684\u8def\u7531\u6a21\u5757\uff08\u8d1f\u8d23\u5206\u914d\u4efb\u52a1\u7684\u90e8\u5206\uff09\u5b9a\u4e86\u89c4\u77e9\uff0c\u901a\u8fc7\u4e0d\u540c\u7a0b\u5ea6\u7684\u7ea6\u675f\uff0c\u8ba9\u5b83\u628a\u4efb\u52a1\u5206\u914d\u5f97\u66f4\u5747\u5300\u4e00\u4e9b\u3002\u5177\u4f53\u7684\u7ea6\u675f\u60c5\u51b5\uff0c\u90fd\u6574\u7406\u5728\u4e0b\u9762\u7684\u8868\u683c\u91cc\u4e86\u3002<\/p><h6><strong><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/a5c8b97258e7736b10affc8ceacc17581746767408975.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/a5c8b97258e7736b10affc8ceacc17581746767408975.webp\" data-ratio=\"0.1919191919191919\" data-type=\"png\" data-w=\"1584\" data-imgfileid=\"100308964\" \/><\/strong><\/h6><h6><strong>\u25b3<\/strong>Balance BSZ\u8868\u793a\u7528\u6765\u8ba1\u7b97\u4e13\u5bb6\u9009\u62e9\u9891\u7387\u7684tokens\u4e2a\u6570<\/h6><p>\u56e2\u961f\u8fd8\u7814\u53d1\u51fa\u4e86\u4e00\u79cd<strong>\u5168\u65b0\u7684EP\u7ec4\u8d1f\u8f7d\u5747\u8861\u635f\u5931\u7b97\u6cd5<\/strong>\u3002<\/p><p>\u548c\u4f20\u7edf\u7684micro-batch\u8f85\u52a9\u635f\u5931\u76f8\u6bd4\uff0c\u5b83\u4e0d\u4f1a\u8fc7\u5ea6\u5f3a\u6c42\u5c40\u90e8\u4efb\u52a1\u5206\u914d\u7684\u7edd\u5bf9\u5747\u8861\uff0c\u907f\u514d\u4e86\u201c\u77eb\u6789\u8fc7\u6b63\u201d\uff1b\u8ddfDP\u7ec4\u7684\u5747\u8861\u635f\u5931\u6bd4\u8d77\u6765\uff0c\u5b83\u5728\u6570\u636e\u4f20\u8f93\u65f6\u8017\u8d39\u7684\u8d44\u6e90\u66f4\u5c11\uff0c\u80fd\u8282\u7701\u4e0d\u5c11\u901a\u4fe1\u6210\u672c\u3002<\/p><p>\u800c\u4e14\u5728\u5bf9\u4e13\u5bb6\u4efb\u52a1\u91cf\u7684\u7ea6\u675f\u7a0b\u5ea6\u4e0a\uff0c\u5b83\u5904\u4e8e\u4e24\u8005\u4e4b\u95f4\uff0c\u662f\u4e2a\u66f4\u6298\u4e2d\u7684\u65b9\u6848\u3002<\/p><p>\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e2a\u65b0\u7b97\u6cd5\u7684\u6548\u679c\uff0c\u56e2\u961f\u5728\u4e00\u4e2a\u603b\u53c2\u6570\u91cf\u8fbe200\u4ebf\uff0820B\uff09\u7684\u5148\u5bfcMoE\u6a21\u578b\u4e0a\uff0c\u4e13\u95e8\u505a\u4e86\u6d88\u878d\u5b9e\u9a8c\uff0c\u5177\u4f53\u60c5\u51b5\u5982\u4e0b\uff1a<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/7dffffa74c4446bea2097bb17c23008d1746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/7dffffa74c4446bea2097bb17c23008d1746767408976.webp\" data-ratio=\"0.3731481481481482\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308974\" \/><\/section><p>\u4e3a\u4e86\u5e94\u5bf9\u4e13\u5bb6\u8d1f\u8f7d\u4e0d\u5747\u7684\u201c\u6728\u6876\u6548\u5e94\u201d\uff0cMoE\u53ef\u4ee5\u91c7\u7528drop-and-pad\u7684\u65b9\u5f0f\u6765\u63d0\u5347\u8bad\u7ec3\u7684\u541e\u5410\u3002<\/p><p>\u56e2\u961f\u9996\u5148\u5728\u4e00\u4e2a20B\u7684\u5148\u5bfcMoE\u4e0a\u5bf9\u6bd4\u4e86\u4e0d\u540c\u4e13\u5bb6\u603b\u6570\u4e0bdrop-and-pad\u548cdropless\u7684\u6027\u80fd\uff1a<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/7af70fff411d4d7040383ce1be41361f1746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/7af70fff411d4d7040383ce1be41361f1746767408976.webp\" data-ratio=\"0.4981481481481482\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308971\" \/><\/section><p>\u7ed3\u679c\u663e\u793a\uff0cdropless\u603b\u662f\u4f18\u4e8edrop-and-pad\u65b9\u6848\u3002<\/p><p>\u5e76\u4e14\u8fd9\u79cd\u6027\u80fd\u7684\u5dee\u8ddd\u4f1a\u968f\u7740\u4e13\u5bb6\u6570\u53d8\u591a\u3001\u6a21\u578b\u53c2\u6570\u53d8\u5927\u800c\u8fdb\u4e00\u6b65\u653e\u5927\u3002<\/p><p>\u56e0\u6b64\u5728\u8bad\u7ec3\u76d8\u53e4Ultra MoE\u65f6\u91c7\u7528\u4e86dropless\u7684\u65b9\u6848\uff0c\u5e76\u91cd\u70b9\u4f18\u5316\u4e86\u8fd9\u4e00\u7b56\u7565\u4e0b\u7684\u8bad\u7ec3\u6548\u7387\u3002<\/p><p>\u5177\u4f53\u800c\u8a00\uff0c\u56e2\u961f\u4ece\u56db\u4e2a\u5173\u952e\u65b9\u5411\u5bf9\u76d8\u53e4Ultra MoE \u6a21\u578b\u8fdb\u884c\u4e86\u5168\u9762\u4f18\u5316\uff0c\u5305\u62ec\u6539\u8fdb\u5e76\u884c\u8ba1\u7b97\u7b56\u7565\u3001\u4f18\u5316\u6570\u636e\u4f20\u8f93\u6548\u7387\u3001\u63d0\u5347\u663e\u5b58\u4f7f\u7528\u6548\u679c\uff0c\u4ee5\u53ca\u8ba9\u4efb\u52a1\u5206\u914d\u66f4\u5747\u5300\u3002<\/p><p>\u5728\u75316000+\u4e2a\u6607\u817eNPU\u7ec4\u6210\u7684\u5927\u578b\u8ba1\u7b97\u96c6\u7fa4\u4e0a\uff0c\u6a21\u578b\u7684<strong>\u7b97\u529b\u5229\u7528\u7387\uff08MFU\uff0c\u5373Model FLOPs Utilization\uff09\u8fbe\u5230\u4e8630.0%<\/strong>\u00a0\uff0c\u548c\u4f18\u5316\u524d\u76f8\u6bd4\uff0c\u63d0\u5347\u5e45\u5ea6\u9ad8\u8fbe58.7%\u3002<\/p><p>\u56e2\u961f\u7528\u4e00\u5957\u80fd\u6a21\u62df\u5168\u6d41\u7a0b\u7684\u6a21\u578b\u4eff\u771f\u7cfb\u7edf\uff0c\u53cd\u590d\u8bd5\u9a8c\u5bfb\u627e\u6700\u4f73\u7684\u5e76\u884c\u8ba1\u7b97\u65b9\u6848\u3002<\/p><p>\u6700\u7ec8\u786e\u5b9a\u7684\u65b9\u6848\u662f\uff1a\u91c7\u752816\u8def\u6d41\u6c34\u7ebf\u5e76\u884c\u30018\u8def\u5f20\u91cf\u5e76\u884c\u30014\u8def\u4e13\u5bb6\u5e76\u884c\u30012\u8def\u865a\u62df\u6d41\u6c34\u7ebf\u5e76\u884c\uff0c\u4ee5\u53ca48\u8def\u6570\u636e\u5e76\u884c\u3002<\/p><p>\u5728\u4e13\u5bb6\u5e76\u884c\u8fd9\u5757\uff0c\u56e2\u961f\u7528\u4e86TP\u62d3\u5c55EP\u7684\u7b56\u7565\u3002<\/p><p>\u7b80\u5355\u6765\u8bf4\uff0c\u5c31\u662f\u8ba9TP\u7ec4\u6765\u5212\u5206\u4e13\u5bb6\u6570\u91cf\uff0c\u8fd9\u6837\u505a\u80fd\u907f\u514d\u56e0\u4e3aTP\u7ec4\u62c6\u5206\u4e13\u5bb6\u53c2\u6570\uff0c\u5bfc\u81f4GMM\u7b97\u5b50\u5728\u5904\u7406\u5c0f\u89c4\u6a21\u4e13\u5bb6\u6570\u636e\u65f6\u6548\u7387\u66b4\u8dcc\u7684\u95ee\u9898\u3002<\/p><p>\u6574\u4e2a\u7cfb\u7edf\u91cc\uff0c\u4e13\u5bb6\u7ec4\u603b\u6570\u662f32\u7ec4\uff08TP \u548c EP \u7ec4\u5408\u8ba1\u7b97\u5f97\u51fa\uff09\uff0c\u4e00\u5171\u5212\u5206\u6210256\u4e2a\u4e13\u5bb6\u3002<\/p><p>\u865a\u62df\u6d41\u6c34\u7ebf\u5e76\u884c\u7b56\u7565\u6548\u679c\u7279\u522b\u597d\uff0c\u4ee5\u524d\u8bad\u7ec3\u65f6\uff0c\u8ba1\u7b97\u8d44\u6e90\u95f2\u7f6e\uff08\u7a7a\u6ce1\u7387\uff09\u7684\u60c5\u51b5\u536018.98%\uff0c\u7528\u4e86\u65b0\u7b56\u7565\u540e\uff0c\u76f4\u63a5\u964d\u523010.49% \u3002<\/p><p>\u540c\u65f6\uff0c\u901a\u8fc7\u5408\u7406\u5206\u914dMTP\u5c42\u548c\u635f\u5931\u51fd\u6570\u5c42\u7684\u4efb\u52a1\uff0c\u628a\u4efb\u52a1\u5206\u914d\u4e0d\u5747\u8861\u5bfc\u81f4\u7684\u8d1f\u8f7d\u6ea2\u51fa\uff0c\u63a7\u5236\u57285%\u4ee5\u5185\uff0c\u5927\u5927\u51cf\u5c11\u4e86\u4efb\u52a1\u5206\u914d\u4e0d\u5747\u5e26\u6765\u7684\u8d1f\u9762\u5f71\u54cd\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b753e3bd499cd7b09fb06612da271f071746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b753e3bd499cd7b09fb06612da271f071746767408976.webp\" data-ratio=\"0.30648148148148147\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308972\" \/><\/section><p>\u4e3a\u4e86\u89e3\u51b3\u5e76\u884c\u6269\u5c55\u4e2d\u7684\u901a\u4fe1\u74f6\u9888\uff0c\u56e2\u961f\u8fd8\u8bbe\u8ba1\u4e86\u4e24\u4e2a\u4e3b\u8981\u6280\u672f\u3002<\/p><p>\u9996\u5148\u5c31\u662f<strong>Hierarchical EP Communication\u5206\u7ea7EP\u901a\u4fe1<\/strong>\u3002<\/p><p>\u76f8\u6bd4\u673a\u5185\u901a\u4fe1\uff0c\u8de8\u673a\u901a\u4fe1\u5e26\u5bbd\u8f83\u4f4e\u3002\u56e2\u961f\u91c7\u7528\u5206\u7ea7EP\u901a\u4fe1\uff0c\u51cf\u5c11\u8de8\u673a\u901a\u4fe1\u91cf\u3002<\/p><p>\u5177\u4f53\u6765\u8bf4\uff0c\u91c7\u7528\u8de8\u673aAllgather \u901a\u4fe1\u5c06\u6240\u6709tokens\u540c\u6b65\u5230\u673a\u5185\uff0c\u7136\u540e\u5728\u673a\u5185\u5bf9token\u6392\u5e8f\u5e76\u91c7\u7528\u673a\u5185AlltoAll\u901a\u4fe1\u5bf9tokens\u91cd\u65b0\u5206\u914d\u3002<\/p><p>\u673a\u5185\u901a\u4fe1\u548c\u673a\u95f4\u901a\u4fe1\u90fd\u53ef\u4ee5\u901a\u8fc7\u524d\u53cd\u5411\u901a\u4fe1\u63a9\u76d6\u6280\u672f\u63a9\u76d6\uff0c\u4ece\u4e0b\u56fe\u7684\u901a\u4fe1\u91cf\u5bf9\u6bd4\u53ef\u4ee5\u770b\u5230\u5206\u7ea7EP\u901a\u4fe1\u5bf9\u8de8\u673a\u901a\u4fe1\u91cf\u51cf\u5c11\u7684\u6548\u679c\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/6920a6770940e98bc4dcd6cde4ca30cb1746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/6920a6770940e98bc4dcd6cde4ca30cb1746767408976.webp\" data-ratio=\"0.2212962962962963\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308970\" \/><\/section><p>\u5176\u6b21\u662f<strong>Adaptive Pipe Overlap Mechanism\u81ea\u9002\u5e94\u524d\u53cd\u5411\u63a9\u76d6\u7b56\u7565<\/strong>\u3002<\/p><p>\u5373\u4f7f\u91c7\u7528\u5206\u7ea7EP\u901a\u4fe1\u7b56\u7565\uff0cEP\u901a\u4fe1\u7684\u8017\u65f6\u5360\u6bd4\u4ecd\u7136\u5f88\u9ad8\u3002\u524d\u53cd\u5411\u7684\u5927\u90e8\u5206EP\u901a\u4fe1\u4e0e\u8ba1\u7b97\u5747\u5177\u6709\u4f9d\u8d56\u5173\u7cfb\uff0c\u81ea\u7136\u63a9\u76d6\u7b56\u7565\u4f1a\u66b4\u9732\u5927\u90e8\u5206EP\u901a\u4fe1\u3002<\/p><p>\u5982\u679c\u91c7\u7528\u901a\u7b97\u878d\u5408\u7b97\u5b50\u7b49\u81ea\u63a9\u76d6\u7b56\u7565\uff0c\u53c8\u4e0d\u53ef\u907f\u514d\u5730\u4f1a\u964d\u4f4e\u8ba1\u7b97\u6548\u7387\u3002<\/p><p>\u56e0\u6b64\uff0c\u56e2\u961f\u91c7\u7528\u57fa\u4e8eVPP\u8c03\u5ea6\u7684\u81ea\u9002\u5e94\u524d\u53cd\u5411\u63a9\u76d6\u7b56\u7565\uff0c\u5b9e\u73b0\u5982\u4e0b\u56fe\u6d41\u7a0b\u7684\u524d\u5411\u8ba1\u7b97\u63a9\u76d6\u53cd\u5411\u901a\u4fe1\uff0c\u53cd\u5411\u8ba1\u7b97\u63a9\u76d6\u524d\u5411\u901a\u4fe1\u3002<\/p><p>\u6838\u5fc3\u8bbe\u8ba1\u5305\u62ec\uff1a\u5229\u7528\u673a\u95f4\u4e0e\u673a\u5185\u901a\u4fe1\u94fe\u8def\u5e26\u5bbd\u72ec\u7acb\u7279\u70b9\u5b9e\u73b0\u673a\u5185\u901a\u4fe1\u4e0e\u673a\u95f4\u901a\u4fe1\u7684\u4e92\u76f8\u63a9\u76d6\uff0c\u5229\u7528\u7b97\u5b50\u7684\u6709\u6548\u6392\u5e03\u7f13\u89e3host bound\uff0c\u5c06\u4e13\u5bb6\u53cd\u5411dw\u8ba1\u7b97\u4e0edx\u8ba1\u7b97\u5206\u79bb\u505a\u66f4\u7ec6\u7c92\u5ea6\u7684\u63a9\u76d6\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/5f6c2e6868aca36c6bcba0664f87a2cd1746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/5f6c2e6868aca36c6bcba0664f87a2cd1746767408976.webp\" data-ratio=\"0.44166666666666665\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308973\" \/><\/section><p>\u5bf9\u663e\u5b58\u8fdb\u884c\u4f18\u5316\u65f6\uff0c\u56e2\u961f\u91c7\u7528\u4e86\u65b0\u7684\u8ba1\u7b97\u65b9\u5f0f\u3002<\/p><p>\u4e0d\u518d\u4f7f\u7528\u4f20\u7edf\u7684\u5168\u91cd\u8ba1\u7b97\uff0c\u800c\u662f\u5bf9\u7ec6\u7c92\u5ea6\u6a21\u5757\uff0c\u50cfMLA\u3001Permute\u548c\u6fc0\u6d3b\u51fd\u6570\u8fdb\u884c\u91cd\u65b0\u8ba1\u7b97\uff0c\u8fd9\u6837\u80fd\u907f\u514d\u989d\u5916\u7684\u8ba1\u7b97\u6d88\u8017\u3002<\/p><p>\u540c\u65f6\uff0c\u8fd0\u7528Tensor Swapping\u6280\u672f\uff0c\u628a\u91cd\u65b0\u8ba1\u7b97\u4e0d\u592a\u5212\u7b97\u7684\u6fc0\u6d3b\u503c\uff0c\u5148\u8f6c\u79fb\u5230CPU\u90a3\u8fb9\uff0c\u7b49\u9700\u8981\u53cd\u5411\u8ba1\u7b97\u65f6\u518d\u63d0\u524d\u53d6\u56de\u6765\uff0c\u8ba9NPU\u5185\u5b58\u5f97\u5230\u66f4\u9ad8\u6548\u7684\u5229\u7528\u3002<\/p><p>\u56e2\u961f\u8fd8\u5728\u7814\u7a76\u65b0\u7684\u663e\u5b58\u8282\u7701\u65b9\u6cd5\uff0c\u51c6\u5907\u628a\u591a\u79cd\u4f18\u5316\u7b56\u7565\u7ec4\u5408\u8d77\u6765\uff0c\u6839\u636e\u4e0d\u540c\u7684\u8bbe\u5907\u914d\u7f6e\uff0c\u627e\u5230\u6700\u9002\u5408\u7684\u7ec4\u5408\uff0c\u65e2\u80fd\u63d0\u9ad8\u663e\u5b58\u5229\u7528\u7387\uff0c\u53c8\u4e0d\u4f1a\u964d\u4f4e\u6a21\u578b\u6027\u80fd\u3002<\/p><p>\u8ba9\u6bcf\u53f0\u8bbe\u5907\u4e0a\u4e13\u5bb6\u5904\u7406\u7684\u4efb\u52a1\u91cf\uff08token \u6570\u91cf\uff09\u5c3d\u91cf\u5747\u5300\uff0c\u80fd\u5927\u5e45\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u3002<\/p><p>\u4e3a\u6b64\uff0c\u56e2\u961f\u8bbe\u8ba1\u4e86\u4e00\u5957\u52a8\u6001\u7684\u8bbe\u5907\u7ea7\u8d1f\u8f7d\u5747\u8861\u673a\u5236\u3002<\/p><p>\u9996\u5148\uff0c\u89c4\u5212\u5668\u5c31\u50cf\u4e00\u4e2a\u201c\u5c0f\u7ba1\u5bb6\u201d\uff0c\u901a\u8fc7\u89c2\u5bdf\u4e00\u6bb5\u65f6\u95f4\u5185\u4e13\u5bb6\u7684\u5de5\u4f5c\u8d1f\u8f7d\u60c5\u51b5\uff0c\u9884\u6d4b\u672a\u6765\u7684\u4efb\u52a1\u91cf\uff0c\u518d\u7528\u8d2a\u5fc3\u7b97\u6cd5\u89c4\u5212\u51fa\u5982\u4f55\u91cd\u65b0\u5206\u914d\u4e13\u5bb6\uff0c\u8ba9\u8bbe\u5907\u95f4\u7684\u4efb\u52a1\u66f4\u5747\u8861\u3002<\/p><p>\u7136\u540e\uff0c\u6267\u884c\u5668\u5b9a\u671f\u884c\u52a8\uff0c\u628a\u4e0d\u540cTransformer\u5c42\u7684\u4e13\u5bb6\u53c2\u6570\u548c\u4f18\u5316\u5668\u72b6\u6001\u5728\u8bbe\u5907\u95f4\u8f6c\u79fb\u3002\u901a\u8fc7\u8fd9\u79cd\u52a8\u6001\u8c03\u6574\uff0c\u6a21\u578b\u7684MFU\u63d0\u9ad8\u4e8610%\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/8945cdb83deb7e5212697b5c0920cbe61746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/8945cdb83deb7e5212697b5c0920cbe61746767408976.webp\" data-ratio=\"0.5055555555555555\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308978\" \/><\/section><p>\u9664\u4e86\u4e0a\u9762\u8fd9\u4e9b\uff0c\u56e2\u961f\u8fd8\u5f00\u53d1\u4e86\u4e00\u4e9b\u4e13\u95e8\u9002\u914d\u6607\u817e\u8bbe\u5907\u7684\u6280\u672f\uff0c\u5305\u62ec\u4e3b\u673a\u7aef\u4f18\u5316\u3001\u8ba1\u7b97\u5378\u8f7d\u4e0e\u6570\u636e\u5171\u4eab\uff0c\u4ee5\u53ca\u878d\u5408\u7b97\u5b50\u3002<\/p><ul class=\"list-paddingleft-1\"><li><section>\u7b97\u5b50\u4e0b\u53d1\u4f18\u5316\uff1a\u4e3a\u4e86\u89e3\u51b3host\u7aef\u6027\u80fd\u74f6\u9888\u95ee\u9898\uff0c\u56e2\u961f\u51cf\u5c11\u4e86\u90a3\u4e9b\u9700\u8981\u9891\u7e41\u540c\u6b65\u64cd\u4f5c\u7684\u7b97\u5b50\uff0c\u907f\u514d\u4e0d\u5fc5\u8981\u7684\u7b49\u5f85\u3002\u540c\u65f6\uff0c\u4f7f\u7528\u7ec6\u7c92\u5ea6CPU\u7ed1\u6838\u6280\u672f\uff0c\u8ba9CPU\u548cNPU\u914d\u5408\u5f97\u66f4\u597d\uff0c\u4efb\u52a1\u4e0b\u53d1\u66f4\u987a\u7545\u3002<\/section><\/li><li><section>\u8ba1\u7b97\u5378\u8f7d\u4e0e\u6570\u636e\u5171\u4eab\uff1a\u5f53\u9047\u5230NPU\u5904\u7406\u8d77\u6765\u6548\u7387\u4f4e\u7684\u6570\u636e\u8ba1\u7b97\uff0c\u6216\u8005\u5728TP\u533a\u57df\u5185\u6570\u636e\u4f20\u8f93\u6162\u7684\u60c5\u51b5\uff0c\u4f5c\u8005\u628a\u8fd9\u4e9b\u4e0d\u9002\u5408NPU\u7684\u8ba1\u7b97\u4ece\u4e3b\u8ba1\u7b97\u6d41\u7a0b\u4e2d\u5206\u79bb\u51fa\u6765\uff0c\u4ea4\u7ed9CPU\u5728\u6570\u636e\u52a0\u8f7d\u65f6\u5904\u7406\u3002\u518d\u7ed3\u5408\u6570\u636e\u5171\u4eab\u6280\u672f\uff0c\u8ba9\u540c\u4e00\u8282\u70b9\u5185\u7684\u8ba1\u7b97\u548c\u6570\u636e\u4f20\u8f93\u901f\u5ea6\u90fd\u5927\u5927\u63d0\u9ad8\u3002<\/section><\/li><li><section>\u878d\u5408\u7b97\u5b50\uff1a\u9664\u4e86\u76d8\u53e4\u7a20\u5bc6\u6a21\u578b\u91cc\u5df2\u6709\u7684FlashAttention \u548c RMSNorm\u878d\u5408\u7b97\u5b50\uff0c\u56e2\u961f\u5728MoE\u6a21\u578b\u91cc\u53c8\u52a0\u5165\u4e86 GMMAdd\u3001Permute\u548cUmpermute\u878d\u5408\u7b97\u5b50\u3002GMMAdd\u878d\u5408\u7b97\u5b50\u628aGroupedMatMul\u7684\u53cd\u5411\u8ba1\u7b97\u548c\u68af\u5ea6\u7d2f\u52a0\u653e\u5728\u4e00\u8d77\u5904\u7406\uff0c\u5229\u7528\u5e76\u884c\u548c\u6d41\u6c34\u7ebf\u6280\u672f\u51cf\u5c11\u8c03\u5ea6\u65f6\u95f4\u3002Permute\u548cUnpermute\u878d\u5408\u7b97\u5b50\u6574\u5408\u4e86\u591a\u79cd\u64cd\u4f5c\uff0c\u80fd\u66f4\u5feb\u5730\u8bfb\u5199\u5185\u5b58\u3002<\/section><\/li><\/ul><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/3eb8af649101eefe7d6a9b4ef14e069c1746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/3eb8af649101eefe7d6a9b4ef14e069c1746767408976.webp\" data-ratio=\"0.319672131147541\" data-type=\"png\" data-w=\"976\" data-imgfileid=\"100308975\" \/><\/section><h2>\u5b9e\u9a8c\u7ed3\u679c<\/h2><p>\u5728\u8bad\u7ec3\u6570\u636e\u96c6\u6784\u5efa\u8fc7\u7a0b\u4e2d\uff0c\u56e2\u961f\u5b9e\u65bd\u4e25\u683c\u7684\u6570\u636e\u8d28\u91cf\u63a7\u5236\uff0c\u5e76\u7740\u91cd\u5f3a\u8c03\u8bed\u6599\u5e93\u7684\u591a\u6837\u6027\u3001\u590d\u6742\u6027\u548c\u5168\u9762\u6027\u3002<\/p><p>\u9488\u5bf9\u957f\u94fe\u601d\u7ef4\u6837\u672c\u5f15\u5165\u7279\u6b8a\u6807\u8bb0\u7b26\u53f7\u5bf9\u63a8\u7406\u8f68\u8ff9\u4e0e\u6700\u7ec8\u7b54\u6848\u8fdb\u884c\u7ed3\u6784\u5316\u5206\u9694\u3002<\/p><p>\u540e\u8bad\u7ec3\u9636\u6bb5\u91c7\u7528\u6307\u4ee4\u5fae\u8c03\u7b56\u7565\uff0c\u6570\u636e\u6db5\u76d6\u9886\u57df\u5e7f\u6cdb\uff0c\u5305\u542b\u901a\u7528\u95ee\u7b54\u3001\u6587\u672c\u751f\u6210\u3001\u8bed\u4e49\u5206\u7c7b\u3001\u4ee3\u7801\u7f16\u7a0b\u3001\u6570\u7406\u903b\u8f91\u63a8\u7406\u53ca\u5de5\u5177\u4f7f\u7528\u7b49\u3002<\/p><p>\u7279\u522b\u5c06\u63a8\u7406\u4e0e\u975e\u63a8\u7406\u6837\u672c\u6bd4\u4f8b\u8bbe\u5b9a\u4e3a3:1\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u63a8\u7406\u6027\u80fd\u3002<\/p><p>\u5b9e\u9a8c\u8868\u660e\uff0c\u76d8\u53e4Ultra MoE\u5bf9\u8bdd\u7248\u672c\u5728\u591a\u9886\u57df\u5747\u5c55\u73b0\u51fa\u5353\u8d8a\u7ade\u4e89\u529b\uff0c\u5728\u5927\u591a\u6570benchmark\u4e0a\u4e0eDeepSeek-R1\u8868\u73b0\u76f8\u5f53\u3002\u6bd4\u5982\u901a\u7528\u7406\u89e3\u4efb\u52a1\uff08\u5982CLUEWSC 94.8\u5206\u3001MMLU 91.5\u5206\uff09\u4e2d\u5c55\u73b0\u5353\u8d8a\u7406\u89e3\u529b\uff0c\u5728\u6570\u5b66\u63a8\u7406\u4e0e\u4ee3\u7801\u751f\u6210\u7b49\u9ad8\u96be\u5ea6\u6d4b\u8bd5\uff08\u5982AIME2024 81.3\u5206\u3001MBPP+ 81.2\u5206\uff09\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u5177\u5907\u7a81\u51fa\u7684\u4ee3\u7801\u4e0e\u6570\u5b66\u89e3\u9898\u80fd\u529b\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b4bcf1366f9403d9a535371ff14bb0231746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/b4bcf1366f9403d9a535371ff14bb0231746767408976.webp\" data-ratio=\"0.4685185185185185\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308979\" \/><\/section><p>\u56e2\u961f\u8fd8\u5bf9\u76d8\u53e4Ultra MoE\u8fdb\u884c\u4e86\u4e13\u5bb6\u4e13\u4e1a\u5ea6\u5206\u6790\u3002<\/p><p>\u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\uff0c\u540c\u4e00\u7f51\u7edc\u5c42\u7684token\u4f1a\u88ab\u4f18\u5148\u8def\u7531\u81f3\u4e0d\u540c\u4e13\u5bb6\uff0c\u4e13\u5bb6\u4e13\u4e1a\u5316\u7a0b\u5ea6\u5b58\u5728\u663e\u8457\u4efb\u52a1\u5dee\u5f02\u6027\u3002<\/p><p>\u8fd9\u8bc1\u5b9e\u4e86\u76d8\u53e4Ultra MoE\u5df2\u5f62\u6210\u663e\u8457\u7684\u4e13\u5bb6\u5dee\u5f02\u5316\uff0c\u8fd9\u79cd\u7279\u6027\u4e0d\u4ec5\u589e\u5f3a\u4e86\u6a21\u578b\u7684\u8868\u8fbe\u80fd\u529b\uff0c\u66f4\u4e3a\u5176\u5353\u8d8a\u6027\u80fd\u63d0\u4f9b\u4e86\u5173\u952e\u652f\u6491\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/ea6e75b65e79e3356620603bdbf2d0811746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/ea6e75b65e79e3356620603bdbf2d0811746767408976.webp\" data-ratio=\"0.6555555555555556\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308977\" \/><\/section><p>\u76d8\u53e4Ultra MoE\u7684MoE\u5c42\u8f93\u51fa\u7531\u5171\u4eab\u4e13\u5bb6\u548c\u8def\u7531\u4e13\u5bb6\u5171\u540c\u8d21\u732e\u7684\u52a0\u6743\u548c\u6784\u6210\u3002<\/p><p>\u56e0\u6b64\uff0c\u4fdd\u6301\u4e8c\u8005\u8f93\u51fa\u7684\u5e73\u8861\u81f3\u5173\u91cd\u8981\u3002<\/p><p>\u4e0b\u56fe\u4e2d\u5c55\u793a\u4e86\u8def\u7531\u4e13\u5bb6\u5728\u5404\u7f51\u7edc\u5c42\u5747\u4fdd\u6301\u7740\u4e0e\u5171\u4eab\u4e13\u5bb6\u76f8\u5f53\u7684\u8d21\u732e\u5f3a\u5ea6\uff0c\u8fd9\u79cd\u5747\u8861\u7684\u534f\u540c\u4f5c\u7528\u6709\u6548\u63d0\u5347\u4e86\u6a21\u578b\u7684\u6574\u4f53\u8868\u5f81\u80fd\u529b\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/f916aa97c4879aa45eea35753d582c591746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/f916aa97c4879aa45eea35753d582c591746767408976.webp\" data-ratio=\"0.4444444444444444\" data-type=\"png\" data-w=\"1080\" data-imgfileid=\"100308976\" \/><\/section><p>\u56e2\u961f\u8fd8\u5206\u6790\u4e86\u4e13\u5bb6\u7684\u5171\u6fc0\u6d3b\u73b0\u8c61\uff0c\u6fc0\u6d3b\u5206\u6570\u8d8a\u9ad8\uff0c\u8bf4\u660e\u4e24\u4e2a\u4e13\u5bb6\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u8d8a\u5f3a\u3002<\/p><p>\u5728\u4e0b\u56fe\u4e2d\uff0c\u9664\u5c11\u6570\u4f8b\u5916\u60c5\u51b5\u5916\uff0c\u8fd9\u4e09\u5c42\u4e2d\u7684\u4e13\u5bb6\u4e4b\u95f4\u5e76\u672a\u51fa\u73b0\u660e\u663e\u7684\u5171\u6fc0\u6d3b\u73b0\u8c61\uff0c\u8fd9\u53cd\u6620\u4e86\u76d8\u53e4Ultra MoE\u7684\u4e13\u5bb6\u5197\u4f59\u5ea6\u8f83\u4f4e\u3002<\/p><section><img decoding=\"async\" class=\"rich_pages wxw-img\" src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/e6c0a9ed0a4df6ca1660e78f41fb64b21746767408976.webp\" data-src=\"https:\/\/simg.baai.ac.cn\/hub-detail\/e6c0a9ed0a4df6ca1660e78f41fb64b21746767408976.webp\" data-ratio=\"0.9853249475890985\" data-type=\"jpeg\" data-w=\"954\" data-imgfileid=\"100308980\" \/><\/section><p>\u4ee5\u4e0a\u4fbf\u662f\u534e\u4e3a\u56fd\u4ea7NPU\u8dd1\u51c6\u4e07\u4ebf\u53c2\u6570\u5927\u6a21\u578b\u80cc\u540e\u7684\u5965\u4e49\u4e86\u3002<\/p><p>\u534e\u4e3a\u76d8\u53e4Ultra MoE\u6280\u672f\u7684\u7a81\u7834\uff0c\u4e0d\u4ec5\u6807\u5fd7\u7740\u56fd\u4ea7\u7b97\u529b\u5e73\u53f0\u5728AI\u5927\u6a21\u578b\u8bad\u7ec3\u9886\u57df\u8fc8\u5165\u4e16\u754c\u9886\u5148\u884c\u5217\uff0c\u66f4\u5f70\u663e\u4e86\u4e2d\u56fd\u79d1\u6280\u81ea\u4e3b\u521b\u65b0\u7684\u5f3a\u5927\u5b9e\u529b\u3002<\/p><p>\u5b83\u8bc1\u660e\u4e86\u4e2d\u56fd\u4f01\u4e1a\u5728\u5168\u7403AI\u7ade\u8d5b\u4e2d\u5df2\u5177\u5907\u4ece\u8ddf\u8dd1\u5230\u5e76\u8dd1\uff0c\u751a\u81f3\u9886\u8dd1\u7684\u5b9e\u529b\u3002<\/p><p>\u672a\u6765\uff0c\u968f\u7740\u6280\u672f\u7684\u6301\u7eed\u8fed\u4ee3\u4e0e\u5e94\u7528\u573a\u666f\u7684\u62d3\u5c55\uff0c\u76d8\u53e4Ultra MoE\u5c06\u4e3a\u5343\u884c\u767e\u4e1a\u7684\u667a\u80fd\u5316\u8f6c\u578b\u6ce8\u5165\u5f3a\u52b2\u52a8\u529b\uff0c\u52a9\u529b\u4e2d\u56fd\u5728\u65b0\u4e00\u8f6e\u79d1\u6280\u9769\u547d\u4e2d\u5360\u636e\u5236\u9ad8\u70b9\uff0c\u4e3a\u4eba\u7c7b\u79d1\u6280\u8fdb\u6b65\u8d21\u732e\u66f4\u591a\u201c\u4e2d\u56fd\u667a\u6167\u201d\u3002<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"<p>\u91d1\u78ca \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n","protected":false},"author":1,"featured_media":3783,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3781","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3781","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=3781"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3781\/revisions"}],"predecessor-version":[{"id":3789,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/3781\/revisions\/3789"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/3783"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=3781"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=3781"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=3781"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}