{"id":147,"date":"2026-05-02T18:09:53","date_gmt":"2026-05-02T10:09:53","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=147"},"modified":"2026-04-23T18:10:00","modified_gmt":"2026-04-23T10:10:00","slug":"%e6%95%b0%e6%8d%ae%e5%b9%b6%e8%a1%8c%e3%80%81%e5%bc%a0%e9%87%8f%e5%b9%b6%e8%a1%8c%e5%92%8c%e6%b5%81%e6%b0%b4%e7%ba%bf%e5%b9%b6%e8%a1%8c%e8%a7%a3%e6%9e%90","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/05\/02\/%e6%95%b0%e6%8d%ae%e5%b9%b6%e8%a1%8c%e3%80%81%e5%bc%a0%e9%87%8f%e5%b9%b6%e8%a1%8c%e5%92%8c%e6%b5%81%e6%b0%b4%e7%ba%bf%e5%b9%b6%e8%a1%8c%e8%a7%a3%e6%9e%90\/","title":{"rendered":"\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u548c\u6d41\u6c34\u7ebf\u5e76\u884c\u89e3\u6790"},"content":{"rendered":"\n<h1 class=\"wp-block-heading\">\u5199\u5728\u524d\u9762<\/h1>\n\n\n\n<p class=\"wp-block-paragraph\">\u76ee\u524d\u4e3a\u6b62\u7684\u6587\u7ae0\u90fd\u5728\u804a\u4e00\u4ef6\u4e8b\uff1a<strong>\u5355\u4e2a\u6a21\u578b\u5185\u90e8\u7684\u201c\u5fae\u89c2\u4e16\u754c\u201d<\/strong> \u2014\u2014Embedding\u600e\u4e48\u5de5\u4f5c\u3001Attention\u600e\u4e48\u8ba1\u7b97\u3001FFN\u600e\u4e48\u5b58\u50a8\u77e5\u8bc6\u3001MoE\u600e\u4e48\u8c03\u5ea6\u4e13\u5bb6\u3002\u4f46\u6211\u4e00\u76f4\u5728\u4e00\u4e2a\u9ed8\u8ba4\u5047\u8bbe\u4e0b\u8bb2\u6545\u4e8b\uff1a\u6a21\u578b\u80fd\u5b8c\u6574\u88c5\u8fdb\u4e00\u53f0GPU\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u73b0\u5b9e\u5f53\u7136\u4e0d\u662f\u8fd9\u6837\u3002\u4eca\u5929\u7684\u5927\u8bed\u8a00\u6a21\u578b\u52a8\u8f84\u4e0a\u5343\u4ebf\u53c2\u6570\uff1a\u4e00\u4e2a7B\u6a21\u578b\u5c31\u5df2\u7ecf\u5403\u6389\u7ea614GB\u663e\u5b58\uff08FP16\u683c\u5f0f\u4e0b\u4ec5\u53c2\u6570\u6743\u91cd\u90e8\u5206\uff09\uff0c\u800c70B\u6a21\u578b\u76f4\u5954140GB\u3002\u5355\u5f20H100\u6700\u9ad8\u4e5f\u53ea\u670980GB\u663e\u5b58\uff0c\u653e\u5927\u6a21\u578b\u540e\u8fde\u4e00\u5f20\u5361\u90fd\u88c5\u4e0d\u4e0b\u3002\u66f4\u522b\u8bf4\u8bad\u7ec3\u65f6\u8fd8\u8981\u989d\u5916\u5b58\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u548c\u4e2d\u95f4\u6fc0\u6d3b\u503c\u2014\u2014\u8fd9\u4e9b\u8f85\u52a9\u6570\u636e\u7684\u663e\u5b58\u9700\u6c42\u5f80\u5f80\u662f\u6a21\u578b\u53c2\u6570\u91cf\u7684\u597d\u51e0\u500d\u3002\u4e8e\u662f\uff0c\u6211\u4eec\u4e0d\u5f97\u4e0d\u7528\u6210\u767e\u4e0a\u5343\u5f20\u663e\u5361\u6765\u534f\u4f5c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4ef6\u4e8b\u542c\u8d77\u6765\u7b80\u5355\uff1a\u201c\u591a\u653e\u51e0\u5f20\u5361\u4e0d\u5c31\u884c\u4e86\uff1f\u201d\u771f\u7684\u505a\u8d77\u6765\uff0c\u96be\u9898\u63a5\u8e35\u800c\u6765\uff1a<strong>\u6a21\u578b\u600e\u4e48\u5207\uff1f\u6570\u636e\u600e\u4e48\u5206\uff1f\u5361\u548c\u5361\u4e4b\u95f4\u600e\u4e48\u201c\u6c9f\u901a\u201d\uff1f\u804a\u5929\u592a\u9891\u7e41\u4f1a\u5835\u4f4f\uff0c\u804a\u592a\u5c11\u4f1a\u7b97\u9519\uff0c\u4e71\u804a\u5929\u4f1a\u5d29\u3002<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4eca\u5929\u7684\u6587\u7ae0\u5c31\u662f\u8981\u8bb2\u6e05\u695a\uff1a\u5de5\u4e1a\u754c\u5230\u5e95\u7528\u4ec0\u4e48\u65b9\u6cd5\u628a\u4e00\u53f0\u88c5\u4e0d\u4e0b\u7684\u6a21\u578b\uff0c\u62c6\u5230\u51e0\u767e\u51e0\u5343\u5f20\u5361\u4e0a\u8bad\u7ec3\uff0c\u53c8\u8981\u5feb\u53c8\u8981\u7a33\u3002\u8fd9\u4e2a\u9886\u57df\u76ee\u524d\u516c\u8ba4\u7684\u201c\u4e09\u677f\u65a7\u201d\u6b63\u662f\u2014\u2014<strong>\u6570\u636e\u5e76\u884c\uff08Data Parallelism\uff09\u3001\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallelism\uff09\u548c\u6d41\u6c34\u7ebf\u5e76\u884c\uff08Pipeline Parallelism\uff09<\/strong>\uff0c\u4e09\u8005\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u5c31\u662f\u73b0\u4ee3\u5927\u6a21\u578b\u8bad\u7ec3\u7684\u201c3D\u5e76\u884c\u201d\u6838\u5fc3\u67b6\u6784\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bfb\u5b8c\u8fd9\u7bc7\u6587\u7ae0\uff0c\u4f60\u4f1a\u660e\u767d\u4e3a\u4ec0\u4e48\u8bad\u7ec3Llama-70B\u9700\u8981\u51e0\u767e\u5f20H100\uff0c\u4ee5\u53ca\u90a3\u4e09\u4e2a\u5b57\u6bcd\u7f29\u5199\u7684\u80cc\u540e\u5404\u81ea\u5728\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898\u3001\u5404\u81ea\u53c8\u6709\u4ec0\u4e48\u5148\u5929\u4e0d\u8db3\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u6570\u636e\u5e76\u884c\uff1a\u6700\u76f4\u89c2\u7684\u201c\u4eba\u591a\u529b\u91cf\u5927\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1.1 \u539f\u59cb\u76f4\u89c9\u4e0e\u6807\u51c6\u6d41\u7a0b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u5e76\u884c\u662f\u6700\u5148\u88ab\u60f3\u5230\u7684\u65b9\u6848\uff1a\u6a21\u578b\u672c\u8eab\u80fd\u4e0d\u80fd\u653e\u8fdbGPU\u4e0d\u91cd\u8981\uff0c\u5148\u628a\u6a21\u578b\u5b8c\u6574\u590d\u5236\u5230\u6bcf\u4e00\u5f20\u5361\u4e0a\uff0c\u7136\u540e\u628a\u8bad\u7ec3\u6570\u636e\u5207\u788e\uff0c\u6bcf\u5f20\u5361\u5404\u81ea\u5403\u4e00\u5c0f\u4efd\u3002\u6bcf\u5f20\u5361\u5728\u201c\u76f8\u540c\u7684\u6a21\u578b\u526f\u672c\u201d\u4e0a\u5206\u522b\u8ba1\u7b97\u68af\u5ea6\uff0c\u6700\u540e\u628a\u6240\u6709\u4eba\u7684\u68af\u5ea6\u6c47\u603b\u8d77\u6765\u6c42\u4e00\u4e2a\u5e73\u5747\uff0c\u7528\u8fd9\u4e2a\u201c\u5e73\u5747\u68af\u5ea6\u201d\u53bb\u7edf\u4e00\u66f4\u65b0\u5404\u5f20\u5361\u4e0a\u7684\u6a21\u578b\u53c2\u6570\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u6d41\u7a0b\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u590d\u5236\u6a21\u578b<\/strong>\uff1a\u5c06\u5b8c\u5168\u76f8\u540c\u7684\u6a21\u578b\u526f\u672c\u52a0\u8f7d\u5230\u6bcf\u5f20GPU\u4e0a\u3002<\/li>\n\n\n\n<li><strong>\u5206\u53d1\u6570\u636e<\/strong>\uff1a\u5c06\u4e00\u4e2aBatch\u7684\u8bad\u7ec3\u6837\u672c\u5747\u5300\u5206\u6210\u82e5\u5e72Micro-Batch\uff0c\u6bcf\u5f20\u5361\u5904\u7406\u5176\u4e2d\u4e00\u4efd\u3002<\/li>\n\n\n\n<li><strong>\u72ec\u7acb\u8ba1\u7b97\u68af\u5ea6<\/strong>\uff1a\u5404GPU\u72ec\u7acb\u5b8c\u6210\u524d\u5411\u4f20\u64ad\u548c\u53cd\u5411\u4f20\u64ad\uff0c\u7b97\u51fa\u5404\u81ea\u7684\u5c40\u90e8\u68af\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u901a\u4fe1\u540c\u6b65\u68af\u5ea6<\/strong>\uff1a\u6267\u884cAll-Reduce\u901a\u4fe1\uff0c\u628a\u6240\u6709GPU\u7684\u68af\u5ea6\u6c47\u603b\u6210\u5168\u5c40\u5e73\u5747\u68af\u5ea6\u3002<\/li>\n\n\n\n<li><strong>\u540c\u6b65\u66f4\u65b0\u53c2\u6570<\/strong>\uff1a\u6bcf\u5f20GPU\u7528\u8fd9\u4e2a\u5168\u5c40\u68af\u5ea6\u7edf\u4e00\u66f4\u65b0\u672c\u5730\u6a21\u578b\u53c2\u6570\u3002<\/li>\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u6548\u679c\u770b\uff0c\u6570\u636e\u5e76\u884c\u76f4\u63a5\u201c\u5e76\u884c\u653e\u5927\u201d\u4e86\u8bad\u7ec3\u541e\u5410\u91cf\u3002\u5b83\u4e5f\u662f\u76ee\u524d\u5b9e\u73b0\u6700\u7b80\u5355\u3001\u5e94\u7528\u6700\u5e7f\u6cdb\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u65b9\u5f0f\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.2 \u770b\u4f3c\u5b8c\u7f8e\u7684\u4ee3\u4ef7\uff1a\u81a8\u80c0\u7684\u901a\u4fe1\u4e0e\u5197\u4f59\u7684\u53c2\u6570<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u5e76\u884c\u7684\u6838\u5fc3\u95ee\u9898\u53ef\u4ee5\u7528\u4e00\u4e2a\u8bcd\u6982\u62ec\uff1a<strong>\u5197\u4f59<\/strong>\u3002\u6a21\u578b\u5fc5\u987b\u88ab\u5b8c\u6574\u590d\u5236\u5230\u6bcf\u5f20\u5361\u4e0a\uff0c\u5bfc\u81f4\u4e00\u4e2a\u76f4\u63a5\u7684\u77db\u76fe\uff1a\u6240\u6709GPU\u7684\u603b\u663e\u5b58\u52a0\u8d77\u6765\u53ef\u80fd\u975e\u5e38\u5927\uff0c\u4f46\u6bcf\u5f20\u5361\u80fd\u770b\u5230\u7684\u201c\u6709\u6548\u663e\u5b58\u201d\u4ecd\u53d7\u9650\u4e8e\u5355\u5361\u5bb9\u91cf\u300270B\u7684\u6a21\u578b\u8981FP16\u5b58\u50a8\uff0c\u5355\u5361\u4e0d\u591f\u5c31\u662f\u4e0d\u591f\uff0c\u52a0\u518d\u591a\u5361\u4e5f\u65e0\u6d4e\u4e8e\u4e8b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u5916\u4e00\u4e2a\u9690\u5f62\u95ee\u9898\u662f\u901a\u4fe1\u5f00\u9500\u3002\u6bcf\u4e00\u6b21\u68af\u5ea6\u540c\u6b65\u90fd\u9700\u8981\u6267\u884cAll-Reduce\u64cd\u4f5c\uff0c\u5176\u6240\u4f20\u8f93\u7684\u6570\u636e\u91cf\u76f8\u5f53\u4e8e\u6a21\u578b\u53c2\u6570\u91cf\u3002\u5f53GPU\u6570\u91cf\u589e\u591a\u3001\u6a21\u578b\u53c2\u6570\u4e0d\u65ad\u81a8\u80c0\u65f6\uff0c\u540c\u6b65\u901a\u4fe1\u53ef\u80fd\u4f1a\u8017\u65f6\u6570\u5341\u6beb\u79d2\u751a\u81f3\u4e0a\u767e\u6beb\u79d2\uff0c\u6210\u4e3a\u5de8\u5927\u7684\u6027\u80fd\u74f6\u9888\u3002\u6240\u4ee5\u6570\u636e\u5e76\u884c\u867d\u7136\u662f\u5fc5\u9009\u9879\uff0c\u4f46\u5728\u8bad\u7ec3\u5de8\u578b\u6a21\u578b\u65f6\u7edd\u4e0d\u53ef\u80fd\u201c\u5355\u6253\u72ec\u6597\u201d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1.3 \u4f18\u5316\u601d\u8def\uff1a\u68af\u5ea6\u7d2f\u79ef\u4e0e\u901a\u4fe1-\u8ba1\u7b97\u91cd\u53e0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9488\u5bf9\u6570\u636e\u5e76\u884c\u7684\u901a\u4fe1\u74f6\u9888\uff0c\u5e38\u89c1\u7684\u4e00\u7c7b\u4f18\u5316\u662f<strong>\u68af\u5ea6\u7d2f\u79ef<\/strong>\uff1a\u4e0d\u4e00\u5b9a\u8981\u6bcf\u4e2aMicro-Batch\u7ed3\u675f\u5c31\u7acb\u523b\u6267\u884cAll-Reduce\uff0c\u53ef\u4ee5\u8ba9\u6bcf\u5f20\u5361\u72ec\u7acb\u8ba1\u7b97\u82e5\u5e72\u4e2aMicro-Batch\u540e\uff0c\u628a\u68af\u5ea6\u7d2f\u52a0\u8d77\u6765\uff0c\u518d\u7edf\u4e00\u505a\u4e00\u6b21\u901a\u4fe1\u540c\u6b65\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6709\u6548\u5168\u5c40Batch Size\u6ca1\u6709\u53d8\uff0c\u4f46\u901a\u4fe1\u9891\u7387\u5927\u5e45\u964d\u4f4e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u53e6\u4e00\u7c7b\u4f18\u5316\u662f\u5728\u5e95\u5c42\u5b9e\u73b0\u4e0a\u505a\u201c\u901a\u4fe1-\u8ba1\u7b97\u91cd\u53e0\u201d\uff1a\u7528PyTorch DDP\u5185\u7f6e\u7684<code>async_op=True<\/code>\u7b49\u673a\u5236\uff0c\u5728\u53cd\u5411\u4f20\u64ad\u8ba1\u7b97\u6700\u540e\u4e00\u5c42\u7684\u540c\u65f6\uff0c\u63d0\u524d\u542f\u52a8\u5e95\u5c42\u901a\u4fe1\u94fe\u8def\u4f20\u8f93\u5df2\u5c31\u7eea\u7684\u68af\u5ea6\u5f20\u91cf\uff0c\u628a\u901a\u4fe1\u5f00\u9500\u201c\u85cf\u201d\u5728\u8ba1\u7b97\u65f6\u95f4\u91cc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0d\u8fc7\uff0c\u65e0\u8bba\u600e\u4e48\u4f18\u5316\uff0c\u6570\u636e\u5e76\u884c\u7684\u6839\u672c\u9650\u5236\u6ca1\u6709\u6d88\u5931\uff1a<strong>\u5f53\u5355\u5361\u663e\u5b58\u6839\u672c\u653e\u4e0d\u4e0b\u5b8c\u6574\u6a21\u578b\u65f6\uff0c\u5b83\u5fc5\u987b\u4e0e\u5176\u4ed6\u5e76\u884c\u65b9\u5f0f\u534f\u540c\u4f5c\u6218\u3002<\/strong><\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u5f20\u91cf\u5e76\u884c\uff1a\u5f53\u6a21\u578b\u672c\u8eab\u5c31\u5927\u5230\u8981\u201c\u6495\u5f00\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">2.1 \u6838\u5fc3\u601d\u60f3\uff1a\u628a\u6743\u91cd\u77e9\u9635\u5207\u6210\u201c\u5e73\u884c\u788e\u7247\u201d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f53Transformer\u91cc\u7684\u4e00\u4e2a\u77e9\u9635\u4e58\u6cd5\u672c\u8eab\u5c31\u5927\u5230\u8fde\u4e00\u5c42\u90fd\u585e\u4e0d\u8fdb\u4e00\u5f20\u5361\u65f6\uff0c\u5c31\u5fc5\u987b\u7528\u66f4\u7ec6\u7c92\u5ea6\u7684\u5207\u6cd5\u2014\u2014<strong>\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallelism, TP\uff09<\/strong> \u5c31\u662f\u7528\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u7684\u6280\u672f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f20\u91cf\u5e76\u884c\u7684\u6838\u5fc3\u662f\u628a\u5355\u4e2a\u5c42\u7684\u6743\u91cd\u5f20\u91cf\uff08\u5982\u5927\u7684\u7ebf\u6027\u5c42\u53c2\u6570\u77e9\u9635W\uff09\u6cbf\u7740\u7279\u5b9a\u7ef4\u5ea6\u5207\u5206\u6210\u591a\u4efd\uff0c\u5206\u522b\u653e\u5230\u4e0d\u540cGPU\u4e0a\uff0c\u6bcf\u4e2aGPU\u53ea\u4fdd\u5b58\u4e00\u90e8\u5206\u6743\u91cd\u3002\u4ee5<code>X \u00d7 W<\/code>\u4e3a\u4f8b\uff1aX\u662f\u8f93\u5165\u5f20\u91cf\uff08\u7ef4\u5ea6[Batch, SeqLen, HiddenSize]\uff09\uff0cW\u662f\u6743\u91cd\u5f20\u91cf\uff08\u7ef4\u5ea6[HiddenSize, HiddenSize&#8217;]\uff09\u3002\u5728\u5355\u5361\u4e0a\uff0c\u8fd9\u4e2a\u4e58\u6cd5\u80fd\u591f\u76f4\u63a5\u5b8c\u6210\u3002\u4f46\u5982\u679cW\u5b9e\u5728\u592a\u5927\u4e86\uff0c\u9700\u8981\u5728\u591a\u5f20\u5361\u4e0a\u201c\u8054\u5408\u8ba1\u7b97\u201d\uff0c\u4e3b\u8981\u6709\u4e24\u79cd\u7ecf\u5178\u5207\u6cd5\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6309\u884c\u5207\u5206<\/strong>\uff1a\u6cbfW\u7684\u884c\u7ef4\u5ea6\u628aW\u5207\u6210\u51e0\u4efd\u3002\u5207W\u7684\u884c\u540e\uff0c\u8f93\u5165X\u4e5f\u9700\u8981\u6309\u5217\u5bf9\u9f50\u5207\u5272\u3002\u6bcf\u5f20\u5361\u62ff\u5206\u7247\u540e\u7684X\u548cW\u5148\u5404\u81ea\u505a\u77e9\u9635\u4e58\u6cd5\uff0c\u5f97\u5230\u90e8\u5206\u7ed3\u679cY_partial\uff1b\u7136\u540e\u6267\u884cAll-Reduce\u901a\u4fe1\uff0c\u628a\u5404\u5361\u7684\u90e8\u5206\u7ed3\u679c\u6c42\u548c\u5230\u6bcf\u4e2aGPU\u4e0a\uff0c\u5f97\u5230\u5b8c\u6574\u8f93\u51faY\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6309\u5217\u5207\u5206<\/strong>\uff1a\u6cbfW\u7684\u5217\u7ef4\u5ea6\u628aW\u5207\u6210\u51e0\u4efd\u3002\u8f93\u5165X\u56e0\u4e3a\u5217\u7ef4\u5ea6\u5e94\u548cW\u5207\u5206\u540e\u7684\u884c\u7ef4\u5ea6\u5bf9\u9f50\uff0c\u6240\u4ee5\u53ef\u4ee5\u76f4\u63a5Broadcast\u5230\u5404\u4e2aGPU\u3002\u6bcf\u5f20\u5361\u5f97\u5230\u5b8c\u6574X\u540e\u548cW\u7684\u5206\u7247X\u76f8\u4e58\uff0c\u5f97\u5230\u5c40\u90e8\u5206\u7247\u8f93\u51faY_partial\uff1b\u6700\u540e\u628a\u6240\u6709Y_partial\u6cbf\u6700\u540e\u4e00\u4e2a\u7ef4\u5ea6\u62fc\u63a5\uff0c\u5168\u90e8\u6c47\u603b\u5373\u53ef\u5f97\u5230\u5b8c\u6574\u8f93\u51faY\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u4e24\u79cd\u5207\u6cd5\u5728Megatron-LM\u7b49\u4e3b\u6d41\u6846\u67b6\u4e2d\u5c31\u662f\u6700\u57fa\u7840\u7684\u64cd\u4f5c\u5355\u5143\u3002LLaMA\u3001Qwen\u3001DeepSeek\u7b49\u5927\u6a21\u578b\u4e2d\u51e0\u4e4e\u6240\u6709\u7684\u8d85\u5927\u77e9\u9635\u4e58\u6cd5\uff0c\u90fd\u662f\u7528\u8fd9\u4e24\u79cd\u5207\u6cd5\u53ca\u5176\u7ec4\u5408\u6765\u5b9e\u73b0\u7684\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">2.2 \u4ee3\u4ef7\uff1a\u901a\u4fe1\u53d8\u5f97\u53c8\u5bc6\u53c8\u7cbe\u7ec6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5f20\u91cf\u5e76\u884c\u6bcf\u4e00\u5c42\u7b97\u5b8c\u90fd\u8981\u901a\u4fe1\u2014\u2014All-Reduce\u6216All-Gather\u2014\u2014\u800c\u4e14\u901a\u4fe1\u7684\u6570\u636e\u91cf\u548c\u5207\u5206\u65b9\u6848\u5bc6\u5207\u76f8\u5173\u3002\u6bcf\u4e24\u5c42\u4e4b\u95f4\u540c\u6b65\u4e00\u6b21\uff0c\u5bfc\u81f4\u901a\u4fe1\u8bbf\u5b58\u91cf\u8fdc\u5927\u4e8e\u6570\u636e\u5e76\u884c\u3002\u800c\u4e14\u56e0\u4e3a\u6d89\u53ca\u8fd8\u539f\u5b8c\u6574\u7684\u8ba1\u7b97\u56fe\uff0cTP\u7684\u9519\u8bef\u5b9e\u73b0\u4f1a\u5bfc\u81f4\u9690\u85cfBug\u5f0f\u7684\u7cbe\u5ea6\u4e0b\u964d\uff0c\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48\u5f20\u91cf\u5e76\u884c\u901a\u5e38\u88ab\u8ba4\u4e3a\u662f\u5b9e\u73b0\u6700\u590d\u6742\u7684\u5927\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5de5\u4e1a\u754c\u7684\u5b9e\u8df5\u662f\uff1a<strong>\u5f20\u91cf\u5e76\u884c\u53ea\u7528\u5728\u5355\u673a\u5185\u90e8\u7684\u9ad8\u901f\u4e92\u8054\uff08\u5982NVLink\u3001NVSwitch\uff09\u4e0a<\/strong>\u3002\u8de8\u673a\u7f51\u7edc\uff08\u5982RoCE\u3001InfiniBand\uff09\u5e26\u5bbd\u8fdc\u4f4e\u4e8e\u673a\u5185\uff0c\u505aTP\u4f1a\u8ba9\u901a\u4fe1\u5f00\u9500\u541e\u566c\u6389\u6240\u6709\u52a0\u901f\u6536\u76ca\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\uff1a\u7528\u201c\u5de5\u5382\u6d41\u6c34\u7ebf\u201d\u4e32\u8d77\u5927\u6a21\u578b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">3.1 \u57fa\u672c\u601d\u60f3\uff1a\u6309\u5c42\u6392\u961f\uff0c\u50cf\u5de5\u5382\u6d41\u6c34\u7ebf\u5206\u914d\u5de5\u4f5c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u6d41\u6c34\u7ebf\u5e76\u884c\u628a\u6a21\u578b\u6309\u5c42\uff08Layer\uff09\u5207\u5206\u5f00\uff0c\u6bcf\u5f20\u5361\u6216\u6bcf\u4e2a\u8bbe\u5907\u8d1f\u8d23\u4e00\u6bb5\u8fde\u7eed\u7684\u201cStage\u201d\uff08\u5c42\u533a\u95f4\uff09<\/strong>\u3002\u6570\u636e\u50cf\u6d41\u6c34\u7ebf\u4e0a\u7684\u96f6\u4ef6\u4e00\u6837\u4f9d\u6b21\u4ece\u7b2c\u4e00\u5f20\u5361\u6d41\u5230\u6700\u540e\u4e00\u5f20\u5361\uff1a\u7b2c\u4e00\u5f20\u5361\u5b8c\u6210\u81ea\u5df1\u7684\u5c42\u7684\u8ba1\u7b97\u540e\uff0c\u628a\u4e2d\u95f4\u7ed3\u679c\u4f20\u7ed9\u4e0b\u4e00\u5f20\u5361\uff1b\u4e0b\u4e00\u5f20\u5361\u63a5\u7740\u7b97\uff0c\u76f4\u5230\u6700\u540e\u4e00\u5c42\u541e\u5410\u51fa\u7ed3\u679c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f46\u539f\u59cb\u6d41\u6c34\u7ebf\u6709\u4e00\u4e2a\u672c\u8d28\u95ee\u9898\uff1a<strong>\u6d41\u6c34\u7ebf\u6c14\u6ce1\uff08Pipeline Bubble\uff09<\/strong>\u3002\u5982\u679c\u628a\u6570\u636e\u4e00\u6b21\u6027\u5b8c\u6574\u8dd1\u5b8c\u524d\u5411\u4f20\u64ad\u518d\u505a\u53cd\u5411\u4f20\u64ad\uff0c\u6bcf\u5f20\u5361\u8ba1\u7b97\u4efb\u52a1\u542f\u52a8\u548c\u7ed3\u675f\u7684\u65f6\u95f4\u5dee\u4f1a\u5bfc\u81f4\u5927\u91cf\u5361\u5904\u5728\u7a7a\u95f2\u7b49\u5f85\u72b6\u6001\uff0c\u5229\u7528\u7387\u6781\u4f4e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e8e\u662f\u7814\u7a76\u8005\u53d1\u73b0\uff0c\u89e3\u51b3\u95ee\u9898\u7684\u5173\u952e\u5c31\u5728\u4e8e\uff1a<strong>\u628a\u4e00\u4e2aBatch\u62c6\u6210\u591a\u4e2aMicro-Batch<\/strong>\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.2 \u4eceGPipe\u52301F1B\uff1a\u6d88\u706d\u6c14\u6ce1\u7684\u6f14\u8fdb\u4e4b\u8def<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GPipe<\/strong>\u662f\u5fae\u8f6f\u57282018\u5e74\u6700\u65e9\u7cfb\u7edf\u63d0\u51faMicro-Batch\u6982\u5ff5\u7684\u5de5\u4f5c\u3002\u5b83\u7684\u8d21\u732e\u5728\u4e8e\u9996\u6b21\u627e\u5230\u4e86\u4e00\u4e2a\u5e72\u51c0\u5229\u843d\u6d88\u706d\u6d41\u6c34\u7ebf\u6c14\u6ce1\u7684\u65b9\u6848\uff1a\u628a\u4e00\u4e2aData Batch\u5207\u6210m\u4e2aMicro-Batch\uff0c\u8ba9\u5b83\u4eec\u7d27\u5bc6\u6392\u961f\u7ecf\u8fc7\u6d41\u6c34\u7ebfStage\uff1b\u6240\u6709Micro-Batch\u7684\u524d\u5411\u4f20\u64ad\u5168\u90e8\u5b8c\u6210\u540e\uff0c\u518d\u7edf\u4e00\u8fdb\u5165\u53cd\u5411\u4f20\u64ad\u3002\u8fd9\u4e2a\u65b9\u6848\u8ba9\u6bcf\u4e2aGPU\u5728\u7edd\u5927\u90e8\u5206\u65f6\u95f4\u5185\u90fd\u6709\u6d3b\u5e72\uff0c\u6c14\u6ce1\u88ab\u538b\u7f29\u5230\u6700\u5c0f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>PipeDream\uff081F1B\u7b56\u7565\uff09<\/strong> \u505a\u4e86\u8fdb\u4e00\u6b65\u4f18\u5316\uff1a\u5b83\u4e0d\u518d\u7b49\u5f85\u6240\u6709Micro-Batch\u7684\u524d\u5411\u4f20\u64ad\u5b8c\u6210\u518d\u7edf\u4e00\u505a\u53cd\u5411\u4f20\u64ad\uff0c\u800c\u662f\u6bcf\u5b8c\u6210\u4e00\u4e2aMicro-Batch\u7684\u524d\u5411\u4f20\u64ad\uff0c\u53ea\u8981\u5bf9\u5e94\u7684\u540e\u5411\u8ba1\u7b97\u6761\u4ef6\u6ee1\u8db3\uff0c\u5c31\u7acb\u523b\u63d2\u5165\u4e00\u6b21\u53cd\u5411\u4f20\u64ad\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u53cd\u5411\u4f20\u64ad\u7684\u8ba1\u7b97\u548c\u524d\u5411\u4f20\u64ad\u5927\u91cf\u91cd\u53e0\uff0c\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u6574\u4f53\u541e\u5410\u91cf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6b64\u5916\uff0c<strong>\u865a\u62df\u6d41\u6c34\u7ebf\u5e76\u884c\uff08VPP\uff09<\/strong> \u7b49\u7b56\u7565\u901a\u8fc7\u5728\u4e00\u4e2aStage\u5185\u90e8\u8fdb\u4e00\u6b65\u5207\u5206\u8ba1\u7b97\u4efb\u52a1\uff0c\u7ee7\u7eed\u7cbe\u7ec6\u5316\u586b\u5145\u6c14\u6ce1\uff0c\u4f7fGPU\u5229\u7528\u7387\u65e0\u9650\u903c\u8fd1100%\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">3.3 \u4f18\u4e0e\u52a3\uff1a\u8de8\u5c42\u901a\u4fe1\u975e\u5e38\u5c11\uff0c\u4f46\u914d\u7f6e\u96be\u5ea6\u9ad8<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d41\u6c34\u7ebf\u5e76\u884c\u7684\u6700\u5927\u4f18\u52bf\u5728\u4e8e\uff1a<strong>\u901a\u4fe1\u53ea\u5728\u76f8\u90bbStage\u4e4b\u95f4\u53d1\u751f<\/strong>\uff0c\u6bcf\u4e00\u4e2aStep\u53ea\u9700\u8981\u4f20\u9012\u4e00\u6b21\u5f20\u91cf\uff0c\u901a\u4fe1\u4f53\u79ef\u6781\u5c0f\uff0c\u5929\u7136\u9002\u5408\u8de8\u673a\u901a\u4fe1\u4e0d\u8db3\u7684\u573a\u666f\u3002\u5f88\u591a\u5343\u4ebf\u4e43\u81f3\u4e07\u4ebf\u7a20\u5bc6\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u90fd\u628a\u6d41\u6c34\u7ebf\u5e76\u884c\u4f5c\u4e3a\u4e3b\u8981\u7684\u8de8\u673a\u5207\u5206\u65b9\u5f0f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6d41\u6c34\u7ebf\u6700\u5927\u7684\u6311\u6218\u662f\u542f\u52a8\u65f6\u7684\u201c\u7a7a\u95f2\u5c3e\u90e8\u201d\u548c\u663e\u5b58\u4e0d\u5bf9\u79f0\u3002\u4e0d\u540cStage\u7684\u663e\u5b58\u5360\u7528\u5929\u7136\u4e0d\u540c\uff1a\u9760\u524d\u7684Stage\u5f80\u5f80\u5b58\u4e86\u66f4\u591a\u7684\u6fc0\u6d3b\u503c\uff0c\u5bfc\u81f4\u6574\u6761\u6d41\u6c34\u7ebf\u7684\u663e\u5b58\u5229\u7528\u7387\u53c2\u5dee\u4e0d\u9f50\u3002\u8d1f\u8f7d\u5747\u8861\u662f\u6d41\u6c34\u7ebf\u8bbe\u7f6e\u4e2d\u6700\u5934\u75bc\u7684\u95ee\u9898\u3002\u6b64\u5916\uff0c\u5982\u679c\u591a\u4e2aStage\u6070\u597d\u843d\u5728\u4e0d\u540c\u7269\u7406\u8282\u70b9\u4e0a\uff0c\u8282\u70b9\u95f4\u901a\u4fe1\u7684\u8d28\u91cf\u4f1a\u4e25\u91cd\u5f71\u54cd\u6d41\u6c34\u7ebf\u6548\u7387\u3002\u5de5\u4e1a\u754c\u901a\u5e38\u7684\u505a\u6cd5\u662f\u4eba\u5de5\u6216\u534a\u81ea\u52a8\u5730\u5c06\u6d41\u6c34\u7ebf\u4e0e\u7f51\u7edc\u62d3\u6251\u7ed1\u5b9a\uff1a\u786e\u4fdd\u8fde\u7eed\u7684Stage\u5c3d\u53ef\u80fd\u653e\u5728\u540c\u4e00\u53f0\u7269\u7406\u670d\u52a1\u5668\u5185\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u30013D\u5e76\u884c\uff1a\u628a\u4e09\u79cd\u65b9\u6cd5\u201c\u62e7\u6210\u4e00\u80a1\u7ef3\u201d<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">4.1 \u4e00\u4e2a\u7edf\u4e00\u7684\u6846\u67b6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5230\u4e862019\u5e74\u540e\uff0c\u5927\u6a21\u578b\u8bad\u7ec3\u9886\u57df\u51dd\u805a\u51fa\u4e00\u4e2a\u5171\u8bc6\uff1a\u5bf9\u4e8e\u8d85\u5927\u6a21\u578b\uff0c\u5355\u9760\u4e00\u79cd\u5e76\u884c\u7b56\u7565\u662f\u4e0d\u591f\u7684\u3002\u4e8e\u662f\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u88ab\u7edf\u4e00\u914d\u7f6e\u5728\u540c\u4e00\u4e2a\u8bad\u7ec3\u4efb\u52a1\u91cc\uff0c\u5f62\u6210\u6240\u8c13<strong>\u201c3D\u5e76\u884c\u201d<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5177\u4f53\u914d\u7f6e\u903b\u8f91\u901a\u5e38\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5f20\u91cf\u5e76\u884c\uff08TP\uff09<\/strong>\uff1a\u7528\u5728<strong>\u5355\u673a\u5185\u90e8<\/strong>\uff0c\u6253\u7834\u8d85\u5927\u77e9\u9635\u4e58\u6cd5\u3002<\/li>\n\n\n\n<li><strong>\u6d41\u6c34\u7ebf\u5e76\u884c\uff08PP\uff09<\/strong>\uff1a\u7528\u5728<strong>\u8de8\u673a\u4e4b\u95f4<\/strong>\uff0c\u628a\u5e9e\u6742\u7684\u5c42\u5e8f\u5217\u62c6\u6210\u591a\u6bb5\u3002<\/li>\n\n\n\n<li><strong>\u6570\u636e\u5e76\u884c\uff08DP\uff09<\/strong>\uff1a\u4f5c\u7528\u5728<strong>\u6574\u4e2a\u96c6\u7fa4\u4e0a<\/strong>\uff0c\u7528\u5197\u4f59\u6570\u636e\u526f\u672c\u63d0\u9ad8\u541e\u5410\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">OpenAI\u7684GPT-3\u3001NVIDIA\u7684Megatron-Turing NLG 530B\uff0c\u4ee5\u53ca\u540e\u6765\u7684LLaMA-70B\/400B\uff0c\u5168\u90e8\u662f\u5728\u8fd9\u4e2a\u6846\u67b6\u4e0a\u8dd1\u51fa\u6765\u7684\u3002DeepSeek-V3\u57283D\u5e76\u884c\u7684\u57fa\u7840\u4e0a\u8fd8\u53e0\u52a0\u4e86\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism\uff09\uff0c\u5f62\u6210\u4e86\u4e94\u7ef4\u5e76\u884c\u7ed3\u6784\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">4.2 \u4e3a\u4ec0\u4e48\u6df7\u5408\u5e76\u884c\u662f\u201c\u521a\u9700\u201d\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u4e00\u5f20\u5361\u8f66\u961f\u7684\u7c7b\u6bd4\u6765\u8bf4\u660e\uff1a\u6570\u636e\u5e76\u884c\u5c31\u50cf\u4e00\u4e2a\u516c\u53f8\u6bcf\u4e00\u8f86\u5361\u8f66\u7684\u578b\u53f7\u548c\u8fd0\u529b\u5b8c\u5168\u76f8\u540c\uff0c\u53ef\u4ee5\u540c\u6b65\u88c5\u5378\u8d27\uff0c\u4f46\u5728\u8d85\u5927\u8d27\u7269\u9762\u524d\u6240\u6709\u8f66\u90fd\u201c\u653e\u4e0d\u4e0b\u201d\u3002\u5f20\u91cf\u5e76\u884c\u76f8\u5f53\u4e8e\u628a\u5355\u4e2a\u8d85\u5927\u8d27\u7269\u5728\u5de5\u5382\u91cc\u5c31\u63d0\u524d\u62c6\u89e3\uff0c\u5206\u522b\u7528\u66f4\u591a\u8f66\u88c5\u3002\u6d41\u6c34\u7ebf\u5e76\u884c\u5219\u5efa\u7acb\u4e86\u4ece\u5305\u88c5\u8f66\u95f4\u5230\u914d\u9001\u4e2d\u5fc3\u7684\u5782\u76f4\u6d41\u6c34\u7ebf\uff0c\u628a\u6b65\u9aa4\u62c6\u5f00\u3002\u4e09\u8005\u534f\u540c\uff0c\u8d27\u7269\u672c\u8eab\u5c31\u4f1a\u5728\u4e0d\u540c\u7684\u7ef4\u5ea6\u4ee5\u4e0d\u540c\u7cbe\u5ea6\u88ab\u62c6\u5206\uff0c\u6700\u7ec8\u6bcf\u6761\u4f20\u9001\u5e26\u4e0a\u7684\u8f66\u90fd\u88c5\u5f97\u4e0b\u3001\u8dd1\u5f97\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u8df5\u8bc1\u660e\uff0c\u8fd9\u79cd\u591a\u7ef4\u5ea6\u534f\u540c\u7684\u6536\u76ca\u5de8\u5927\uff1aDeepSeek-R1\u572832\u8282\u70b9\u96c6\u7fa4\uff08\u6bcf\u8282\u70b98\u5361A100\uff09\u4f7f\u75283D\u5e76\u884c\u52a0ZeRO-3\u4f18\u5316\u540e\uff0c\u76f8\u6bd4\u7eaf\u6570\u636e\u5e76\u884c\u65b9\u6848\uff0c\u5185\u5b58\u5360\u7528\u964d\u4f4e62%\uff0c\u901a\u4fe1\u5f00\u9500\u51cf\u5c1141%\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u524d\u6cbf\u6f14\u8fdb\uff1aDP\u7684\u201c\u63a5\u73ed\u4eba\u201d\u4e0e\u5e76\u884c\u7684\u65b0\u7ef4\u5ea6<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e09\u79cd\u57fa\u7840\u5e76\u884c\u7b56\u7565\u5efa\u7acb\u540e\uff0c\u6574\u4e2a\u9886\u57df\u7684\u524d\u6cbf\u65b9\u5411\u4e00\u76f4\u5728\u6269\u5c55\u65b0\u7684\u7ef4\u5ea6\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u96f6\u5197\u4f59\u6570\u636e\u5e76\u884c\uff08ZeRO\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">ZeRO\u4ece\u6839\u672c\u4e0a\u56de\u7b54\u4e86\u201c\u6570\u636e\u5e76\u884c\u7684\u663e\u5b58\u5197\u4f59\u80fd\u5426\u88ab\u6d88\u706d\u201d\u7684\u95ee\u9898\u3002\u5b83\u628a\u4f20\u7edf\u6570\u636e\u5e76\u884c\u4e2d\u6bcf\u5f20\u5361\u90fd\u8981\u5b8c\u6574\u4fdd\u5b58\u4e09\u4efd\u6570\u636e\u2014\u2014\u6a21\u578b\u53c2\u6570\u3001\u68af\u5ea6\u3001\u4f18\u5316\u5668\u72b6\u6001\u2014\u2014\u7684\u201c\u5168\u5458\u590d\u5236\u201d\u6a21\u5f0f\u62c6\u89e3\u4e3a\u5206\u5e03\u5f0f\u5207\u5206\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>ZeRO-1\uff08\u4f18\u5316\u5668\u72b6\u6001\u5206\u7247\uff09<\/strong>\uff1a\u53ea\u5206\u7247\u4f18\u5316\u5668\u72b6\u6001\u3002<\/li>\n\n\n\n<li><strong>ZeRO-2\uff08\u68af\u5ea6+\u4f18\u5316\u5668\u72b6\u6001\u5206\u7247\uff09<\/strong>\uff1a\u68af\u5ea6\u4e5f\u5206\u7247\u5b58\u653e\u3002<\/li>\n\n\n\n<li><strong>ZeRO-3\uff08\u5168\u90e8\u6a21\u578b\u72b6\u6001\u5206\u7247\uff09<\/strong>\uff1a\u4ece\u524d\u5411\u4f20\u64ad\u4e2d\u8ba1\u7b97\u51fa\u7684\u6bcf\u4e00\u5757\u6a21\u578b\u53c2\u6570\u548c\u6fc0\u6d3b\u503c\uff0c\u6309\u9700\u8981\u4e34\u65f6\u6536\u96c6\uff0c\u7528\u5b8c\u7acb\u5373\u91ca\u653e\u201c\u501f\u7528\u201d\u7684\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">ZeRO\u51e0\u4e4e\u6d88\u9664\u4e86\u6570\u636e\u5e76\u884c\u7684\u5185\u5b58\u5197\u4f59\uff0c\u8ba9\u5355\u5361\u4e0a\u65e0\u6cd5\u627f\u8f7d\u7684\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u591a\u5361\u6570\u636e\u5e76\u884c\u8bad\u7ec3\u5f97\u4ee5\u5b9e\u73b0\u3002\u5176\u5347\u7ea7\u7248\u672c<strong>ZeRO++<\/strong>\u5f15\u5165\u4e86\u91cf\u5316\u901a\u4fe1\u7b49\u6280\u672f\uff0c\u5c06\u901a\u4fe1\u91cf\u964d\u81f31\/4\uff0c\u541e\u5410\u63d0\u5347\u81f3\u539f\u6765\u76842.16\u500d\uff1b<strong>ZeRO-Offload<\/strong> \u7b49\u53d8\u4f53\u5219\u53ef\u4ee5\u5728\u8bad\u7ec3\u4e2d\u52a8\u6001\u5c06\u90e8\u5206\u6570\u636e\u6682\u5b58\u5728CPU\u5185\u5b58\u6216NVMe\u786c\u76d8\u4e0a\uff0c\u8fdb\u4e00\u6b65\u6253\u7834\u7269\u7406\u663e\u5b58\u9650\u5236\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 \u5e8f\u5217\u5e76\u884c\u4e0e\u4e0a\u4e0b\u6587\u5e76\u884c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u968f\u7740\u5927\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u4ece4K\u5411128K\u751a\u81f31M\u6269\u5c55\uff0c\u201cActivation\u6fc0\u6d3b\u503c\u201d\u6210\u4e3a\u6bd4\u53c2\u6570\u66f4\u5938\u5f20\u7684\u663e\u5b58\u5360\u7528\u6765\u6e90\u3002\u6fc0\u6d3b\u663e\u5b58\u548c\u5e8f\u5217\u957f\u5ea6\u5448\u7ebf\u6027\u6b63\u76f8\u5173\uff0c\u8d85\u957f\u5e8f\u5217\u4f1a\u5bfc\u81f4\u5355\u5361\u5373\u4f7f\u88c5\u5f97\u4e0b\u6a21\u578b\u4e5f\u88c5\u4e0d\u4e0b\u4e2d\u95f4\u6fc0\u6d3b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5e8f\u5217\u5e76\u884c\uff08Sequence Parallelism\uff09<\/strong> \u628a\u8f93\u5165\u5e8f\u5217\u6cbf\u5e8f\u5217\u957f\u5ea6\u7ef4\uff08splitting by seq_len\uff09\u5207\u5206\u5230\u591a\u5f20\u5361\u4e0a\u5e76\u884c\u8ba1\u7b97\uff0cLayerNorm\u3001Dropout\u7b49\u9010\u5e8f\u5217\u64cd\u4f5c\u5929\u7136\u53ef\u4ee5\u5206\u7247\u3002\u8fd9\u4e00\u6280\u672f\u4e0e\u5f20\u91cf\u5e76\u884c\u642d\u914d\u65f6\u80fd\u6781\u5927\u964d\u4f4e\u6fc0\u6d3b\u503c\u5185\u5b58\u538b\u529b\uff0c\u6210\u4e3a\u8fd1\u5e74\u8d85\u957f\u5e8f\u5217\u8bad\u7ec3\u7684\u6807\u51c6\u5b9e\u8df5\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4e0a\u4e0b\u6587\u5e76\u884c\uff08Context Parallelism\uff09<\/strong> \u5219\u662f\u628a\u6ce8\u610f\u529b\u5c42Q\u3001K\u3001V\u5207\u957f\u5e8f\u5217\uff0c\u7528Ring Self-Attention\u7b49\u65b9\u5f0f\u907f\u514d\u6784\u5efa\u5b8c\u6574\u7684N\u00d7N\u957fAttention\u77e9\u9635\uff0c\u652f\u6301\u8bad\u7ec3\u767e\u4e07Token\u957f\u5ea6\u7684\u5e8f\u5217\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 \u81ea\u52a8\u5e76\u884c\u4e0e\u81ea\u9002\u5e94\u8c03\u5ea6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u624b\u52a8\u914d\u7f6e\u5e76\u884c\u7b56\u7565\u9700\u8981\u6781\u5176\u4e30\u5bcc\u7684\u5148\u9a8c\u7ecf\u9a8c\uff0c\u4eba\u8111\u5f88\u96be\u7b97\u51faTP=4\u3001PP=8\u3001DP=64\u4e14\u663e\u5b58\u4e0d\u7206\u7684\u6700\u4f18\u7ec4\u5408\u30022025\u5e74\u540e\uff0c\u9886\u57df\u91cc\u51fa\u73b0\u4e86\u4e00\u4e9b\u81ea\u52a8\u5316\u641c\u7d22\u5e76\u884c\u914d\u7f6e\u7684\u5de5\u4f5c\u2014\u2014\u7ed9\u5b9a\u6a21\u578b\u7ed3\u6784\u548c\u786c\u4ef6\u96c6\u7fa4\u4fe1\u606f\uff0c\u81ea\u52a8\u8ba1\u7b97\u51fa\u6700\u4f18\u7684\u6df7\u5408\u5e76\u884c\u7b56\u7565\u53ca\u5176\u8d44\u6e90\u5206\u914d\uff0c\u5728\u8bad\u7ec3\u541e\u5410\u4e0a\u83b7\u5f97\u8fdc\u8d85\u4eba\u5de5\u8bbe\u5b9a\u7684\u6536\u76ca\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001\u4e09\u6761\u8def\u5f84\u7684\u201c\u6838\u5fc3\u5dee\u5f02\u201d<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">\u7ef4\u5ea6<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u6570\u636e\u5e76\u884c\uff08DP\uff09<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u5f20\u91cf\u5e76\u884c\uff08TP\uff09<\/th><th class=\"has-text-align-left\" data-align=\"left\">\u6d41\u6c34\u7ebf\u5e76\u884c\uff08PP\uff09<\/th><\/tr><\/thead><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u5207\u5206\u5bf9\u8c61<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6570\u636e\uff08\u6cbfBatch\u5207\u5206\uff09<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6bcf\u4e00\u4e2a\u6743\u91cd\u5f20\u91cf\uff08\u6309\u7ef4\u5ea6\u5207\u5206\uff09<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6a21\u578b\u5c42\uff08\u6309Stage\u5207\u5206\uff09<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u6bcf\u5361\u62e5\u6709<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u5b8c\u6574\u6a21\u578b\u526f\u672c<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6bcf\u4e2a\u5927\u77e9\u9635\u7684\u5c40\u90e8\u5207\u7247<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u4e00\u4e2a\u8fde\u7eed\u7684\u5c42\u533a\u95f4<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u901a\u4fe1\u5bc6\u96c6\u5ea6<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6bcfStep\u540c\u6b65\uff08All-Reduce\uff09<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6bcf\u5c42\u5185\u591a\u6b21\u901a\u4fe1<\/td><td class=\"has-text-align-left\" data-align=\"left\">Stage\u8fb9\u754c\u6781\u5c11\u91cf\u901a\u4fe1<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u5b9e\u73b0\u590d\u6742\u5ea6<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6781\u4f4e<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6700\u9ad8<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8f83\u9ad8\uff08\u9700\u914d\u7f6e\u6c14\u6ce1\uff09<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u9002\u5408\u7684\u786c\u4ef6\u62d3\u6251<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8de8\u673a\u53ef\u63a5\u53d7<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u5fc5\u987b\u673a\u5185\u9ad8\u901f\u4e92\u8054<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u8de8\u673a\/\u8282\u70b9\u5b8c\u7f8e\u5951\u5408<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u9002\u5408\u7684\u6a21\u578b\u89c4\u6a21<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u5c0f\u6a21\u578b\uff0c\u6216\u5df2\u6709ZeRO\u652f\u63f4<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u771f\u6b63\u91cd\u91cf\u7ea7\u77e9\u9635\u4e58\u6cd5<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u5343\u4ebf\u3001\u4e07\u4ebf\u7a20\u5bc6\u5c42\u6a21\u578b<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u4e3b\u8981\u74f6\u9888<\/strong><\/td><td class=\"has-text-align-left\" data-align=\"left\">\u901a\u4fe1\u3001\u663e\u5b58\u5197\u4f59<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u901a\u4fe1\u5ef6\u8fdf\uff0c\u8ba1\u7b97\u548c\u901a\u4fe1\u5f3a\u8026\u5408<\/td><td class=\"has-text-align-left\" data-align=\"left\">\u6c14\u6ce1\u3001\u8d1f\u8f7d\u4e0d\u5747\u3001\u6fc0\u6d3b\u5e26\u5bbd<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u603b\u7ed3\u4e0e\u5c55\u671b\uff1a3D\u5e76\u884c\u4e4b\u540e\uff0c\u8fd8\u6709\u4ec0\u4e48\u65b0\u6545\u4e8b\uff1f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u4e09\u8005\u6784\u6210\u4e86\u73b0\u4ee3\u5927\u6a21\u578b\u8bad\u7ec3\u5206\u5e03\u5f0f\u7cfb\u7edf\u6700\u7ecf\u5178\u7684\u201c\u4e09\u677f\u65a7\u201d\u3002\u4ece\u65e9\u671f\u7684\u5404\u505a\u5404\u7684\uff0c\u5230\u5982\u4eca\u88ab\u6709\u673a\u878d\u5408\u5230\u7edf\u4e00\u7684\u201c3D\u5e76\u884c\u201d\u6846\u67b6\uff0c\u5b83\u4eec\u4e0d\u4ec5\u89e3\u51b3\u4e86\u7b97\u529b\u89c4\u6a21\u5316\u7684\u95ee\u9898\uff0c\u66f4\u53cd\u8fc7\u6765\u6df1\u523b\u5f71\u54cd\u4e86\u5927\u6a21\u578b\u7684\u8bbe\u8ba1\u54f2\u5b66\u2014\u2014MoE\u7684\u51fa\u73b0\u5c31\u662f\u5728\u7b97\u6cd5\u5c42\u9762\u4e3b\u52a8\u56de\u7b54\u4e86\u4e00\u4e2a\u95ee\u9898\uff1a\u201c\u5982\u4f55\u5728\u4e00\u4e2a\u5206\u5e03\u5f0f\u7684\u4e16\u754c\u5b9e\u73b0\u66f4\u9ad8\u7684\u8ba1\u7b97\u4e13\u95e8\u5316\u3002\u201d<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c55\u671b\u672a\u6765\uff0c\u591a\u4e2a\u524d\u6cbf\u65b9\u5411\u6b63\u5728\u63a8\u8fdb\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u81ea\u52a8\u641c\u7d22\u6700\u4f18\u6df7\u5408\u5e76\u884c\u7b56\u7565<\/strong>\uff1a\u4ece\u4eba\u8111\u8bbe\u7f6e\u5230\u7cfb\u7edf\u4f18\u5316\u7684\u8303\u5f0f\u8f6c\u79fb\uff0cGalvatron\u7b49\u6846\u67b6\u5df2\u5b9e\u73b0\u52a8\u6001\u9009\u62e9\u5e76\u884c\u7b56\u7565\u3002<\/li>\n\n\n\n<li><strong>\u7b97\u5b58\u5206\u79bb\u67b6\u6784\u4e0e\u5f02\u6784\u5185\u5b58\u4ea4\u6362\uff08ZeRO-Offload\u4e0eSSD\/NVMe\u52a0\u901f\uff09<\/strong> \uff1a\u8fdb\u4e00\u6b65\u628a\u6a21\u578b\u72b6\u6001\u6269\u5c55\u5230CPU\u751a\u81f3SSD\u4e0a\uff0c\u7528\u8ba1\u7b97\u6362\u663e\u5b58\u3002<\/li>\n\n\n\n<li><strong>\u4e13\u5bb6\u5e76\u884c\u3001\u56fe\u50cf\u5e76\u884c\u3001\u591a\u6a21\u6001\u7684\u5927\u89c4\u6a21\u5e76\u884c\u8bad\u7ec3<\/strong>\uff1a\u5728MoE\u548c\u89c6\u89c9\u6a21\u578b\u4e2d\u589e\u52a0\u4e86\u65b0\u7684\u5e76\u884c\u7ef4\u5ea6\uff0c\u4e0e3D\u5e76\u884c\u5e76\u7acb\u5f62\u62104D\/5D\u6df7\u5408\u5e76\u884c\u3002<\/li>\n\n\n\n<li><strong>\u5f02\u6784\u82af\u7247\u7684\u5168\u7403\u5316\u8c03\u5ea6<\/strong>\uff1a\u5c06\u540c\u4e00\u8bad\u7ec3\u4efb\u52a1\u8de8\u4e0d\u540c\u7b97\u529b\u3001\u4e0d\u540c\u4e92\u8054\u901f\u5ea6\u7684GPU\u534f\u540c\u8c03\u5ea6\u6210\u6700\u4f18\u5206\u5e03\u6001\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e32\u8054\u7cfb\u5217\u77e5\u8bc6\u603b\u7ed3<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6211\u4eec\u76ee\u524d\u8d70\u8fc7\u7684\u8def\u7ebf\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Tokenizer \u2192 Embedding \u2192 RoPE<\/strong><\/li>\n\n\n\n<li><strong>Attention\uff08MHA\/MQA\/GQA\uff09 \u2192 FFN\uff08SwiGLU\uff09<\/strong><\/li>\n\n\n\n<li><strong>Pre-Norm \/ Post-Norm \/ \u6b8b\u5dee\u8fde\u63a5<\/strong><\/li>\n\n\n\n<li><strong>MoE\u67b6\u6784\uff08\u4e13\u5bb6\u4e0e\u8def\u7531\u673a\u5236\uff09<\/strong><\/li>\n\n\n\n<li><strong>KV Cache\u4e0ePagedAttention<\/strong><\/li>\n\n\n\n<li><strong>\u6570\u636e\u5e76\u884c\u3001\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\uff08\u672c\u6587<\/strong>)<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u5199\u5728\u524d\u9762 \u76ee\u524d\u4e3a\u6b62\u7684\u6587\u7ae0\u90fd\u5728\u804a\u4e00\u4ef6\u4e8b\uff1a\u5355\u4e2a\u6a21\u578b\u5185\u90e8\u7684\u201c\u5fae\u89c2\u4e16\u754c\u201d \u2014\u2014Embedding\u600e\u4e48\u5de5\u4f5c\u3001Attent [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":148,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[48,37,47,46],"class_list":["post-147","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-data-parallel","tag-llm","tag-pipeline-parallel","tag-tensor-parallel"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/147","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=147"}],"version-history":[{"count":2,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/147\/revisions"}],"predecessor-version":[{"id":150,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/147\/revisions\/150"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/148"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=147"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=147"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=147"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}