{"id":480,"date":"2025-09-15T18:21:04","date_gmt":"2025-09-15T10:21:04","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=480"},"modified":"2025-11-09T04:37:29","modified_gmt":"2025-11-08T20:37:29","slug":"bytescale-communication-efficient-scaling-of-llm-training-with-a-2048k-context-length-on-16384-gpus","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=480","title":{"rendered":"ByteScale: Communication-Efficient Scaling of LLM Training with a 2048K Context Length on 16384 GPUs"},"content":{"rendered":"\n

Hao Ge (Peking University); Junda Feng, Qi Huang (ByteDance Inc.); Fangcheng Fu (Shanghai Jiao Tong University); Xiaonan Nie, Lei Zuo, Haibin Lin (ByteDance Inc.); Bin Cui (Peking University); Xin Liu (ByteDance Inc.)<\/em><\/p>\n\n\n\n

ByteScale: Communication-Efficient Scaling of LLM Training with a 2048K Context Length on 16384 GPUs<\/a><\/p>\n\n\n\n

\u62a5\u9053\u4eba\uff1a\u65b9\u660e\u4fca\uff08SNG \u4e09\u5e74\u7ea7\u7855\u58eb\u751f\uff09<\/em><\/p>\n\n\n\n

\"ByteScale:<\/figure>\n\n\n\n
\n\n\n\n

\u4ecb\u7ecd<\/strong><\/h2>\n\n\n\n

\u672c\u6587\u91cd\u70b9\u7814\u7a76\u968f\u7740\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u5230\u6570\u5341\u4e07\u751a\u81f3\u6570\u767e\u4e07token\u7ea7\u522b\u65f6\uff0c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u8bad\u7ec3\u4e2d\u51fa\u73b0\u7684\u901a\u4fe1\u74f6\u9888\u95ee\u9898\u3002\u73b0\u6709\u7684\u65b9\u6cd5\u8981\u4e48\u4f9d\u8d56\u6570\u636e\u5e76\u884c<\/mark>\uff0c\u8981\u4e48\u4f9d\u8d56\u4e0a\u4e0b\u6587\u5e76\u884c<\/mark>\uff0c\u4f46\u4e24\u8005\u90fd\u5b58\u5728\u4e25\u91cd\u7684\u6548\u7387\u95ee\u9898\u3002\u6570\u636e\u5e76\u884c\u5728\u5e8f\u5217\u957f\u5ea6\u589e\u52a0\u65f6\u901a\u4fe1\u6210\u672c\u6025\u5267\u4e0a\u5347\uff0c\u800c\u4e0a\u4e0b\u6587\u5e76\u884c\u5728\u5904\u7406\u957f\u77ed\u5e8f\u5217\u6df7\u5408\u65f6\u4f1a\u9020\u6210\u8d44\u6e90\u6d6a\u8d39\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u95ee\u9898\u7684\u6839\u6e90\u5728\u4e8e\u5f53\u524d\u5e76\u884c\u5316\u7b56\u7565<\/mark>\u7684\u50f5\u5316\uff0c\u5373\u5bf9\u6240\u6709\u5e8f\u5217\u91c7\u7528\u76f8\u540c\u7684\u5206\u533a\u65b9\u6848\uff0c\u5bfc\u81f4\u8bbe\u5907\u5229\u7528\u7387\u4f4e\u548c\u8de8\u8bbe\u5907\u7b49\u5f85\u65f6\u95f4\u8fc7\u957f\u3002\u8fd9\u4e00\u5206\u6790\u4e3aByteCycle\u7684\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u52a8\u673a\u3002<\/p>\n\n\n\n

\"\u56fe\u7247\"<\/figure>\n\n\n\n

\u6838\u5fc3\u601d\u60f3\u4e0e\u8d21\u732e<\/strong><\/h2>\n\n\n\n
\"ByteScale:<\/figure>\n\n\n\n

ByteCycle\u7684\u6838\u5fc3\u601d\u60f3\u662f\u63d0\u51fa\u4e00\u79cd\u65b0\u7684\u6df7\u5408\u5e76\u884c\u8303\u5f0f\uff0c\u5c06\u6570\u636e\u5e76\u884c\u548c\u4e0a\u4e0b\u6587\u5e76\u884c\u878d\u5408\u5230\u4e00\u4e2a\u7edf\u4e00\u7684\u901a\u4fe1\u6a21\u578b\u4e2d\uff0c\u5e76\u6839\u636e\u5e8f\u5217\u957f\u5ea6\u52a8\u6001\u8c03\u6574\u3002\u5176\u8d21\u732e\u662f\u591a\u65b9\u9762\u7684\uff1a\u9996\u5148\uff0c\u5b83\u5f15\u5165\u4e86\u4e00\u79cd\u9009\u62e9\u6027\u5378\u8f7d\u673a\u5236<\/mark>\uff0c\u4f7f\u5f97\u77ed\u5e8f\u5217\u53ef\u4ee5\u5b8c\u5168\u5728\u5355\u4e2aGPU\u4e0a\u5904\u7406\uff0c\u65e0\u9700\u4e0d\u5fc5\u8981\u7684\u8de8\u8bbe\u5907\u540c\u6b65\uff0c\u800c\u4ec5\u5bf9\u957f\u5e8f\u5217\u8fdb\u884c\u5206\u5e03\u5f0f\u5904\u7406\uff1b\u5176\u6b21\uff0c\u5b83\u5f00\u53d1\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u5fae\u6279\u91cf\u8c03\u5ea6<\/mark>\u65b9\u6cd5\uff0c\u786e\u4fdd\u5373\u4f7f\u5728\u5f02\u6784\u5e8f\u5217\u957f\u5ea6\u4e0b\u4e5f\u80fd\u5b9e\u73b0\u8d1f\u8f7d\u5747\u8861\uff0c\u4ece\u800c\u6700\u5c0f\u5316\u6d41\u6c34\u7ebf\u6c14\u6ce1\u548c\u7a7a\u95f2\u65f6\u95f4\uff1b\u7b2c\u4e09\uff0c\u5b83\u63d0\u4f9b\u4e86\u4e00\u79cd\u8f7b\u91cf\u7ea7\u7684\u6570\u636e\u611f\u77e5\u901a\u4fe1\u4f18\u5316\u5668\uff0c\u4f7f\u8be5\u65b9\u6848\u80fd\u591f\u6269\u5c55\u5230\u6570\u4e07\u4e2aGPU\u3002\u8fd9\u4e9b\u8d21\u732e\u5e76\u975e\u5b64\u7acb\u7684\u6280\u5de7\uff0c\u800c\u662f\u4ee3\u8868\u4e86\u5206\u5e03\u5f0f\u8bad\u7ec3\u8bbe\u8ba1\u7684\u4e00\u79cd\u8303\u5f0f\u8f6c\u53d8\uff0c\u91cd\u65b0\u5b9a\u4e49\u4e86\u5728\u6781\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u4e2d\u5e94\u5982\u4f55\u534f\u8c03\u5e76\u884c\u5316\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u8bc4\u4f30<\/strong><\/h2>\n\n\n\n

\u4f5c\u8005\u5728\u753116,384\u4e2aGPU\u7ec4\u6210\u7684\u5927\u89c4\u6a21\u96c6\u7fa4\u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u6db5\u76d6\u4e86\u4ece7B\u523070B\u53c2\u6570\u7684LLaMA\u7b49\u7a20\u5bc6\u6a21\u578b\uff0c\u4ee5\u53caMistral 8\u00d77B\u548c8\u00d722B\u7b49\u7a00\u758f\u6df7\u5408\u4e13\u5bb6\uff08MoE\uff09\u6a21\u578b\uff0c\u4e0a\u4e0b\u6587\u957f\u5ea6\u6700\u9ad8\u6269\u5c55\u52302,048K\u3002\u5728GitHub\u6570\u636e\u96c6\u4e0a\uff0c\u57fa\u7ebf\u7cfb\u7edf\u5728\u4e0a\u4e0b\u6587\u957f\u5ea6\u52a0\u500d\u65f6\u541e\u5410\u91cf\u51e0\u4e4e\u51cf\u534a\uff0c\u800cByteCycle\u7684\u541e\u5410\u91cf\u4ec5\u4e0b\u964d1.08\u500d\uff0c\u663e\u793a\u51fa\u5728\u957f\u4e0a\u4e0b\u6587\u4e0b\u7684\u5353\u8d8a\u7a33\u5b9a\u6027\u3002\u5728Byted\u6570\u636e\u96c6\u4e0a\uff0c\u5c3d\u7ba1\u5e8f\u5217\u5206\u5e03\u5e26\u6765\u4e86\u989d\u5916\u6311\u6218\uff0cByteCycle\u4ecd\u7136\u663e\u8457\u4f18\u4e8e\u73b0\u6709\u7cfb\u7edf\u3002\u62a5\u544a\u7684\u6700\u5927\u52a0\u901f\u6bd4\u8fbe\u5230\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u76847.89\u500d\uff0c\u8fdc\u8d85\u73b0\u6709\u6700\u5148\u8fdb\u57fa\u7ebf\uff0c\u4e3a\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u8bbe\u7acb\u4e86\u65b0\u7684\u6807\u6746\u3002\u8fdb\u4e00\u6b65\u7684\u6d88\u878d\u7814\u7a76\u8bc1\u5b9e\uff0c\u9009\u62e9\u6027\u5378\u8f7d\u3001\u5747\u8861\u8c03\u5ea6\u548c\u8fdc\u7a0b\u6570\u636e\u52a0\u8f7d\u5404\u81ea\u90fd\u505a\u51fa\u4e86\u53ef\u8861\u91cf\u7684\u8d21\u732e\uff0c\u786e\u4fdd\u6027\u80fd\u63d0\u5347\u662f\u6574\u4f53\u8bbe\u8ba1\u7684\u76f4\u63a5\u7ed3\u679c\uff0c\u800c\u975e\u67d0\u4e00\u4f18\u5316\u7684\u5076\u7136\u4ea7\u7269\u3002\u603b\u4e4b\uff0c\u8bc4\u4f30\u7ed3\u679c\u8868\u660eByteCycle\u662f\u4e00\u4e2a\u7a33\u5065\u4e14\u53ef\u6269\u5c55\u7684\u6846\u67b6\uff0c\u4ece\u6839\u672c\u4e0a\u6539\u53d8\u4e86\u957f\u4e0a\u4e0b\u6587LLM\u8bad\u7ec3\u7684\u53ef\u80fd\u6027\u3002<\/p>\n\n\n\n

\u95ee\u7b54\u73af\u8282<\/strong><\/h2>\n\n\n\n

Q1\uff1a \u4f60\u4eec\u63d0\u51fa\u4e86\u6df7\u5408\u6570\u636e\u5e76\u884c\u548c\u52a8\u6001\u9009\u62e9\u6027\u5378\u8f7d\uff0c\u5e76\u572816,000\u4e2a\u540c\u6784CPU\u4e0a\u7814\u7a76\u4e86\u6027\u80fd\u3002\u90a3\u5bf9\u4e8e\u5f02\u6784GPU\u96c6\u7fa4\uff08\u4f8b\u5982\u663e\u5b58\u5927\u5c0f\u3001\u4e92\u8fde\u901f\u5ea6\u4e0d\u540c\uff09\uff0c\u6027\u80fd\u654f\u611f\u6027\u6709\u4ec0\u4e48\u7ecf\u9a8c\u6216\u770b\u6cd5\u5417\uff1f<\/p>\n\n\n\n

A1:  \u6027\u80fd\u786e\u5b9e\u4f1a\u53d7\u5230\u5f71\u54cd\uff0c\u56e0\u4e3a\u901a\u4fe1\u5e26\u5bbd\u548c\u8ba1\u7b97\u80fd\u529b\u5b58\u5728\u5dee\u5f02\u3002\u5173\u952e\u5728\u4e8e\u5982\u4f55\u66f4\u597d\u5730\u5b9e\u73b0\u901a\u4fe1\u548c\u8ba1\u7b97\u7684\u91cd\u53e0\u3002\u5982\u679c\u4f7f\u7528\u4e0d\u540c\u7c7b\u578b\u7684GPU\uff0c\u53ef\u80fd\u9700\u8981\u8c03\u6574\u6bcf\u4e2arank\u5904\u7406\u7684token\u6570\u91cf\uff0c\u4ee5\u66f4\u597d\u5730\u5b9e\u73b0\u91cd\u53e0\u3002<\/p>\n\n\n\n

Q2:  \u4f60\u4eec\u7684\u5de5\u4f5c\u4e3b\u8981\u9488\u5bf9LLM\u8bad\u7ec3\u5e76\u4f18\u5316\u4e86\u901a\u4fe1\u4e0e\u8ba1\u7b97\u7684\u91cd\u53e0\uff0c\u90a3\u4e48\u7c7b\u4f3c\u7684\u601d\u8def\u80fd\u5426\u5e94\u7528\u5230LLM\u63a8\u7406\u9636\u6bb5\uff1f<\/p>\n\n\n\n

A2: \u53ef\u4ee5\u7684\uff0c\u7c7b\u4f3c\u7684\u601d\u8def\u540c\u6837\u9002\u7528\u4e8e\u63a8\u7406\uff0c\u7279\u522b\u662f\u957f\u4e0a\u4e0b\u6587\u63a8\u7406\u3002\u5728\u63a8\u7406\u65f6\u4e5f\u6709\u5982\u4e0a\u4e0b\u6587\u5e76\u884c\u3001\u5e8f\u5217\u5e76\u884c\u7b49\u65b9\u6cd5\uff0c\u4f46\u9700\u8981\u7279\u522b\u6ce8\u610f\u957f\u5e8f\u5217\u548c\u77ed\u5e8f\u5217\u4e4b\u95f4\u7684\u5e73\u8861\u3002<\/p>\n\n\n\n

\u4e2a\u4eba\u89c2\u70b9<\/strong><\/h2>\n\n\n\n

\u8fd9\u7bc7\u6587\u7ae0\u7684\u6838\u5fc3\u601d\u60f3\u2014\u2014\u6839\u636e\u5e8f\u5217\u5b9e\u9645\u957f\u5ea6\u52a8\u6001\u7ec4\u88c5\u901a\u4fe1\u7ec4<\/mark>\uff0c\u800c\u975e\u9759\u6001\u7f51\u683c\uff0c\u7ed9\u6211\u7559\u4e0b\u4e86\u6df1\u523b\u5370\u8c61\u3002\u8fd9\u4e0d\u4ec5\u4ec5\u662f\u4e00\u9879\u4f18\u5316\u2014\u2014\u66f4\u662f\u4e00\u79cd\u8303\u5f0f\u8f6c\u53d8\uff0c\u5b83\u627f\u8ba4\u4e86\u771f\u5b9e\u6570\u636e\u4e2d\u56fa\u6709\u7684\u5f02\u6784\u6027\u3002\u6211\u8ba4\u4e3a\u5206\u5e03\u5f0f\u8bad\u7ec3\u7cfb\u7edf\u7684\u672a\u6765\u5e94\u8be5\u662f\u50cf\u6587\u7ae0\u63d0\u51fa\u7684\u90a3\u6837\uff1a\u81ea\u9002\u5e94\u3001\u6570\u636e\u611f\u77e5\u4e14\u6d41\u52a8\u7684\u3002\u7136\u800c\uff0c\u867d\u7136\u5b9e\u9a8c\u8fbe\u5230\u4e867.89\u500d\u7684\u663e\u8457\u52a0\u901f\u6bd4\u63d0\u5347\uff0c\u4f46\u662f\u6211\u8ba4\u4e3a\u6211\u4eec\u54c8\u8fd8\u9700\u8981\u5173\u6ce8\u5176\u52a8\u6001\u8c03\u5ea6\u673a\u5236\u7684\u7cfb\u7edf\u6027\u5f00\u9500\u3002\u8bba\u6587\u51fa\u8272\u5730\u4f18\u5316\u4e86GPU\u65f6\u95f4\uff0c\u4f46\u5e76\u6ca1\u6709\u8ba8\u8bba\u5176CPU\u7aef\u7684\u6210\u672c\uff1a\u5305\u62ec\u6301\u7eed\u5206\u6790\u5e8f\u5217\u3001\u4e3a\u6bcf\u4e00\u4e2a\u5fae\u6279\u6b21\u7528\u516c\u5f0f(3)\u6c42\u89e3\u4f18\u5316\u95ee\u9898\u3001\u4ee5\u53ca\u901a\u8fc7Ray\u534f\u8c03\u8fdc\u7a0b\u6570\u636e\u52a0\u8f7d\u5668\u3002\u5728\u4ed6\u4eec16K GPU\u7684\u5de8\u5927\u89c4\u6a21\u4e0b\uff0c\u8fd9\u79cd\u5143\u6570\u636e\u7ba1\u7406\u548c\u8c03\u5ea6\u51b3\u7b56\u7684\u5f00\u9500\u53ef\u80fd\u4f1a\u6210\u4e3a\u4e00\u4e2a\u4e0d\u53ef\u5ffd\u89c6\u7684\u74f6\u9888\uff0c\u53ef\u80fd\u8fdb\u4e00\u6b65\u9650\u5236\u53ef\u6269\u5c55\u6027\uff0c\u6216\u4f7f\u7cfb\u7edf\u5bf9\u4e3b\u673a\u6027\u80fd\u5f02\u5e38\u654f\u611f\u3002<\/p>\n\n\n\n

\u6587\u7ae0\u5730\u5740\uff1a https:\/\/mp.weixin.qq.com\/mp\/appmsgalbum?action=getalbum&__biz=Mzg2Mjg0ODcyNQ==&scene=1&album_id=4157934849875214355<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

Hao Ge (Peking University); Junda Feng, Qi Huang (ByteDance Inc.); Fangcheng Fu (Shanghai Jiao Tong University); Xiaonan Nie, Lei Zuo, Haibin Lin (ByteDance Inc.); Bin Cui (Peking University); Xin Liu (ByteDance Inc.) ByteScale: Communicati … \u7ee7\u7eed\u9605\u8bfb →<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":538,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,23],"tags":[],"class_list":["post-480","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng","category-23"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/480","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=480"}],"version-history":[{"count":4,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/480\/revisions"}],"predecessor-version":[{"id":540,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/480\/revisions\/540"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/538"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=480"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=480"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=480"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}