{"id":1349,"date":"2026-05-18T11:48:45","date_gmt":"2026-05-18T03:48:45","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1349"},"modified":"2026-05-18T11:48:45","modified_gmt":"2026-05-18T03:48:45","slug":"megascale-infer-efficient-mixture-of-experts-model-serving-with-disaggregated-expert-parallelism","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1349","title":{"rendered":"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism"},"content":{"rendered":"\n<p>SIGCOMM 2025<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>1. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u75db\u70b9\uff1aMoE \u63a8\u7406\u7684\u6548\u7387\u74f6\u9888<\/strong><\/h4>\n\n\n\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u5165\u201c\u4e07\u4ebf\u53c2\u6570\u65f6\u4ee3\u201d\uff0c\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Mixture-of-Experts, MoE\uff09\u56e0\u5176\u80fd\u5728\u589e\u52a0\u53c2\u6570\u91cf\u7684\u540c\u65f6\u4fdd\u6301\u8f83\u4f4e\u7684\u6fc0\u6d3b\u8ba1\u7b97\u91cf\uff0c\u6210\u4e3a\u4e86\u6784\u5efa\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u7684\u9996\u9009\u67b6\u6784\u3002\u7136\u800c\uff0c\u5728\u5b9e\u9645\u63a8\u7406\u573a\u666f\u4e2d\uff0cMoE \u6a21\u578b\u9762\u4e34\u7740\u4e25\u5cfb\u7684\u6311\u6218\u3002\u4f20\u7edf\u7684\u63a8\u7406\u7cfb\u7edf\u901a\u5e38\u5c06\u6ce8\u610f\u529b\u673a\u5236\uff08Attention\uff09\u548c\u5168\u8fde\u63a5\u5c42\uff08FFN\/Experts\uff09\u7ed1\u5b9a\u5728\u540c\u4e00\u5757 GPU \u4e0a\u5904\u7406\uff0c\u8fd9\u79cd\u201c\u8026\u5408\u67b6\u6784\u201d\u5bfc\u81f4\u4e86\u4e25\u91cd\u7684\u8d44\u6e90\u9519\u914d\uff1aAttention \u6a21\u5757\u901a\u5e38\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u7684\uff0c\u800c MoE \u4e2d\u7684\u4e13\u5bb6\u5c42\uff08FFN\uff09\u7531\u4e8e\u5176\u7a00\u758f\u6fc0\u6d3b\u7684\u7279\u6027\uff0c\u8f6c\u53d8\u4e3a\u663e\u5b58\u5e26\u5bbd\u5bc6\u96c6\u578b\uff08Memory-bound\uff09\uff0c\u5bfc\u81f4 GPU \u5229\u7528\u7387\u4f4e\u4e0b\uff0c\u8fd0\u8425\u6210\u672c\u9ad8\u6602\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0cMoE \u6a21\u578b\u7684\u4e13\u5bb6\u5e76\u884c\uff08Expert Parallelism\uff09\u5728\u5927\u89c4\u6a21\u90e8\u7f72\u65f6\u4f1a\u4ea7\u751f\u5de8\u5927\u7684\u5168\u5f00\u9500\u901a\u4fe1\uff08All-to-All communication\uff09\u3002\u5728\u4f20\u7edf\u7684\u8026\u5408\u5e76\u884c\u7b56\u7565\u4e0b\uff0c\u8fd9\u79cd\u901a\u4fe1\u5f00\u9500\u5f80\u5f80\u4e0e\u8ba1\u7b97\u91cd\u53e0\u4e0d\u8db3\uff0c\u8fdb\u4e00\u6b65\u62c9\u957f\u4e86\u63a8\u7406\u5ef6\u8fdf\u3002\u5bf9\u4e8e\u50cf DeepSeek-V3 \u8fd9\u6837\u62e5\u6709\u6570\u767e\u4e2a\u4e13\u5bb6\u7684\u6a21\u578b\uff0c\u5982\u4f55\u5728\u4fdd\u8bc1\u541e\u5410\u91cf\u7684\u540c\u65f6\u964d\u4f4e\u5ef6\u8fdf\uff0c\u5e76\u89e3\u51b3\u4e13\u5bb6\u8d1f\u8f7d\u4e0d\u5747\uff08Expert Unbalance\uff09\u5e26\u6765\u7684\u957f\u5c3e\u6548\u5e94\uff0c\u662f\u5f53\u524d\u5206\u5e03\u5f0f\u63a8\u7406\u9886\u57df\u4e9f\u5f85\u89e3\u51b3\u7684\u5e95\u5c42\u96be\u9898\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"317\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-1-1024x317.png\"  class=\"wp-image-1350\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-1-1024x317.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-1-300x93.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-1-768x238.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/1-1.png 1143w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe\" \/><\/figure>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"597\" height=\"339\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2-1.png\"  class=\"wp-image-1351\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2-1.png 597w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/2-1-300x170.png 300w\" sizes=\"auto, (max-width: 597px) 100vw, 597px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe1\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe1\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>2. \u6838\u5fc3\u521b\u65b0\uff1a\u4e13\u5bb6\u5e76\u884c\u4e0e\u6ce8\u610f\u529b\u673a\u5236\u7684\u89e3\u8026\uff08Disaggregation\uff09<\/strong><\/h4>\n\n\n\n<p>MegaScale-Infer \u63d0\u51fa\u4e86\u4e00\u4e2a\u9769\u547d\u6027\u7684\u7cfb\u7edf\u6846\u67b6\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u662f\u201c\u6ce8\u610f\u529b\u4e0e\u4e13\u5bb6\u7684\u89e3\u8026\u90e8\u7f72\u201d\u3002\u8be5\u7cfb\u7edf\u5c06\u6a21\u578b\u6bcf\u4e00\u5c42\u4e2d\u7684 Attention \u6a21\u5757\u548c FFN \u4e13\u5bb6\u6a21\u5757\u5265\u79bb\uff0c\u5206\u522b\u90e8\u7f72\u5728\u4e0d\u540c\u7684 GPU \u8d44\u6e90\u6c60\u4e2d\u3002\u8fd9\u79cd\u8bbe\u8ba1\u5141\u8bb8\u9488\u5bf9\u4e24\u7c7b\u6a21\u5757\u7684\u4e0d\u540c\u7279\u6027\u8fdb\u884c\u72ec\u7acb\u7f29\u653e\uff1aAttention \u6c60\u53ef\u4ee5\u914d\u7f6e\u8ba1\u7b97\u80fd\u529b\u66f4\u5f3a\u7684\u786c\u4ef6\u5e76\u91c7\u7528\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallelism\uff09\uff0c\u800c\u4e13\u5bb6\u6c60\u5219\u53ef\u4ee5\u90e8\u7f72\u5728\u663e\u5b58\u5bb9\u91cf\u5927\u3001\u5e26\u5bbd\u9ad8\u7684\u786c\u4ef6\u4e0a\uff0c\u91c7\u7528\u66f4\u5927\u89c4\u6a21\u7684\u4e13\u5bb6\u5e76\u884c\uff08EP\uff09\u3002\u901a\u8fc7\u8fd9\u79cd\u5f02\u6784\u90e8\u7f72\uff0c\u7cfb\u7edf\u80fd\u591f\u5b9e\u73b0\u8ba1\u7b97\u8d44\u6e90\u4e0e\u6a21\u578b\u9700\u6c42\u7684\u7cbe\u51c6\u5339\u914d\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"576\" height=\"360\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3-1.png\"  class=\"wp-image-1352\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3-1.png 576w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/3-1-300x188.png 300w\" sizes=\"auto, (max-width: 576px) 100vw, 576px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe2\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe2\" \/><\/figure>\n<\/div>\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u89e3\u8026\u5e26\u6765\u7684\u6a21\u5757\u95f4\u901a\u4fe1\u538b\u529b\uff0cMegaScale-Infer \u5f15\u5165\u4e86\u201c\u4e52\u4e53\u6d41\u6c34\u7ebf\u5e76\u884c\u201d\uff08Ping-Pong Pipeline Parallelism\uff09\u3002\u7cfb\u7edf\u5c06\u4e00\u4e2a\u8bf7\u6c42\u6279\u6b21\uff08Batch\uff09\u5212\u5206\u4e3a\u591a\u4e2a\u5fae\u6279\u6b21\uff08Micro-batches\uff09\uff0c\u5e76\u5728 Attention \u6c60\u548c\u4e13\u5bb6\u6c60\u4e4b\u95f4\u8fdb\u884c\u7a7f\u68ad\u5f0f\u63a8\u7406\u3002\u5f53\u7b2c\u4e00\u7ec4\u5fae\u6279\u6b21\u5728\u4e13\u5bb6\u6c60\u8fdb\u884c\u4e13\u5bb6\u8ba1\u7b97\u65f6\uff0c\u7b2c\u4e8c\u7ec4\u5fae\u6279\u6b21\u53ef\u4ee5\u5728 Attention \u6c60\u8fdb\u884c\u9884\u5904\u7406\u3002\u8fd9\u79cd\u8bbe\u8ba1\u6781\u5927\u5730\u91cd\u53e0\u4e86\u8de8\u6c60\u901a\u4fe1\u4e0e\u5185\u90e8\u8ba1\u7b97\u7684\u65f6\u95f4\uff0c\u786e\u4fdd\u4e86\u6d41\u6c34\u7ebf\u7684\u9ad8\u6548\u8fd0\u8f6c\u3002\u540c\u65f6\uff0c\u9488\u5bf9 MoE \u7684\u7a00\u758f\u6027\uff0c\u7cfb\u7edf\u8bbe\u8ba1\u4e86\u52a8\u6001\u8c03\u5ea6\u673a\u5236\uff0c\u786e\u4fdd\u8bf7\u6c42\u80fd\u591f\u5b9e\u65f6\u5206\u53d1\u5230\u7a7a\u95f2\u7684\u4e13\u5bb6\u5b9e\u4f8b\u4e0a\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"585\" height=\"324\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4-1.png\"  class=\"wp-image-1353\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4-1.png 585w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/4-1-300x166.png 300w\" sizes=\"auto, (max-width: 585px) 100vw, 585px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe3\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe3\" \/><\/figure>\n<\/div>\n\n\n<h4 class=\"wp-block-heading\"><strong>3. \u7cfb\u7edf\u4f18\u5316\uff1a\u52a8\u6001\u8d1f\u8f7d\u5747\u8861\u4e0e\u901a\u4fe1\u9690\u85cf<\/strong><\/h4>\n\n\n\n<p>\u5728\u89e3\u8026\u67b6\u6784\u7684\u57fa\u7840\u4e0a\uff0cMegaScale-Infer \u8fdb\u4e00\u6b65\u9488\u5bf9\u4e13\u5bb6\u8d1f\u8f7d\u4e0d\u5747\u8861\u95ee\u9898\u8fdb\u884c\u4e86\u6df1\u5ea6\u4f18\u5316\u3002\u5728 MoE \u63a8\u7406\u4e2d\uff0c\u67d0\u4e9b\u201c\u70ed\u95e8\u4e13\u5bb6\u201d\u53ef\u80fd\u4f1a\u88ab\u9891\u7e41\u8c03\u7528\uff0c\u800c\u5176\u4ed6\u4e13\u5bb6\u5219\u76f8\u5bf9\u95f2\u7f6e\u3002\u7cfb\u7edf\u901a\u8fc7\u5f15\u5165\u201c\u4e13\u5bb6\u526f\u672c\uff08Expert Replication\uff09\u201d\u548c\u201c\u52a8\u6001\u8d1f\u8f7d\u611f\u77e5\u8c03\u5ea6\u201d\u6280\u672f\uff0c\u80fd\u591f\u6839\u636e\u5b9e\u65f6\u6d41\u91cf\u81ea\u52a8\u8c03\u6574\u4e0d\u540c\u4e13\u5bb6\u7684\u526f\u672c\u6570\u91cf\u3002\u8fd9\u610f\u5473\u7740\u7cfb\u7edf\u53ef\u4ee5\u52a8\u6001\u5730\u5728\u4e13\u5bb6\u6c60\u4e2d\u91cd\u5206\u914d\u8d44\u6e90\uff0c\u5c06\u8d1f\u8f7d\u538b\u529b\u4ece\u8fc7\u70ed\u8282\u70b9\u8f6c\u79fb\u5230\u7a7a\u95f2\u8282\u70b9\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u4e86\u63a8\u7406\u7684\u957f\u5c3e\u5ef6\u8fdf\uff08Tail Latency\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"218\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/5-2-1024x218.png\"  class=\"wp-image-1354\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/5-2-1024x218.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/5-2-300x64.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/5-2-768x164.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/5-2.png 1125w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe4\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<p>\u5728\u901a\u4fe1\u5c42\u9762\uff0cMegaScale-Infer \u5bf9 All-to-All \u7b97\u5b50\u8fdb\u884c\u4e86\u5b9a\u5236\u5316\u6539\u9020\u3002\u5b83\u5229\u7528\u89e3\u8026\u67b6\u6784\u4e0b\u7684\u7f51\u7edc\u62d3\u6251\u7279\u6027\uff0c\u4f18\u5316\u4e86\u8de8\u8282\u70b9\u7684\u6570\u636e\u4ea4\u6362\u8def\u5f84\u3002\u901a\u8fc7\u4e0e\u5b57\u8282\u8df3\u52a8\u5185\u90e8\u5927\u89c4\u6a21\u751f\u4ea7\u73af\u5883\u7684\u7ed3\u5408\uff0c\u7cfb\u7edf\u80fd\u591f\u5229\u7528\u9ad8\u901f RDMA \u7f51\u7edc\uff0c\u5b9e\u73b0\u8ba1\u7b97\u3001\u663e\u5b58\u62f7\u8d1d\u4e0e\u7f51\u7edc\u4f20\u8f93\u7684\u4e09\u91cd\u91cd\u53e0\u3002\u8fd9\u79cd\u6df1\u5ea6\u7684\u5de5\u7a0b\u4f18\u5316\u4f7f\u5f97\u7cfb\u7edf\u5728\u5904\u7406\u8d85\u5927 Batch Size \u65f6\uff0c\u4f9d\u7136\u80fd\u4fdd\u6301\u7ebf\u6027\u589e\u957f\u7684\u541e\u5410\u80fd\u529b\uff0c\u907f\u514d\u4e86\u4f20\u7edf EP \u5e76\u884c\u4e2d\u5e38\u89c1\u7684\u901a\u4fe1\u5d29\u584c\u73b0\u8c61\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\"><strong>4. \u5b9e\u9a8c\u7ed3\u8bba\u4e0e\u5b66\u672f\u610f\u4e49<\/strong><\/h4>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u5728\u90e8\u7f72 DeepSeek \u6216 Switch Transformer \u7b49\u5927\u89c4\u6a21 MoE \u6a21\u578b\u65f6\uff0cMegaScale-Infer \u76f8\u6bd4\u4e8e\u4f20\u7edf\u7684 vLLM \u6216 DeepSpeed-Inference \u65b9\u6848\uff0c\u5b9e\u73b0\u4e86 <strong>1.5x \u5230 3.2x<\/strong> \u7684\u541e\u5410\u91cf\u63d0\u5347\uff0c\u540c\u65f6\u663e\u8457\u964d\u4f4e\u4e86\u5355\u4f4d\u8bf7\u6c42\u7684\u6210\u672c\u3002\u5728\u7aef\u5230\u7aef\u7684\u751f\u4ea7\u73af\u5883\u6d4b\u8bd5\u4e2d\uff0c\u8be5\u7cfb\u7edf\u5c55\u73b0\u4e86\u6781\u5f3a\u7684\u6269\u5c55\u6027\uff0c\u80fd\u591f\u5e73\u6ed1\u652f\u6491\u4ece\u6570\u5341\u4e2a\u5230\u6570\u5343\u4e2a GPU \u7684\u63a8\u7406\u96c6\u7fa4\u89c4\u6a21\u3002\u7279\u522b\u662f\u5728\u5f02\u6784\u786c\u4ef6\u73af\u5883\u4e0b\uff0c\u89e3\u8026\u67b6\u6784\u7684\u7075\u6d3b\u6027\u5f97\u5230\u4e86\u5145\u5206\u9a8c\u8bc1\uff0c\u8bc1\u660e\u4e86\u6df7\u5408\u4f7f\u7528\u9ad8\u6027\u80fd\u8ba1\u7b97\u663e\u5361\u4e0e\u9ad8\u5e26\u5bbd\u663e\u5b58\u663e\u5361\u7684\u53ef\u884c\u6027\u3002<\/p>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"570\" height=\"327\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6-1.png\"  class=\"wp-image-1355\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6-1.png 570w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/6-1-300x172.png 300w\" sizes=\"auto, (max-width: 570px) 100vw, 570px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe5\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe5\" \/><\/figure>\n<\/div>\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"268\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-1-1024x268.png\"  class=\"wp-image-1357\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-1-1024x268.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-1-300x79.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-1-768x201.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/7-1.png 1110w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe6\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe6\" \/><\/figure>\n\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"581\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1-1024x581.png\"  class=\"wp-image-1358\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1-1024x581.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1-300x170.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1-768x436.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/9-1.png 1068w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe7\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe7\" \/><\/figure>\n<\/div>\n\n<div class=\"wp-block-image\">\n<figure class=\"aligncenter size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"243\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1-1024x243.png\"  class=\"wp-image-1359\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1-1024x243.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1-300x71.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1-768x182.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2026\/05\/10-1.png 1125w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe8\" alt=\"MegaScale-Infer: Efficient Mixture-of-Experts Model Serving with Disaggregated Expert Parallelism\u63d2\u56fe8\" \/><\/figure>\n<\/div>\n\n\n<p><strong>\u7ed3\u8bba\u4e0e\u542f\u793a<\/strong>\uff1aMegaScale-Infer \u7684\u6210\u529f\u6807\u5fd7\u7740\u5206\u5e03\u5f0f\u63a8\u7406\u7cfb\u7edf\u6b63\u4ece\u201c\u901a\u7528\u5e76\u884c\u201d\u5411\u201c\u67b6\u6784\u611f\u77e5\u5e76\u884c\u201d\u6f14\u8fdb\u3002\u5b83\u901a\u8fc7\u89e3\u8026\u8fd9\u4e00\u624b\u6bb5\uff0c\u6253\u7834\u4e86\u5355\u4e00 GPU \u5185\u90e8\u7684\u8d44\u6e90\u7ea6\u675f\u74f6\u9888\u3002\u5bf9\u4e8e\u7814\u7a76\u4eba\u5458\u800c\u8a00\uff0c\u8fd9\u7bc7\u8bba\u6587\u63d0\u4f9b\u4e86\u5904\u7406\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u7684\u65b0\u601d\u8def\u2014\u2014\u5373\u4e0d\u4ec5\u8981\u4f18\u5316\u7b97\u6cd5\u5c42\u9762\u7684\u7a00\u758f\u6027\uff0c\u66f4\u8981\u4ece\u7cfb\u7edf\u67b6\u6784\u5c42\u9762\u53bb\u9002\u914d\u8fd9\u79cd\u7a00\u758f\u6027\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63a2\u7d22\u5982\u4f55\u5728\u89e3\u8026\u73af\u5883\u4e0b\u8fdb\u884c\u81ea\u52a8\u5316\u7684\u6a21\u578b\u5207\u5206\u4e0e\u8d44\u6e90\u7f16\u6392\uff0c\u4ee5\u53ca\u5982\u4f55\u5c06\u8be5\u67b6\u6784\u63a8\u5e7f\u5230\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u957f\u5e8f\u5217\u63a8\u7406\u4e2d\u3002<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>SIGCOMM 2025 1. \u7814\u7a76\u80cc\u666f\u4e0e\u6838\u5fc3\u75db\u70b9\uff1aMoE \u63a8\u7406\u7684\u6548\u7387\u74f6\u9888 \u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u5165\u201c\u4e07\u4ebf\u53c2\u6570\u65f6\u4ee3\u201d\uff0c\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Mixture-of-Experts, MoE\uff09\u56e0\u5176\u80fd\u5728\u589e\u52a0\u53c2\u6570\u91cf\u7684\u540c\u65f6\u4fdd\u6301\u8f83\u4f4e\u7684\u6fc0\u6d3b\u8ba1\u7b97\u91cf\uff0c\u6210\u4e3a\u4e86\u6784\u5efa\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u7684\u9996\u9009\u67b6\u6784\u3002\u7136\u800c\uff0c\u5728\u5b9e\u9645\u63a8\u7406\u573a\u666f\u4e2d\uff0cMoE \u6a21\u578b\u9762\u4e34\u7740\u4e25\u5cfb\u7684\u6311\u6218\u3002\u4f20\u7edf\u7684\u63a8\u7406\u7cfb\u7edf\u901a\u5e38\u5c06\u6ce8\u610f\u529b\u673a\u5236\uff08Attention\uff09\u548c\u5168\u8fde\u63a5\u5c42\uff08FFN\/Experts\uff09\u7ed1\u5b9a\u5728\u540c\u4e00\u5757 GPU \u4e0a\u5904\u7406\uff0c\u8fd9\u79cd\u201c\u8026\u5408\u67b6\u6784\u201d\u5bfc\u81f4\u4e86\u4e25\u91cd\u7684\u8d44\u6e90\u9519\u914d\uff1aAttentio &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=1349\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":5,"featured_media":1360,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[17,1],"tags":[13],"class_list":["post-1349","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-17","category-uncategorized","tag-13"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1349","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\/5"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1349"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1349\/revisions"}],"predecessor-version":[{"id":1361,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1349\/revisions\/1361"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1360"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1349"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1349"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1349"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}