
{"id":10383,"date":"2026-07-21T07:15:27","date_gmt":"2026-07-20T23:15:27","guid":{"rendered":"https:\/\/infernews.com\/blog\/tech-report-expanded-hyper-connections\/"},"modified":"2026-07-21T07:15:27","modified_gmt":"2026-07-20T23:15:27","slug":"tech-report-expanded-hyper-connections","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/tech-report-expanded-hyper-connections\/","title":{"rendered":"xHC \u9ede\u6a23\u628a Transformer \u6b98\u5dee\u6d41\u64f4\u5230 16 \u8def"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/xhc_overview.jpg\" alt=\"xHC architecture overview\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576 Hyper-Connections (HC) \u60f3\u518d\u5f80\u4e0a\u52a0\u6b98\u5dee\u6d41\u6578\u91cf\uff0c\u5361\u4f4d\u5514\u4fc2\u7406\u5ff5\uff0c\u800c\u4fc2\u6210\u672c\u540c\u8cc7\u8a0a\u958b\u59cb\u91cd\u8907\u3002xHC \u5c6c\u65bc\u6a21\u578b\u7d50\u69cb\u7814\u7a76\u9805\u76ee\uff0c\u91dd\u5c0d Transformer residual stream \u64f4\u5c55\u5230\u66f4\u591a\u5e73\u884c streams \u6642\uff0c\u9ede\u6a23\u907f\u514d\u6548\u76ca\u905e\u6e1b\u540c\u8a08\u7b97\u91cf\u66b4\u589e\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">xHC \u5514\u4fc2\u628a\u6240\u6709 streams \u90fd\u5bc6\u96c6\u66f4\u65b0\uff0c\u800c\u4fc2\u4fdd\u7559\u5c0d\u5168\u90e8 N=16 streams \u7684\u8b80\u53d6\uff0c\u518d\u53ea\u5c0d k=4 \u500b active streams \u505a\u7a00\u758f\u66f4\u65b0\u3002\u5481\u6a23\u4e00\u4f86\uff0cHC-family \u5728 N&gt;4 \u6642\u5e38\u898b\u7684 O(N^3C) residual-mapping \u6210\u672c\uff0c\u88ab\u58d3\u5230 O(k^3C)\uff1b\u53e6\u4e00\u908a\u518d\u7528 temporal feature augmentation\uff0c\u88dc\u56de\u55ae\u4e00 write-back vector \u9935\u5514\u98fd\u591a\u8def streams \u7684\u554f\u984c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">xHC \u662f\u9996\u500b\u5728 HC-family \u88e1\u9762\u628a\u6709\u6548\u64f4\u5c55\u63a8\u5230 N=16 \u7684\u505a\u6cd5\uff0c\u4e3b\u6253\u5834\u666f\u4fc2\u60f3\u55ba width\u3001depth \u4e4b\u5916\uff0c\u518d\u589e\u52a0\u4e00\u689d memory-scaling \u8ef8\u3002\u5c0d\u7814\u7a76 Transformer \u67b6\u69cb\u3001\u8a13\u7df4\u5927\u6a21\u578b\uff0c\u6216\u8005\u60f3\u7406\u89e3\u7a00\u758f\u66f4\u65b0\u9ede\u6a23\u63db\u53d6\u66f4\u9ad8\u64f4\u5c55\u6027\u7684\u5718\u968a\uff0c\u5462\u500b\u9805\u76ee\u6709\u53c3\u8003\u50f9\u503c\uff1b\u800c xHC-Flash \u5247\u9032\u4e00\u6b65\u70ba\u90e8\u7f72\u8003\u616e\uff0c\u900f\u904e\u8de8\u9023\u7e8c sublayers \u5171\u7528 full-state \u904b\u7b97\uff0c\u6e1b\u5c11 memory traffic\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u578b\u7d50\u69cb<\/strong>\uff0c\u8655\u7406\u7684\u662f Transformer \u6b98\u5dee\u6d41\u64f4\u5c55\u5230\u591a\u8def\u5f8c\u7684\u6548\u7387\u8207\u6709\u6548\u6027\u554f\u984c\u3002<\/li>\n<li>\u4e3b\u8981\u5dee\u7570\u5728\u65bc\u300c\u5168\u91cf\u8b80\u53d6\u3001\u5c40\u90e8\u66f4\u65b0\u300d\uff1b\u5514\u4fc2\u76f2\u76ee\u52a0 streams\uff0c\u800c\u4fc2\u63a7\u5236\u771f\u6b63\u9700\u8981\u66f4\u65b0\u7684\u8def\u6578\u3002<\/li>\n<li>temporal feature augmentation \u7528 causal depthwise convolutions \u63d0\u4f9b\u591a\u5c3a\u5ea6\u5c40\u90e8\u7279\u5fb5\uff0c\u4ee4\u65b0\u589e streams \u5187\u5481\u6613\u8b8a\u6210\u5197\u9918\u3002<\/li>\n<li>xHC-Flash \u53cd\u6620\u9805\u76ee\u5514\u53ea\u505c\u7559\u55ba\u7406\u8ad6\u8a2d\u8a08\uff0c\u800c\u4fc2\u6709\u9867\u53ca\u8f03\u5927\u898f\u6a21\u8a13\u7df4\u540c\u90e8\u7f72\u6642\u7684\u8a18\u61b6\u9ad4\u6d41\u91cf\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cd\u9ede\u653e\u55ba 18B scale \u7684\u8a13\u7df4\u640d\u5931\u8207\u4e0b\u6e38\u8868\u73fe\u3002\u6578\u5b57\u7d30\u7bc0\u672a\u5728\u7247\u6bb5\u4e2d\u5b8c\u6574\u5c55\u958b\uff0c\u4f46\u65b9\u5411\u5f88\u6e05\u695a\uff1axHC \u60f3\u8b49\u660e\uff0c\u6b98\u5dee\u6d41\u53ef\u4ee5\u64f4\u5f97\u66f4\u95ca\uff0c\u800c\u4e14\u5514\u9700\u8981\u7528\u6210\u500d\u5bc6\u96c6\u66f4\u65b0\u53bb\u63db\u3002\u5c0d\u95dc\u6ce8\u6a21\u578b\u7d50\u69cb\u5275\u65b0\u7684\u4eba\u569f\u8b1b\uff0c\u5b83\u6700\u503c\u5f97\u7747\u7684\u5730\u65b9\u5514\u4fc2\u55ae\u4e00 benchmark \u5206\u6578\uff0c\u800c\u4fc2\u628a HC \u8207 mHC \u63a8\u904e N=4 \u4e4b\u5f8c\uff0c\u4ecd\u7136\u7dad\u6301\u53ef\u8a08\u7b97\u3001\u53ef\u64f4\u5c55\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/aHapBean\/xHC\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.14530\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u60f3\u628a Transformer \u505a\u5f97\u66f4\u8070\u660e\uff0c\u672a\u5fc5\u8981\u4e00\u5473\u52a0\u6df1\u52a0\u95ca\u3002xHC \u63c0\u5497\u53e6\u4e00\u689d\u8def\uff1a\u628a\u6b98\u5dee\u6d41\u64f4\u6210 16 \u8def\uff0c\u4f46\u53ea\u66f4\u65b0\u5176\u4e2d 4 \u8def\uff0c\u63db\u53d6\u66f4\u5be6\u969b\u7684\u64f4\u5c55\u6548\u7387\u3002<\/p>\n","protected":false},"author":8,"featured_media":10382,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,127,199],"tags":[],"class_list":["post-10383","post","type-post","status-publish","format-standard","hentry","category-133","category-127","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10383","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10383"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10383\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10382"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10383"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10383"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10383"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}