{"id":12949,"date":"2026-10-06T04:07:24","date_gmt":"2026-10-05T20:07:24","guid":{"rendered":"https:\/\/infernews.com\/blog\/gated-token-recurrence-for-efficient-dense-prediction\/"},"modified":"2026-10-06T04:07:24","modified_gmt":"2026-10-05T20:07:24","slug":"gated-token-recurrence-for-efficient-dense-prediction","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/gated-token-recurrence-for-efficient-dense-prediction\/","title":{"rendered":"GTR \u4ee5\u5faa\u74b0\u67b6\u69cb\u58d3\u4f4e\u9ad8\u89e3\u50cf\u5ea6\u8996\u89ba\u63a8\u7406\u6210\u672c"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/teaser.jpg\" alt=\"teaser\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9ad8\u89e3\u50cf\u5ea6\u5f71\u50cf\u4ee4\u5168\u57df softmax attention \u7684\u8a08\u7b97\u91cf\u96a8 token \u6578\u91cf\u5e73\u65b9\u589e\u52a0\uff0cGTR \u4ee5\u7121 softmax \u7684\u5faa\u74b0\u5f0f\u8996\u89ba\u9aa8\u5e79\u8655\u7406\u9019\u500b\u74f6\u9838\uff0c\u652f\u63f4\u7269\u4ef6\u5075\u6e2c\u3001\u5206\u5272\u3001\u59ff\u614b\u3001\u6df1\u5ea6\u53ca\u5b9a\u5411\u5075\u6e2c\u3002\u5b83\u5c6c\u65bc\u8996\u89ba\u6a21\u578b\uff0c\u5be6\u969b\u89e3\u6c7a\u540c\u4e00\u7d44\u5f71\u50cf\u7279\u5fb5\u5982\u4f55\u517c\u9867\u591a\u7a2e\u5bc6\u96c6\u9810\u6e2c\u4efb\u52d9\u8207\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GTR \u7531 12 \u5c64 Gated Linear Attention\uff08GLA\uff09\u7d44\u6210\uff0c\u6bcf\u5c64\u8f2a\u63db\u4e8c\u7dad\u6383\u63cf\u65b9\u5411\uff0c\u518d\u914d\u5408\u52a0\u5165 3\u00d73 depthwise convolution \u7684 Spatial SwiGLU\uff0c\u88dc\u8db3\u9130\u8fd1 patch token \u7684\u5c40\u90e8\u8cc7\u8a0a\u3002\u6a21\u578b\u4ee5 stride-16 \u55ae\u5c3a\u5ea6\u7279\u5fb5\u5716\u904b\u4f5c\uff0c\u4e26\u900f\u904e\u8f15\u91cf\u4e09\u5c3a\u5ea6 projector \u548c query-based head \u8f38\u51fa\u4e0d\u540c\u4efb\u52d9\u6240\u9700\u7d50\u679c\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4ee5\u55ae\u6b21\u5e73\u65b9 L2 loss\uff0c\u5c07\u5b78\u751f\u6a21\u578b\u6700\u7d42 patch token \u5c0d\u9f4a detection-specialized DINOv3 teacher<\/li>\n<li>\u540c\u4e00\u500b GTR backbone \u652f\u63f4\u516d\u985e\u5bc6\u96c6\u9810\u6e2c\u4efb\u52d9<\/li>\n<li>RTX 4090\u3001FP16\u3001batch 1 \u4e0b\uff0cGTR-L \u9054\u5230 58.9 COCO box AP\uff0cmedian forward pass \u70ba 1.908 ms<\/li>\n<li>\u63d0\u4f9b chunkwise CUDA operator\uff0c\u4ea6\u652f\u63f4 TensorRT \u90e8\u7f72\u5230 DRIVE AGX Thor<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5009\u5eab\u5305\u542b\u74b0\u5883\u6e96\u5099\u3001\u8cc7\u6599\u8655\u7406\u3001\u8a13\u7df4\u3001\u8a55\u4f30\u3001\u63a8\u7406\u53ca TensorRT \u90e8\u7f72\u6d41\u7a0b\uff0c\u9069\u5408\u7814\u7a76\u9ad8\u6548\u8996\u89ba\u9aa8\u5e79\u3001\u8eca\u8f09\u611f\u77e5\u6216\u9700\u8981\u591a\u4efb\u52d9\u5171\u7528\u6a21\u578b\u7684\u5718\u968a\u3002\u4ee3\u50f9\u662f CUDA \u5c08\u7528\u7b97\u5b50\u53ca\u7de8\u8b6f\u6d41\u7a0b\u63d0\u9ad8\u4e86\u786c\u4ef6\u4f9d\u8cf4\uff1b\u5b98\u65b9\u5ef6\u9072\u6578\u5b57\u4f86\u81ea RTX 4090 \u6216 DRIVE AGX Thor \u7684\u7279\u5b9a\u8a2d\u5b9a\uff0c\u63db\u7528\u5176\u4ed6\u88dd\u7f6e\u6642\u4ecd\u9700\u91cd\u65b0\u91cf\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/intellindust-ai-lab.github.io\/projects\/GTR\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Intellindust-AI-Lab\/GTR\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>GTR \u4ee5\u7121 softmax \u7684\u5faa\u74b0\u5f0f\u8996\u89ba\u9aa8\u5e79\u8655\u7406\u591a\u7a2e\u5bc6\u96c6\u9810\u6e2c\uff0c\u5728\u901f\u5ea6\u8207\u7cbe\u5ea6\u4e4b\u9593\u5c0b\u627e\u66f4\u5be6\u7528\u7684\u5e73\u8861\u3002<\/p>\n","protected":false},"author":8,"featured_media":12948,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,165,76,149,127,179,30,163,150],"tags":[],"class_list":["post-12949","post","type-post","status-publish","format-standard","hentry","category-133","category-165","category-76","category-149","category-127","category-nvidia","category-image","category-163","category-150"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12949","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=12949"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12949\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12948"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12949"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12949"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12949"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}