{"id":13135,"date":"2026-10-10T01:13:34","date_gmt":"2026-10-09T17:13:34","guid":{"rendered":"https:\/\/infernews.com\/blog\/tokenrouter\/"},"modified":"2026-10-10T01:54:27","modified_gmt":"2026-10-09T17:54:27","slug":"tokenrouter","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/tokenrouter\/","title":{"rendered":"TokenRouter \u8b93\u591a\u500b\u6a21\u578b\u9010 Token \u5354\u4f5c"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/10\/TokenRouter-Token.jpg\" alt=\"TokenRouter Logo\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u4e00\u500b\u56de\u7b54\u9700\u8981\u5c0f\u578b\u6a21\u578b\u5feb\u901f\u751f\u6210\uff0c\u4ea6\u8981\u5927\u578b\u6a21\u578b\u8655\u7406\u6307\u5b9a Token \u6642\uff0c\u670d\u52d9\u7cfb\u7d71\u8981\u540c\u6642\u61c9\u4ed8\u6a21\u578b\u5207\u63db\u3001\u4e0d\u540c\u6b65\u89e3\u78bc\u548c\u4e0d\u898f\u5247\u8acb\u6c42\u3002TokenRouter \u5c6c\u65bc\u958b\u6e90 serving system\uff0c\u5c08\u9580\u8655\u7406 token-level LLM routing\uff0c\u8b93\u591a\u500b\u6a21\u578b\u5728\u55ae\u6b21\u56de\u61c9\u5167\u5354\u4f5c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u958b\u767c\u8005\u53ea\u9700\u900f\u904e route()\u3001send() \u548c receive() \u63cf\u8ff0\u8def\u7531\u908f\u8f2f\uff0c\u7cfb\u7d71\u4fbf\u8ca0\u8cac\u8acb\u6c42\u4ea4\u63a5\u3001Tokenization\u3001Detokenization\u3001\u6279\u6b21\u8abf\u5ea6\u53ca\u6a21\u578b\u57f7\u884c\u3002\u5b83\u63a1\u7528\u7368\u7acb subserver\u3001\u975e\u540c\u6b65 model-centric execution \u548c delayed batching\uff0c\u8b93\u5feb\u6a21\u578b\u7e7c\u7e8c\u89e3\u78bc\uff0c\u6162\u6a21\u578b\u8655\u7406\u88ab\u8def\u7531\u7684 Token\uff1bPython 3.10\u3001YAML \u914d\u7f6e\u3001OpenAI-compatible chat completions \u548c Python engine \u5247\u63d0\u4f9b\u4e86\u8f03\u76f4\u63a5\u7684\u6e2c\u8a66\u5165\u53e3\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u652f\u63f4\u540c\u4e00\u56de\u61c9\u5167\u6309 Token \u6216\u7247\u6bb5\u5207\u63db\u6a21\u578b<\/li>\n<li>\u5167\u7f6e scheduler\u3001model executor \u53ca CUDA graph \u652f\u63f4<\/li>\n<li>\u53ef\u7528\u55ae\u6a5f\u6216\u8de8\u7bc0\u9ede\u914d\u7f6e\u4e0d\u540c\u6a21\u578b<\/li>\n<li>\u6db5\u84cb\u4e94\u7a2e\u8def\u7531\u7b97\u6cd5\u3001\u4e09\u985e\u5de5\u4f5c\u8ca0\u8f09\u53ca\u591a\u7d44\u6a21\u578b\u914d\u5c0d<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6e2c\u8a66\u7d50\u679c\u986f\u793a\uff0cTokenRouter \u5728\u4e0d\u540c\u7b97\u6cd5\u3001\u5de5\u4f5c\u8ca0\u8f09\u548c\u6a21\u578b\u914d\u5c0d\u4e0b\uff0c\u89e3\u78bc\u541e\u5410\u91cf\u8f03\u73fe\u6709\u7cfb\u7d71\u9ad8 2.01\u201364.15 \u500d\u3002\u9019\u500b\u5dee\u8ddd\u6703\u53d7\u786c\u4ef6\u3001\u6279\u6b21\u898f\u6a21\u3001\u8def\u7531\u983b\u7387\u53ca\u6a21\u578b\u7d44\u5408\u5f71\u97ff\uff0c\u4e0d\u80fd\u76f4\u63a5\u8996\u70ba\u6240\u6709\u5834\u666f\u7684\u56fa\u5b9a\u63d0\u5347\uff1b\u9805\u76ee\u63d0\u4f9b\u7684 Qwen3-0.6B \u8207 Qwen3-32B \u914d\u7f6e\uff0c\u8f03\u9069\u5408\u7528\u4f86\u7406\u89e3\u5c0f\u578b\u8207\u5927\u578b\u6a21\u578b\u5354\u4f5c\u7684\u6d41\u7a0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9700\u8981\u7814\u7a76 token-level routing\u3001\u6e2c\u8a66\u63a8\u7406\u5de5\u4f5c\u6d41\uff0c\u6216\u71df\u904b\u591a\u6a21\u578b\u670d\u52d9\u7684\u5718\u968a\u6703\u8f03\u5bb9\u6613\u53d7\u76ca\u3002TokenRouter \u628a\u8def\u7531\u7b56\u7565\u8207\u57f7\u884c\u5f15\u64ce\u5206\u958b\uff0c\u4fdd\u7559\u8f03\u5927\u5f48\u6027\uff0c\u4f46\u4ea6\u610f\u5473\u7740\u4f7f\u7528\u8005\u4ecd\u8981\u81ea\u884c\u8a2d\u8a08\u7b97\u6cd5\u3001\u6e96\u5099\u6a21\u578b\u7bc0\u9ede\u548c\u8abf\u6821\u5206\u6563\u5f0f\u901a\u8a0a\uff0c\u4e26\u975e\u5373\u63d2\u5373\u7528\u7684\u6a21\u578b\u670d\u52d9\u5e73\u53f0\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/fuvty.github.io\/thinking_yard_project_page\/projects\/tokenrouter\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/thu-nics\/TokenRouter\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u540c\u4e00\u500b\u56de\u7b54\u53ef\u6309 Token \u5728\u5927\u5c0f\u6a21\u578b\u4e4b\u9593\u5207\u63db\uff0cTokenRouter \u4ee5\u5c08\u7528\u670d\u52d9\u5f15\u64ce\u8655\u7406\u8def\u7531\u3001\u4ea4\u63a5\u8207\u6279\u6b21\u8abf\u5ea6\u3002<\/p>\n","protected":false},"author":8,"featured_media":13134,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","wpai_meta_description":"","footnotes":""},"categories":[133,76,185,179,152,213],"tags":[],"class_list":["post-13135","post","type-post","status-publish","format-standard","hentry","category-133","category-76","category-qwen","category-nvidia","category-python","category-tokenize"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13135","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=13135"}],"version-history":[{"count":2,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13135\/revisions"}],"predecessor-version":[{"id":13138,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/13135\/revisions\/13138"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/13134"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=13135"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=13135"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=13135"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}