{"id":71228,"date":"2026-02-03T12:25:18","date_gmt":"2026-02-03T04:25:18","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/71228.html"},"modified":"2026-02-03T12:25:18","modified_gmt":"2026-02-03T04:25:18","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e5%bf%85%e5%ad%a6%ef%bc%9a%e4%b8%93%e4%b8%9a%e6%9c%af%e8%af%ad%e4%b8%8e%e6%a6%82%e5%bf%b5%e5%bc%95%e5%af%bc","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/71228.html","title":{"rendered":"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc"},"content":{"rendered":"<p>\u672c\u6587\u4f5c\u8005&#xff1a;kaiyuan \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0a&#xff09; \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0b)<\/p>\n<p>\u672c\u6587\u603b\u7ed3\u4e86LLM\u63a8\u7406\u76f8\u5173\u7684\u4e00\u4e9b\u57fa\u7840\u77e5\u8bc6&#xff0c;\u5305\u62ec\u7ecf\u5e38\u7528\u5230\u7684\u4e13\u4e1a\u672f\u8bed&#xff08;\u7f29\u5199\u3001\u540d\u8bcd&#xff09;\u3001\u5173\u952e\u7279\u6027\u3001\u5f15\u64ce\u53c2\u6570\u7b49&#xff0c;\u5e2e\u52a9\u8bfb\u8005\u5bf9\u8be5\u65b9\u5411\u6709\u4e2a\u521d\u6b65\u4e86\u89e3\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aa42182.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h2>1 \u5173\u952e\u7279\u6027<\/h2>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aa6bcbe.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>Continus-batching&#xff1a;\u6301\u7eed\u5730\u5f80GPU\u4e2d\u9001\u5165\u8bf7\u6c42\u6570\u636e&#xff0c;\u800c\u4e0d\u662f\u79bb\u6563\u7684\u8fdb\u884c\u6570\u636e\u63a8\u7406\u3002\u4e00\u4e2a\u8bf7\u6c42\u7ed3\u675f\u7acb\u523b\u4e0b\u53d1\u65b0\u7684\u8bf7\u6c42&#xff0c;\u4fdd\u8bc1GPU\u5229\u7528\u7387&#xff0c;\u53c2\u8003<sup>[1]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aa8f17e.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u9759\u6001vs\u6301\u7eed batching\u7684\u5bf9\u6bd4<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aac9339.gif\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> continuous batching<\/p>\n<ul>\n<li>Paged Attention&#xff1a;\u5206\u9875\u6ce8\u610f\u529b\u673a\u5236\u3002\u5c06Attention\u8fd0\u7b97\u4e2d\u7684KV\u503c\u6309\u7167\u865a\u62df\u6620\u5c04\u7684\u65b9\u5f0f\u7ba1\u7406\u8d77\u6765&#xff0c;\u53ef\u964d\u4f4e\u5185\u5b58\u788e\u7247\u5316&#xff0c;\u53c2\u8003<sup>[2]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042515-698178ab08c35.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>Copy-on-write&#xff1a;\u5199\u65f6\u590d\u5236\u3002\u591a\u4e2a\u8bf7\u6c42\u590d\u7528\u4e00\u4e2a\u672a\u4f7f\u7528\u5b8c\u7684block\u524d\u7f00&#xff0c;\u5f53\u8bf7\u6c42\u5199\u5165\u6570\u636e\u4e0d\u540c\u65f6\u624d\u4e3a\u8be5\u8bf7\u6c42\u590d\u5236\u6570\u636e\u5230\u4e00\u4e2a\u65b0\u7684block\u4e2d\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042515-698178ab35051.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>Radix Attetnion&#xff1a;\u7528\u57fa\u6570\u6811\u6570\u636e\u7ed3\u6784\u6784\u5efaattention\u7684KV cache\u7ba1\u7406\u3002\u7528\u952e\u503c\u7f13\u5b58\u5f20\u91cf\u4ee5\u975e\u8fde\u7eed\u7684\u5206\u9875\u5e03\u5c40\u5b58\u50a8&#xff0c;\u5176\u4e2d\u6bcf\u9875\u7684\u5927\u5c0f\u7b49\u540c\u4e8e\u4e00\u4e2atoken\u3002\u53c2\u8003<sup>[3]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042515-698178ab549d3.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u91c7\u7528LRU\u7b56\u7565\u7684\u57fa\u6570\u6811\u793a\u4f8b<\/p>\n<ul>\n<li>\n<p>PD Disaggregation&#xff1a;PD\u5206\u79bb\u90e8\u7f72\u3002\u6a21\u578b\u8ba1\u7b97\u8fc7\u7a0b\u62c6\u6210\u4e24\u6b65:P&#xff08;prefill&#xff0c;\u7b97\u529b\u5bc6\u96c6&#xff09;\u3001D&#xff08;decode&#xff0c;\u5b58\u50a8\u5bc6\u96c6&#xff09;&#xff0c;\u5c06P\u548cD\u62c6\u5206\u90e8\u7f72\u63d0\u5347GPU\u5229\u7528\u7387\u3002\u53c2\u8003<sup>[4]<\/sup>\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042515-698178abc4dd4.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<\/li>\n<li>\n<p>AFD&#xff08;Attention-FFN Disaggregation&#xff09;&#xff1a;Attention\u4e0eFFN\u5206\u79bb\u3002\u5c06Attention\u4e0eFFN\u5b50\u6a21\u5757\u5206\u522b\u90e8\u7f72\u5230\u4e0d\u540c\u8bbe\u5907\u4e0a&#xff0c;\u901a\u8fc7\u8c03\u6574\u4e0d\u540c\u7684A\u4e0eF\u914d\u6bd4\u53ef\u5b9e\u73b0\u8f83\u9ad8\u63a8\u7406\u6027\u80fd\u3002\u53c2\u8003<sup>[5]<\/sup>\u3002<\/p>\n<\/li>\n<li>\n<p>APC&#xff08;Auto Prefix Cache&#xff09;&#xff1a;\u81ea\u52a8\u524d\u7f00\u7f13\u5b58&#xff0c;\u590d\u7528\u5386\u53f2\u7684KV cache\u4fe1\u606f&#xff0c;\u964d\u4f4eprefill\u9636\u6bb5\u7684\u8ba1\u7b97\u91cf\u4ece\u800c\u63d0\u5347\u9996token\u6027\u80fd\u3002\u53c2\u8003<sup>[6]<\/sup>\u3002<\/p>\n<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac03bef.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p>Speculative Decoding&#xff1a;\u6295\u673a\u63a8\u7406\/\u63a8\u6d4b\u89e3\u7801\u3002\u7528\u8349\u7a3f\u6a21\u578b\u751f\u6210\u6570\u636e&#xff0c;\u7136\u540e\u7528\u53c2\u6570\u91cf\u8f83\u5927\u7684\u6a21\u578b\u8fdb\u884c\u7ed3\u679c\u9a8c\u8bc1\/\u751f\u6210\u3002\u53c2\u8003&lt;LLM\u63d0\u901f\u5229\u5668&#xff1a;\u6295\u673a\u63a8\u7406\u7684\u539f\u7406\u4e0e\u5e38\u89c1\u65b9\u6848&gt;\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac1bec6.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> LLM-1\u8d1f\u8d23\u8f93\u51fa\u3001LLM-2\u8d1f\u8d23\u6821\u9a8c<\/p>\n<p>Chunked Prefill: \u5206\u5757\u9884\u586b\u5145&#xff0c;\u5206\u5757\u9884\u586b\u5145\u5141\u8bb8\u5c06\u5927\u7684prefill\u5206\u5757\u6210\u66f4\u5c0f\u7684\u5757&#xff08;\u5207\u5206\u5e8f\u5217\u957f\u5ea6&#xff09;&#xff0c;\u5e76\u5c06\u5b83\u4eec\u4e0e\u89e3\u7801\u8bf7\u6c42\u4e00\u8d77\u6279\u5904\u7406\u3002\u53c2\u80031<sup>[7]<\/sup>,2<sup>[8]<\/sup>\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac47455.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills <\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac76b3c.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> Chunked prefill \u5728vLLM\u63a8\u7406\u6846\u67b6\u4e2d\u7684\u5e94\u7528<\/p>\n<ul>\n<li>W8A16&#xff1a;\u4e00\u79cd\u91cf\u5316\u7684\u683c\u5f0f&#xff0c;W8\u8868\u793a\u6743\u91cd&#xff08;Weight&#xff09;\u4f7f\u7528 8 \u4f4d\u6574\u6570&#xff08;INT8&#xff09;\u8fdb\u884c\u91cf\u5316&#xff0c;A16&#xff1a;\u8868\u793a\u6fc0\u6d3b\u503c&#xff08;Activation&#xff09;\u4fdd\u7559\u8f83\u9ad8\u7684\u7cbe\u5ea6&#xff0c;\u901a\u5e38\u4f7f\u7528 16\u4f4d\u6d6e\u70b9\u6570&#xff08;FP16&#xff09;\u6216\u8005\u6df7\u5408\u7cbe\u5ea6\u3002\u4e0e\u4e4b\u7c7b\u4f3c\u8fd8\u6709W8A8&#xff0c;W4A16\u7b49&#xff0c;\u540e\u9762\u53ef\u8ddf\u53c2\u6570\u7c7b\u578b&#xff0c;\u5982w8a8_int8\u3002\u53c2\u80031<sup>[9]<\/sup>&#xff0c;2<sup>[10]<\/sup>\u3002<\/li>\n<li>Multi-Lora(Low-Rank Adaptation)&#xff1a;\u591aLoRA\u9002\u914d\u5668\u5171\u7528\u57fa\u7840\u6a21\u578b&#xff0c;\u5728\u62e5\u6709\u4e00\u4e2a\u57fa\u7840\u9884\u8bad\u7ec3\u6a21\u578b\u4e0e\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u5206\u522b\u5fae\u8c03\u7684\u591a\u4e2a\u7279\u5b9aLoRA\u9002\u914d\u5668\u7684\u60c5\u51b5\u4e0b&#xff0c;\u591aLoRA\u670d\u52a1\u673a\u5236\u80fd\u6839\u636e\u4f20\u5165\u8bf7\u6c42\u52a8\u6001\u9009\u62e9\u6240\u9700\u7684LoRA\u6a21\u5757\u3002\u53c2\u8003Lora<sup>[11]<\/sup>&#xff0c;Multi-LoRA<sup>[12]<\/sup>\u3002<\/li>\n<li>Guided Decoder&#xff1a;\u5f15\u5bfc\u89e3\u7801\u5668\u3002\u7ea6\u675f\u6a21\u578b\u8f93\u51fa&#xff0c;\u4f7f\u5176\u4e25\u683c\u9075\u5faa\u9884\u5b9a\u4e49\u7684\u683c\u5f0f\u6216\u8bed\u6cd5\u89c4\u5219&#xff08;\u5982 JSON\u3001SQL\u3001\u6b63\u5219\u8868\u8fbe\u5f0f\u7b49&#xff09;&#xff0c;\u4ece\u800c\u751f\u6210\u7ed3\u6784\u5316\u3001\u53ef\u63a7\u7684\u6587\u672c\u8f93\u51fa\u3002<\/li>\n<li>Function Call\/Tool Call&#xff1a;\u5de5\u5177\u8c03\u7528\u3002\u5229\u7528LLM\u7684\u5f15\u5bfc\u89e3\u7801\u8f93\u51fa\u652f\u6301\u51fd\u6570\u8c03\u7528\u6240\u9700\u53c2\u6570\u683c\u5f0f&#xff0c;\u4f7f\u5f97LLM\u80fd\u591f\u8c03\u7528\u5de5\u5177&#xff0c;\u53c2\u80031<sup>[13]<\/sup>,2<sup>[14]<\/sup>\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac86f1d.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<p> \u51fd\u6570\u8c03\u7528\u7684\u793a\u4f8b<\/p>\n<ul>\n<li>DP\/SP\/TP\/PP: \u63a8\u7406\u5e76\u884c\u7684\u5e38\u89c1\u65b9\u5f0f&#xff0c;\u5305\u62ec&#xff0c;\u6570\u636e\u5e76\u884c&#xff08;DP&#xff09;\u3001\u5e8f\u5217\u5e76\u884c(SP)\u3001\u5f20\u91cf\u5e76\u884c&#xff08;TP&#xff09;\u3001\u5c42\u5e76\u884c&#xff08;PP&#xff09;&#xff0c;\u53c2\u8003<sup>[15]<\/sup>\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ac9d91f.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<p> \u4e00\u822c\u5207batch\u4e3aDP\u3001\u5207\u5e8f\u5217\u4e3aSP\u3001\u5207\u9690\u85cf\u5c42\u5c3a\u5bf8\u4e3aTP\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178aca8515.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u6df7\u5408\u5e76\u884c\u7684\u793a\u610f\u56fe<\/p>\n<ul>\n<li>TBO\/DBO&#xff08;Two Batch Overlap\/Dual Batch Overlap&#xff09;: \u8bf7\u6c42\u62c6\u5206\u4e3a\u5fae\u6279\u6b21&#xff0c;\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u4e0e\u5206\u53d1&#xff08;dispatch&#xff09;\/\u7ec4\u5408&#xff08;combine&#xff09;\u64cd\u4f5c\u4ea4\u9519\u8fdb\u884c&#xff0c;\u8ba1\u7b97\u6d41\u548c\u901a\u4fe1\u6d41\u4ea4\u9519\u6267\u884c&#xff0c;\u63d0\u5347\u7b97\u529b\u5229\u7528\u7387\u3002\u53c2\u80031<sup>[16]<\/sup>,21<sup>[17]<\/sup>\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178acba0a1.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<p> TBO\u4e0b\u53d1\u793a\u610f\u56fe<\/p>\n<ul>\n<li>FA(FlashAttention): \u4e00\u79cd\u52a0\u901fAttention\u8fd0\u7b97\u7684\u7b97\u6cd5&#xff0c;\u4e3b\u8981\u662f\u4e24\u4e2a\u7248\u672cFA1\u3001FA2\u3002FA1<sup>[18]<\/sup>\u4f7f\u7528online softmax\u6280\u5de7&#xff0c;\u5c06\u81ea\u6ce8\u610f\u529b\u4e2d\u591a\u4e2a\u64cd\u4f5c\u6b65\u9aa4\u878d\u5408\u4e3a\u5355\u4e2a\u7b97\u5b50&#xff0c;\u964d\u4f4e\u4e86\u4ece\u5168\u5c40\u5185\u5b58(HBM)\u4e2d\u8bfb\/\u5b58\u6ce8\u610f\u529b\u77e9\u9635\u7684\u5f00\u9500\u3002FA2<sup>[19]<\/sup> , \u8fdb\u4e00\u6b65\u4f18\u5316\u6027\u80fd&#xff0c;\u91c7\u7528\u66f4\u5408\u7406\u7684\u5206\u5757\u7b56\u7565\u5e76\u51cf\u5c11\u975e\u5f20\u91cf\u8fd0\u7b97\u64cd\u4f5c\u6570\u91cf&#xff0c;\u4ee5\u7f13\u89e3NVIDIA A\/H\u67b6\u6784\u4e2d\u975e\u5f20\u91cf\u8ba1\u7b97\u6838\u5fc3\u6027\u80fd\u504f\u4f4e\u7684\u95ee\u9898\u3002 <img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178acc5d6d.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<p> FA2<\/p>\n<p>Flash-Decoding: \u9488\u5bf9\u957f\u5e8f\u5217\u63a8\u7406\u89e3\u7801\u7684Attention\u8ba1\u7b97\u65b9\u5f0f&#xff0c;\u53c2\u8003<sup>[20]<\/sup>\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042516-698178ace6083.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u957f\u5e8f\u5217\u5206\u5757\u8fd0\u7b97<\/p>\n<ul>\n<li>FIA(Flash Infer Attention)&#xff1a;\u5728LLM\u6846\u67b6\u4e2d\u4f7f\u7528\u7684\u4e00\u79cd\u63a8\u7406\u5e93&#xff0c;\u9488\u5bf9\u63a8\u7406\u7684\u4e0d\u540c\u9636\u6bb5&#xff08;prefill\/append\/decode&#xff09;\u5b9e\u73b0\u4e86\u6027\u80fd\u66f4\u4f18\u7684FlashAttention\u7b97\u5b50&#xff0c;\u53c2\u8003<sup>[21]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042517-698178ad2ff8c.webp\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> FIA<\/p>\n<ul>\n<li>LinearAttention&#xff1a;\u7ebf\u6027\u6ce8\u610f\u529b(Linear Attention)\u662fLLM\u4e2d\u4e00\u79cd\u9ad8\u6027\u80fd(\u5b58\u50a8\u66f4\u5c0f\u3001\u8ba1\u7b97\u66f4\u5feb)\u8303\u5f0f\u3002\u91c7\u7528\u7ebf\u6027\u6ce8\u610f\u529b\u7684\u6df7\u5408\u6a21\u578b\u7684\u7cbe\u5ea6\u5728\u4e0d\u65ad\u63d0\u5347&#xff0c;\u67d0\u4e9b\u65b9\u9762\u5df2\u8fbe\u5230\u6807\u51c6\u6ce8\u610f\u529b(Full\/Standard Attention)\u7684\u6c34\u5e73\u3002\u53c2\u8003<sup>[22]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042517-698178ad65c91.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> SSMs\u53d1\u5c55\u5386\u7a0b<\/p>\n<ul>\n<li>\n<p>DPLB(Data parallel Load Balancing)&#xff1a;\u6570\u636e\u5e76\u884c\u8d1f\u8f7d\u5747\u8861&#xff0c;\u4e00\u79cd\u8c03\u5ea6\u673a\u5236\u4fdd\u8bc1\u5206\u53d1\u5230\u5b9e\u4f8b\u4e0a\u7684\u63a8\u7406\u8bf7\u6c42\u8d1f\u8f7d\u5747\u8861&#xff0c;\u4ece\u800c\u63d0\u5347\u6574\u4e2a\u6a21\u5757\u7684\u541e\u5410\/\u6548\u7387&#xff0c;\u53c2\u8003<sup>[23]<\/sup>\u3002 <img decoding=\"async\" src=\"2026-02-034r3iee4ozeq.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<\/li>\n<li>\n<p>EPLB(Expert Parallelism Load Balancer)&#xff1a;\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861&#xff0c;\u5728MoE\u6a21\u578b\u4e2d\u8ba9\u4e0d\u540c\u8bbe\u5907\u4e0a\u7684\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861&#xff0c;\u53c2\u8003<sup>[24]<\/sup>\u3002<\/p>\n<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-030bnwwearwnj.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> DPLB &#043; EPLB<\/p>\n<h2>2 \u6027\u80fd\u6307\u6807<\/h2>\n<p>\u63a8\u7406\u7684\u6027\u80fd\u6307\u6807\u7528\u4e8e\u8bc4\u4ef7\u63a8\u7406\u7cfb\u7edf\u7684\u80fd\u529b&#xff0c;\u901a\u8fc7tokens\u751f\u6210\u901f\u7387\u6765\u8861\u91cf\u670d\u52a1\u7684\u8d28\u91cf\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-02-03tkclszx1soz.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>TTFT(Time To First Token):\u9996token\u751f\u6210\u7684\u65f6\u95f4&#xff0c;\u7528\u4e8e\u8861\u91cfprefill\u6027\u80fd\u7684\u6307\u6807\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03qt1pktyi45y.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>E2EL(End-to-End Latency): \u7aef\u5230\u7aef\u8bf7\u6c42\u65f6\u5ef6\u3002\u4ece\u8f93\u5165\u63d0\u793a\u8bcd\u5230\u751f\u6210\u6240\u6709\u7ed3\u679c\u5e76\u8fd4\u56de\u7ed3\u675f\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03rxhuhbol4s1.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>ITL(Inter token Latency): \u89e3\u7801&#xff08;decode&#xff09;\u9636\u6bb5\u6bcf\u4e2atoken\u7684\u751f\u6210\u65f6\u95f4\u3002\u4e00\u822c\u516c\u5f0f<sup>[25]<\/sup>&#xff1a;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>         I<\/p>\n<p>         T<\/p>\n<p>         L<\/p>\n<p>         &#061;<\/p>\n<p>           E<\/p>\n<p>           2<\/p>\n<p>           E<\/p>\n<p>           L<\/p>\n<p>           \u2212<\/p>\n<p>           T<\/p>\n<p>           T<\/p>\n<p>           F<\/p>\n<p>           T<\/p>\n<p>           n<\/p>\n<p>           \u2212<\/p>\n<p>           1<\/p>\n<p>        ITL &#061; \\\\frac{E2EL- TTFT}{n-1}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0785em\">I<\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">T<\/span><span class=\"mord mathnormal\">L<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1.2757em;vertical-align: -0.4033em\"><\/span><span class=\"mord\"><span class=\"mopen nulldelimiter\"><\/span><span class=\"mfrac\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8723em\"><span class=\"\" style=\"top: -2.655em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">n<\/span><span class=\"mbin mtight\">\u2212<\/span><span class=\"mord mtight\">1<\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -3.23em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"frac-line\" style=\"border-bottom-width: 0.04em\"><\/span><\/span><span class=\"\" style=\"top: -3.394em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0576em\">E<\/span><span class=\"mord mtight\">2<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0576em\">E<\/span><span class=\"mord mathnormal mtight\">L<\/span><span class=\"mbin mtight\">\u2212<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.1389em\">TTFT<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.4033em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><span class=\"mclose nulldelimiter\"><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-030n2ts5hrcfr.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>TBT(Time Between Tokens): \u751f\u6210token\u4e4b\u95f4\u7684\u65f6\u95f4\u5dee&#xff0c;\u4e00\u822c\u6307\u67d0\u4e2atoken\u7684\u751f\u6210\u65f6\u95f4\u3002\u516c\u5f0f&#xff1a;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>         T<\/p>\n<p>         B<\/p>\n<p>          T<\/p>\n<p>          i<\/p>\n<p>         &#061;<\/p>\n<p>         l<\/p>\n<p>         a<\/p>\n<p>         t<\/p>\n<p>         e<\/p>\n<p>         n<\/p>\n<p>         c<\/p>\n<p>          y<\/p>\n<p>          i<\/p>\n<p>         \u2212<\/p>\n<p>         l<\/p>\n<p>         a<\/p>\n<p>         t<\/p>\n<p>         e<\/p>\n<p>         n<\/p>\n<p>         c<\/p>\n<p>          y<\/p>\n<p>           i<\/p>\n<p>           \u2212<\/p>\n<p>           1<\/p>\n<p>        TBT_i &#061; latency_i &#8211; latency_{i-1}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8333em;vertical-align: -0.15em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0502em\">TB<\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">T<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.1389em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">i<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.8889em;vertical-align: -0.1944em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0197em\">l<\/span><span class=\"mord mathnormal\">a<\/span><span class=\"mord mathnormal\">t<\/span><span class=\"mord mathnormal\">e<\/span><span class=\"mord mathnormal\">n<\/span><span class=\"mord mathnormal\">c<\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">y<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">i<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><span class=\"mbin\">\u2212<\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.9028em;vertical-align: -0.2083em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0197em\">l<\/span><span class=\"mord mathnormal\">a<\/span><span class=\"mord mathnormal\">t<\/span><span class=\"mord mathnormal\">e<\/span><span class=\"mord mathnormal\">n<\/span><span class=\"mord mathnormal\">c<\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0359em\">y<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.0359em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mbin mtight\">\u2212<\/span><span class=\"mord mtight\">1<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2083em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n<li>TPOT(Time Per Output Token): \u6240\u6709tokens\u751f\u6210\u7684\u5e73\u5747\u65f6\u95f4&#xff0c;\u5305\u62ec\u9996token\u3002\u516c\u5f0f&#xff1a;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>         T<\/p>\n<p>         P<\/p>\n<p>         O<\/p>\n<p>         T<\/p>\n<p>         &#061;<\/p>\n<p>           E<\/p>\n<p>           2<\/p>\n<p>           E<\/p>\n<p>           L<\/p>\n<p>          n<\/p>\n<p>        TPOT &#061; \\\\frac{E2EL}{n}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">TPOT<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1.2173em;vertical-align: -0.345em\"><\/span><span class=\"mord\"><span class=\"mopen nulldelimiter\"><\/span><span class=\"mfrac\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8723em\"><span class=\"\" style=\"top: -2.655em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">n<\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -3.23em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"frac-line\" style=\"border-bottom-width: 0.04em\"><\/span><\/span><span class=\"\" style=\"top: -3.394em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0576em\">E<\/span><span class=\"mord mtight\">2<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0576em\">E<\/span><span class=\"mord mathnormal mtight\">L<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.345em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><span class=\"mclose nulldelimiter\"><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n<\/ul>\n<p>\u6ce8\u610f&#xff1a;\u4e00\u4e9b\u5e94\u7528\u573a\u666f\u4e0b&#xff08;\u6bd4\u5982vLLM\u4e2d<sup>[26]<\/sup>&#xff09;&#xff0c;ITL\u91c7\u7528\u4e86TBT\u7684\u8ba1\u7b97\u65b9\u5f0f&#xff0c;TPOT\u91c7\u7528\u4e86ITL\u7684\u8ba1\u7b97\u65b9\u5f0f\u3002<\/p>\n<ul>\n<li>QPS(Queries Per Second): \u6bcf\u79d2\u5904\u7406\u7684\u8bf7\u6c42\u6570\u91cf&#xff1b;\u516c\u5f0f:<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>         Q<\/p>\n<p>         P<\/p>\n<p>         S<\/p>\n<p>         &#061;<\/p>\n<p>          T<\/p>\n<p>            \u03a3<\/p>\n<p>             r<\/p>\n<p>             e<\/p>\n<p>             q<\/p>\n<p>             u<\/p>\n<p>             e<\/p>\n<p>             s<\/p>\n<p>             t<\/p>\n<p>             &#061;<\/p>\n<p>             0<\/p>\n<p>            T<\/p>\n<p>           l<\/p>\n<p>           a<\/p>\n<p>           t<\/p>\n<p>           e<\/p>\n<p>           n<\/p>\n<p>           c<\/p>\n<p>           y<\/p>\n<p>           (<\/p>\n<p>           i<\/p>\n<p>           )<\/p>\n<p>        QPS &#061; \\\\frac{T}{\\\\Sigma_{request&#061;0}^{T} latency(i)}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8778em;vertical-align: -0.1944em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0576em\">QPS<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1.5806em;vertical-align: -0.7083em\"><\/span><span class=\"mord\"><span class=\"mopen nulldelimiter\"><\/span><span class=\"mfrac\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8723em\"><span class=\"\" style=\"top: -2.607em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mtight\"><span class=\"mord mtight\">\u03a3<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.8329em\"><span class=\"\" style=\"top: -2.1885em;margin-left: 0em;margin-right: 0.0714em\"><span class=\"pstrut\" style=\"height: 2.5em\"><\/span><span class=\"sizing reset-size3 size1 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">re<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0359em\">q<\/span><span class=\"mord mathnormal mtight\">u<\/span><span class=\"mord mathnormal mtight\">es<\/span><span class=\"mord mathnormal mtight\">t<\/span><span class=\"mrel mtight\">&#061;<\/span><span class=\"mord mtight\">0<\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -2.8448em;margin-right: 0.0714em\"><span class=\"pstrut\" style=\"height: 2.5em\"><\/span><span class=\"sizing reset-size3 size1 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.1389em\">T<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.4504em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0197em\">l<\/span><span class=\"mord mathnormal mtight\">a<\/span><span class=\"mord mathnormal mtight\">t<\/span><span class=\"mord mathnormal mtight\">e<\/span><span class=\"mord mathnormal mtight\">n<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0359em\">cy<\/span><span class=\"mopen mtight\">(<\/span><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mclose mtight\">)<\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -3.23em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"frac-line\" style=\"border-bottom-width: 0.04em\"><\/span><\/span><span class=\"\" style=\"top: -3.394em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.1389em\">T<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.7083em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><span class=\"mclose nulldelimiter\"><\/span><\/span><\/span><\/span><\/span><\/span><\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-032hnovj2gp1i.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>TPS(TokensPerSecond): \u6bcf\u79d2\u541e\u5410\u91cf\u7684\u603b\u8f93\u51fa token \u6570; \u516c\u5f0f<sup>[27]<\/sup>&#xff1a;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>         T<\/p>\n<p>         P<\/p>\n<p>         S<\/p>\n<p>         &#061;<\/p>\n<p>           n<\/p>\n<p>            t<\/p>\n<p>            o<\/p>\n<p>            k<\/p>\n<p>            e<\/p>\n<p>            n<\/p>\n<p>            s<\/p>\n<p>            T<\/p>\n<p>            y<\/p>\n<p>           \u2212<\/p>\n<p>            T<\/p>\n<p>            x<\/p>\n<p>        TPS &#061; \\\\frac{n_{tokens}}{T_y &#8211; T_x}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.0576em\">TPS<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1.2596em;vertical-align: -0.5423em\"><\/span><span class=\"mord\"><span class=\"mopen nulldelimiter\"><\/span><span class=\"mfrac\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.7173em\"><span class=\"\" style=\"top: -2.655em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.1389em\">T<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.1645em\"><span class=\"\" style=\"top: -2.357em;margin-left: -0.1389em;margin-right: 0.0714em\"><span class=\"pstrut\" style=\"height: 2.5em\"><\/span><span class=\"sizing reset-size3 size1 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0359em\">y<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2819em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mbin mtight\">\u2212<\/span><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.1389em\">T<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.1645em\"><span class=\"\" style=\"top: -2.357em;margin-left: -0.1389em;margin-right: 0.0714em\"><span class=\"pstrut\" style=\"height: 2.5em\"><\/span><span class=\"sizing reset-size3 size1 mtight\"><span class=\"mord mathnormal mtight\">x<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.143em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -3.23em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"frac-line\" style=\"border-bottom-width: 0.04em\"><\/span><\/span><span class=\"\" style=\"top: -3.4159em\"><span class=\"pstrut\" style=\"height: 3em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">n<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3448em\"><span class=\"\" style=\"top: -2.3488em;margin-left: 0em;margin-right: 0.0714em\"><span class=\"pstrut\" style=\"height: 2.5em\"><\/span><span class=\"sizing reset-size3 size1 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">t<\/span><span class=\"mord mathnormal mtight\">o<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0315em\">k<\/span><span class=\"mord mathnormal mtight\">e<\/span><span class=\"mord mathnormal mtight\">n<\/span><span class=\"mord mathnormal mtight\">s<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.1512em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.5423em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><span class=\"mclose nulldelimiter\"><\/span><\/span><\/span><\/span><\/span><\/span> &#xff0c;\u5355\u4f4dtok\/s\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03z50nllmjaek.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> TPS\u8ba1\u7b97\u793a\u610f<\/p>\n<ul>\n<li>QPM(Queries Per Minute): \u6bcf\u5206\u949f\u5904\u7406\u7684\u8bf7\u6c42\u7684\u6570\u91cf\u3002<\/li>\n<li>TP90(Top Percentile): \u81f3\u5c11\u670990%\u6216\u800599%\u7684\u8bf7\u6c42\u6ee1\u8db3\u8be5\u6761\u4ef6&#xff0c;\u7c7b\u4f3c\u6307\u6807\u8fd8\u6709TP50\u3001TP99\u3002<\/li>\n<li>RPS(Requests per Second): \u6bcf\u79d2\u8bf7\u6c42\u6570&#xff0c;\u7528\u4e8e\u63a7\u5236\u6d4b\u8bd5\u65f6\u7684\u8bf7\u6c42\u6ce8\u5165\u901f\u7387&#xff0c;\u4e5f\u662f\u541e\u5410\u91cf\u6d4b\u8bd5\u7684\u91cd\u8981\u53c2\u8003\u6307\u6807&#xff0c;\u5355\u4f4dreq\/s\u3002<\/li>\n<li>Ramp Up: \u722c\u5761\u6d4b\u8bd5&#xff0c;\u5728\u4fee\u6539RPS\u6765\u6d4b\u8bd5\u670d\u52a1\u6027\u80fd\u3002<\/li>\n<li>SLO(Service Level Objective)&#xff1a;\u670d\u52a1\u8d28\u91cf\u76ee\u6807,\u662f\u786e\u4fdd\u4e3a\u5ba2\u6237\u63d0\u4f9b\u4f18\u8d28\u670d\u52a1\u7684\u5173\u952e\u3002\u4f8b\u5982\u8fc7\u53bb\u4e00\u6bb5\u65f6\u95f4\u5185\u8bf7\u6c42\u662f\u5426\u90fd\u6ee1\u8db3TP99&#xff0c;TPS&#061;20tok\/s\u3002<\/li>\n<li>MFU(Model Flops Utilization)&#xff1a;\u662f\u8861\u91cf\u6a21\u578b\u5bf9GPU\u7b97\u529b\u8d44\u6e90\u4f7f\u7528\u6548\u7387\u7684\u4e00\u4e2a\u6307\u6807<sup>[28]<\/sup>\u3002<\/li>\n<\/ul>\n<p>\u91c7\u6837\u76f8\u5173\u6982\u5ff5&#xff1a;<\/p>\n<ul>\n<li>Temperature&#xff1a;\u6e29\u5ea6&#xff0c;\u64cd\u4f5c\u7528\u4e8e\u8c03\u6574logits\u7684\u6982\u7387\u5206\u5e03\u6574\u4f53\u60c5\u51b5&#xff0c;\u80fd\u8ba9\u6982\u7387\u5206\u5e03\u53d8\u5f97\u5c16\u9510\u6216\u8005\u5e73\u5766\u3002<\/li>\n<li>TopK&#xff1a;\u6982\u7387\u6392\u5e8f\u540e&#xff0c;\u4fdd\u7559\u6982\u7387\u6700\u5927\u7684K\u4e2a\u503c\u3002\u53d6\u503c\u8303\u56f4\u662f1~vocab_size&#xff08;-1\u8868\u793a\u7981\u7528&#xff09;\u3002<\/li>\n<li>TopP&#xff1a;\u6982\u7387\u6392\u5e8f\u540e&#xff0c;\u53d6\u7d2f\u79ef\u6982\u7387\u5230P\u7684\u503c&#xff0c;\u5148\u6392\u5e8f <span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>          p<\/p>\n<p>           (<\/p>\n<p>           1<\/p>\n<p>           )<\/p>\n<p>         \u2265<\/p>\n<p>          p<\/p>\n<p>           (<\/p>\n<p>           2<\/p>\n<p>           )<\/p>\n<p>         \u2265<\/p>\n<p>         \u22ef<\/p>\n<p>         \u2265<\/p>\n<p>          p<\/p>\n<p>           (<\/p>\n<p>           V<\/p>\n<p>           )<\/p>\n<p>        p_{(1)} \\\\geq p_{(2)} \\\\geq \\\\cdots \\\\geq p_{(V)}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.9912em;vertical-align: -0.3552em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3448em\"><span class=\"\" style=\"top: -2.5198em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mopen mtight\">(<\/span><span class=\"mord mtight\">1<\/span><span class=\"mclose mtight\">)<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3552em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2265<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.9912em;vertical-align: -0.3552em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3448em\"><span class=\"\" style=\"top: -2.5198em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mopen mtight\">(<\/span><span class=\"mord mtight\">2<\/span><span class=\"mclose mtight\">)<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3552em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2265<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.7719em;vertical-align: -0.136em\"><\/span><span class=\"minner\">\u22ef<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2265<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.7858em;vertical-align: -0.3552em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3448em\"><span class=\"\" style=\"top: -2.5198em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mopen mtight\">(<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.2222em\">V<\/span><span class=\"mclose mtight\">)<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3552em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff0c;\u7136\u540e\u7b97\u7d2f\u52a0\u503c&#xff1a;<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"><\/p>\n<p>          C<\/p>\n<p>          k<\/p>\n<p>         &#061;<\/p>\n<p>          \u2211<\/p>\n<p>           i<\/p>\n<p>           &#061;<\/p>\n<p>           1<\/p>\n<p>          k<\/p>\n<p>          p<\/p>\n<p>           (<\/p>\n<p>           i<\/p>\n<p>           )<\/p>\n<p>        C_k &#061; \\\\sum_{i&#061;1}^k p_{(i)}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8333em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0715em\">C<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3361em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.0715em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0315em\">k<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">&#061;<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 1.3442em;vertical-align: -0.3552em\"><\/span><span class=\"mop\"><span class=\"mop op-symbol small-op\" style=\"position: relative;top: 0em\">\u2211<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.989em\"><span class=\"\" style=\"top: -2.4003em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mrel mtight\">&#061;<\/span><span class=\"mord mtight\">1<\/span><\/span><\/span><\/span><span class=\"\" style=\"top: -3.2029em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0315em\">k<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.2997em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.1667em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3448em\"><span class=\"\" style=\"top: -2.5198em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mopen mtight\">(<\/span><span class=\"mord mathnormal mtight\">i<\/span><span class=\"mclose mtight\">)<\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3552em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>&#xff0c; \u4fdd\u7559<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\"><\/p>\n<p>          C<\/p>\n<p>          k<\/p>\n<p>         \u2265<\/p>\n<p>         P<\/p>\n<p>        C_k \\\\geq P<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8333em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\" style=\"margin-right: 0.0715em\">C<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3361em\"><span class=\"\" style=\"top: -2.55em;margin-left: -0.0715em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0315em\">k<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2265<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord mathnormal\" style=\"margin-right: 0.1389em\">P<\/span><\/span><\/span><\/span><\/span>\u7684\u524dK\u4e2a\u503c\u3002topP\u53d6\u503c\u8303\u56f40~1\u3002<\/li>\n<li>MinP&#xff1a;\u4fdd\u7559\u6240\u6709\u6982\u7387\u81f3\u5c11\u4e3a\u6700\u9ad8\u6982\u7387\u7684P\u500d\u7684\u5019\u9009\u8bcd&#xff0c;\u4fdd\u7559\u6240\u6709\u6ee1\u8db3 (<span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\n<p>          p<\/p>\n<p>          i<\/p>\n<p>         \u2265<\/p>\n<p>         minp<\/p>\n<p>         \u00d7<\/p>\n<p>          p<\/p>\n<p>          max<\/p>\n<p>        p_i \\\\geq\\\\text{minp}\\\\times p_{\\\\text{max}}<\/p>\n<p>     <\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8304em;vertical-align: -0.1944em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.3117em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mathnormal mtight\">i<\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><span class=\"mrel\">\u2265<\/span><span class=\"mspace\" style=\"margin-right: 0.2778em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.8623em;vertical-align: -0.1944em\"><\/span><span class=\"mord text\"><span class=\"mord\">minp<\/span><\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><span class=\"mbin\">\u00d7<\/span><span class=\"mspace\" style=\"margin-right: 0.2222em\"><\/span><\/span><span class=\"base\"><span class=\"strut\" style=\"height: 0.625em;vertical-align: -0.1944em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">p<\/span><span class=\"msupsub\"><span class=\"vlist-t vlist-t2\"><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.1514em\"><span class=\"\" style=\"top: -2.55em;margin-left: 0em;margin-right: 0.05em\"><span class=\"pstrut\" style=\"height: 2.7em\"><\/span><span class=\"sizing reset-size6 size3 mtight\"><span class=\"mord mtight\"><span class=\"mord text mtight\"><span class=\"mord mtight\">max<\/span><\/span><\/span><\/span><\/span><\/span><span class=\"vlist-s\">\u200b<\/span><\/span><span class=\"vlist-r\"><span class=\"vlist\" style=\"height: 0.15em\"><span class=\"\"><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>) \u7684\u5019\u9009\u8bcd&#xff0c;min_p\u7684\u53d6\u503c\u8303\u56f4&gt;0\u3002<\/li>\n<li>Frequency Penalty&#xff1a;\u9891\u7387\u60e9\u7f5a&#xff0c;\u5bf9\u51fa\u73b0\u8fc7\u7684\u8bcd&#xff0c;\u6839\u636e\u5176\u51fa\u73b0\u9891\u7387\u964d\u4f4elogits\u503c&#xff0c;\u9891\u7387\u8d8a\u9ad8\u8870\u51cf\u8d8a\u4e25\u91cd\u3002<\/li>\n<li>Presence Penalty&#xff1a;\u5b58\u5728\u60e9\u7f5a&#xff0c;\u5bf9\u51fa\u73b0\u8fc7\u7684\u8bcd&#xff0c;\u5728logits\u4e2d\u51cf\u53bb\u4e00\u4e2a\u76f8\u5e94\u60e9\u7f5a\u503c&#xff0c;\u6bcf\u4e2a\u8bcd\u81f3\u591a\u60e9\u7f5a\u4e00\u6b21\u3002<\/li>\n<li>Repetition Penalty&#xff1a;\u91cd\u590d\u60e9\u7f5a&#xff0c;\u5bf9\u91cd\u590d\u51fa\u73b0\u7684\u8bcd\u8fdb\u884c\u8870\u51cf&#xff0c;\u7c7b\u4f3c\u9891\u7387\u5904\u7406\u3002<\/li>\n<li>Beam Search&#xff1a;\u675f\u641c\u7d22&#xff0c;\u662f\u4e00\u79cd\u7ed3\u5408topK\u548c\u526a\u679d\u7684\u641c\u7d22\u7b97\u6cd5&#xff0c;\u6bcf\u6b21\u4fdd\u7559\u675f\u5bbd&#xff08;beam width&#xff09;k\u4e2a\u7ed3\u679c\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-034lqy0s3u4fu.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u675f\u5bbd&#061;2\u7684beam search\u793a\u610f<\/p>\n<p>\u91c7\u6837\u53c2\u6570\u7684\u8be6\u7ec6\u89e3\u91ca\u53c2\u8003LLM\u63a8\u7406\u57fa\u7840&#xff1a;\u91c7\u6837(Sampling)\u5e38\u89c1\u77e5\u8bc6\u6982\u89c8<\/p>\n<h2>3 \u5f15\u64ce\u4e0e\u6846\u67b6<\/h2>\n<ul>\n<li>vLLM&#xff1a;\u4e00\u6b3e\u4e3b\u6d41\u7684LLM\u63a8\u7406\u6846\u67b6&#xff0c;\u7531\u52a0\u5dde\u5927\u5b66\u4f2f\u514b\u5229\u5206\u6821(UCB)\u7684\u5929\u7a7a\u8ba1\u7b97\u5b9e\u9a8c\u5ba4\u63d0\u51fa&#xff0c;\u6700\u521d\u662f\u57fa\u4e8ePagedAttention\u5f00\u53d1&#xff0c;\u4e3b\u6253\u4e00\u4e2a\u529f\u80fd\u5168&#xff0c;\u751f\u6001\u597d\u3002\u53c2\u80031<sup>[29]<\/sup>,2<sup>[30]<\/sup>\u3002 <img decoding=\"async\" src=\"2026-02-03baw3rei0t2u.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<p> vLLM\u6267\u884c\u6d41\u7a0b&#xff08;\u591aDP\u573a\u666f&#xff09;<\/p>\n<ul>\n<li>SGLang&#xff1a;\u53e6\u4e00\u6b3e\u4e3b\u6d41\u7684LLM\u63a8\u7406\u6846\u67b6&#xff0c;\u7531UCB&#043;\u65af\u5766\u798f\u63d0\u51fa&#xff0c;\u6700\u521d\u662f\u57fa\u4e8eRadixAttetnion\u5f00\u53d1&#xff0c;\u4e3b\u6253\u4e00\u4e2a\u901f\u5ea6\u5feb&#xff0c;\u53c2\u80031<sup>[31]<\/sup>,2<sup>[32]<\/sup>\u3002<\/li>\n<li>Dynamo&#xff1a;NVIDIA\u81ea\u7814\u7684\u4e00\u6b3e\u9488\u5bf9\u5206\u5e03\u5f0f\u63a8\u7406\u7cfb\u7edf\u7684\u89e3\u51b3\u65b9\u6848&#xff0c;\u53c2\u80031<sup>[33]<\/sup>,2<sup>[34]<\/sup> ,3<sup>[35]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03h0htv0lbklk.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> Dynamo\u67b6\u6784\u56fe\u6982\u89c8<\/p>\n<ul>\n<li>AIBrix&#xff1a;LLM\u63a8\u7406\u7684\u4e91\u539f\u751f\u89e3\u51b3\u65b9\u6848&#xff0c;\u7531\u5b57\u8282\u8df3\u52a8\u63a8\u51fa&#xff0c;\u53c2\u80031<sup>[36]<\/sup>,2<sup>[37]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03wcj23bj3dps.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> AIBrix\u67b6\u6784\u56fe\u6982\u89c8<\/p>\n<p>\u5e38\u89c1\u670d\u52a1\u53c2\u6570&#xff08;vLLM&#xff09;&#xff1a;<\/p>\n<ul>\n<li>max-num-seqs&#xff1a;\u6bcf\u6b21\u8fed\u4ee3\u7684\u6700\u5927\u5e8f\u5217\u6570\u3002<\/li>\n<li>max-num-batched-tokens&#xff1a;\u6bcf\u6b21\u8fed\u4ee3\u7684\u6700\u5927\u6279\u5904\u7406 token \u6570\u3002\u63a7\u5236\u603b token \u6570\u3002<\/li>\n<li>max-model-len&#xff1a;\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u5982\u679c\u672a\u6307\u5b9a&#xff0c;\u5219\u4ece\u6a21\u578b\u914d\u7f6e\u4e2d\u81ea\u52a8\u63a8\u5bfc\u3002\u6700\u5927\u8f93\u5165&#043;\u8f93\u51fa tokens\u3002<\/li>\n<li>max-input-length&#xff1a;\u8f93\u5165\u9650\u5236\u957f\u5ea6\u73b0\u9700\u3002<\/li>\n<li>load-format&#xff1a;\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u7684\u683c\u5f0f\u3002\u5176\u4e2d&#034;dummy&#034;&#xff1a;\u7528\u968f\u673a\u503c\u521d\u59cb\u5316\u6743\u91cd&#xff0c;\u4e3b\u8981\u7528\u4e8e\u6027\u80fd\u5206\u6790\u3002<\/li>\n<li>dtype&#xff1a;\u6a21\u578b\u6743\u91cd\u548c\u6fc0\u6d3b\u7684\u6570\u636e\u7c7b\u578b\u3002\u53ef\u9009\u503c: auto, half, float16, bfloat16, float, float32<\/li>\n<li>gpu-memory-utilization&#xff1a; \u7528\u4e8e\u6a21\u578b\u6267\u884c\u5668\u7684 GPU \u5185\u5b58\u6bd4\u4f8b&#xff0c;\u5e76\u4e0d\u662f\u63a8\u7406\u7684\u6574\u4f53\u4f7f\u7528\u91cf&#xff0c;\u53c2\u770b<sup>[38]<\/sup>\u3002<\/li>\n<li>block-size&#xff1a;kv cache\u4e2dblock\u7684\u5927\u5c0f&#xff0c;\u8fde\u7eedtoken\u7684\u6570\u91cf\u3002<\/li>\n<li>enforce-eager&#xff1a;PyTorch\u59cb\u7ec8\u4f7f\u7528 eager \u6a21\u5f0f \u3002<\/li>\n<li>pipeline-parallel-size, pp&#xff1a;\u6d41\u6c34\u7ebf\u9636\u6bb5\u7684\u6570\u91cf\u3002<\/li>\n<li>tensor-parallel-size, tp&#xff1a;\u5f20\u91cf\u5e76\u884c\u526f\u672c\u7684\u6570\u91cf\u3002<\/li>\n<li>data-parallel-size&#xff1a;\u6570\u636e\u5e76\u884c\u526f\u672c\u7684\u6570\u91cf\u3002<\/li>\n<li>enable-expert-parallel&#xff1a;\u5bf9 MoE \u5c42\u4f7f\u7528\u4e13\u5bb6\u5e76\u884c\u800c\u4e0d\u662f\u5f20\u91cf\u5e76\u884c\u3002EP&#061;DP*TP&#xff0c;\u53c2\u8003<sup>[39]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03tba5dks4m12.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> MoE\u4e2dDP&#061;2\u3001TP&#061;2\u3001EP&#061;4\u3002<\/p>\n<p>\u5e38\u89c1\u670d\u52a1\u53c2\u6570&#xff08;SGLang&#xff09;&#xff1a;<\/p>\n<p><img decoding=\"async\" src=\"2026-02-03czipev0q01a.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u5185\u5b58\u548c\u8c03\u5ea6<\/p>\n<p>\u6ce8\u610f&#xff1a;\u5f15\u64ce\u7248\u672c\u7684\u4e0d\u540c\u53c2\u6570\u7565\u6709\u5dee\u5f02&#xff0c;\u5177\u4f53\u53c2\u770b\u5b98\u7f51\u4ecb\u7ecdvLLM<sup>[40]<\/sup>\u3001SGLang<sup>[41]<\/sup>\u3002<\/p>\n<p>vLLM\u5f15\u64ce\u4e3b\u8981\u6a21\u5757&#xff1a;<\/p>\n<ul>\n<li>\n<p>Scheduler\u8c03\u5ea6\u5668&#xff0c;\u6839\u636e\u7cfb\u7edf\u8d44\u6e90\u548c\u8bf7\u6c42\u60c5\u51b5\u7ec4\u7ec7\u6bcf\u6b21\u63a8\u7406\u9700\u8981\u8ba1\u7b97\u7684\u6570\u636e\u3002\u53c2\u8003<sup>[42]<\/sup>\u3002 <img decoding=\"async\" src=\"2026-02-03jugoxbll1w5.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<\/li>\n<li>\n<p>KV cache\u7ba1\u7406&#xff1a;\u4ee5PagedAttention\u4e3a\u57fa\u7840\u8fdb\u884c\u6784\u5efa&#xff0c;\u5206\u4e86\u903b\u8f91\u5c42\u4e0e\u7269\u7406\u5c42&#xff0c;\u8be5\u65b9\u5f0f\u7c7b\u4f3c\u4e8e\u64cd\u4f5c\u7cfb\u7edf\u7684\u865a\u62df\u5185\u5b58&#xff08;virtual memory&#xff09;\u7ba1\u7406\u3002\u53c2\u8003<sup>[43]<\/sup>\u3002 <img decoding=\"async\" src=\"2026-02-03r1vskhmk310.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<\/li>\n<li>\n<p>Engine core\u6267\u884c\u5355\u5143&#xff1a;\u6267\u884c\u8c03\u5ea6\u5668\u4e0b\u53d1\u7684\u5355\u6b65\u4efb\u52a1&#xff0c;\u5305\u62ecKV\u4f20\u8f93\u3001\u6a21\u578b\u6267\u884c&#xff0c;\u53c2\u8003<sup>[44]<\/sup>\u3002<\/p>\n<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-033gktq412a35.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h2>4 \u8f6f\/\u786c\u4ef6\u77e5\u8bc6\u70b9<\/h2>\n<ul>\n<li>CoT(Chain of Thought):\u601d\u7ef4\u94fe&#xff0c;\u63d0\u793a\u4e2d\u5f15\u5bfc\u6216\u8981\u6c42\u6a21\u578b\u5c55\u793a\u5176\u9010\u6b65\u7684\u63a8\u7406\u8fc7\u7a0b&#xff0c;LLM\u7684\u6df1\u5ea6\u601d\u8003&#xff0c;\u53c2\u8003<sup>[45]<\/sup><\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-03j0vplqjwmjw.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<p> \u4e00\u822c\u8f93\u51fa\u4e0eCoT\u8f93\u51fa\u7684\u5dee\u5f02<\/p>\n<ul>\n<li>FCFS(First-Come, First-Served): \u5148\u5230\u5148\u670d\u52a1&#xff0c;\u4e00\u79cd\u8c03\u5ea6\u7b56\u7565\u3002<\/li>\n<li>GPTQ(Generative Pre-trained Transformer Quntanization): \u9488\u5bf9GPT\u7684\u4e00\u79cd\u91cf\u5316\u65b9\u6cd5&#xff0c;\u57fa\u4e8e\u8fd1\u4f3c\u4e8c\u9636\u4fe1\u606f\u7684\u65b0\u578b\u5355\u6b21\u6743\u91cd\u91cf\u5316\u65b9\u6cd5,\u53c2\u8003<sup>[46]<\/sup>\u3002<\/li>\n<li>AWQ(Activation-aware Weight Quantization): \u6fc0\u6d3b\u611f\u77e5\u7684\u6743\u91cd\u91cf\u5316\u65b9\u6cd5\u3002\u6839\u636e\u6fc0\u6d3b\u503c\u5206\u5e03\u6765\u5bf9\u6743\u91cd\u8fdb\u884c\u91cf\u5316&#xff0c;\u91cf\u5316\u7c92\u5ea6\u662f\u901a\u9053\u7ea7\u7684&#xff08;\u533a\u522b\u4e8e\u6574\u5c42\u7528\u4e00\u4e2a\u7cfb\u6570&#xff09;&#xff0c;\u53c2\u8003<sup>[47]<\/sup>\u3002<\/li>\n<li>Llama-2-7b-hf&#xff1a;Hugging Face\u683c\u5f0f\u7684Llama2\u6a21\u578b&#xff0c;\u53c2\u65702billion\u3002<\/li>\n<li>Round Robin&#xff1a;\u8f6e\u8be2\u7b97\u6cd5\u3002\u4f8b\u5982&#xff0c;\u5047\u8bbe\u65f6\u95f4\u7247\u4e3a10\u6beb\u79d2&#xff0c;\u8fdb\u7a0bA\u3001B\u3001C\u6309\u987a\u5e8f\u8fdb\u5165\u5c31\u7eea\u961f\u5217&#xff0c;\u64cd\u4f5c\u7cfb\u7edf\u4f1a\u5148\u8ba9\u8fdb\u7a0bA\u8fd0\u884c10\u6beb\u79d2&#xff0c;\u7136\u540e\u5207\u6362\u5230\u8fdb\u7a0bB\u8fd0\u884c10\u6beb\u79d2&#xff0c;\u63a5\u7740\u662f\u8fdb\u7a0bC&#xff0c;\u4e4b\u540e\u518d\u56de\u5230\u8fdb\u7a0bA&#xff0c;\u5982\u6b64\u5faa\u73af\u3002\u8fd9\u79cd\u7b97\u6cd5\u7684\u4f18\u70b9\u662f\u516c\u5e73&#xff0c;\u6bcf\u4e2a\u8fdb\u7a0b\u90fd\u80fd\u83b7\u5f97\u4e00\u5b9a\u7684\u65f6\u95f4\u7247\u6765\u8fd0\u884c&#xff0c;\u4e0d\u4f1a\u51fa\u73b0\u9965\u997f\u73b0\u8c61&#xff08;\u8fdb\u7a0b\u957f\u65f6\u95f4\u5f97\u4e0d\u5230CPU\u8c03\u5ea6&#xff09;\u3002<\/li>\n<li>EOS&#xff1a;\u7279\u6b8a\u7684\u5e8f\u5217\u7ed3\u675f\u6807\u8bb0&#xff0c;\u4e00\u822c\u751f\u6210token\u65f6\u9047\u5230EOS\u4f1a\u505c\u6b62\u8f93\u51fa\u3002<\/li>\n<li>logprobs&#xff1a;\u5bf9\u6570\u6982\u7387&#xff0c;\u8ba1\u7b97\u65b9\u5f0fln\u00a7&#xff0c;\u6982\u7387\u7d2f\u79ef\u76f8\u4e58\u65f6\u80fd\u591f\u907f\u514d\u9762\u6ea2\u51fa&#xff08;\u8d85\u8fc7\u6d6e\u70b9\u8fdb\u5ea6\u7684\u8868\u8fbe\u8303\u56f4&#xff09;&#xff0c;\u6bd4\u59820.9^200\u3002<\/li>\n<li>Reasoning Parser: \u63a8\u5bfc\u8fc7\u7a0b\u89e3\u6790\u5668&#xff0c;\u53ef\u4ece\u7ed3\u679c\u4e2d\u5206\u79bb\u51fa\u63a8\u7406\u8fc7\u7a0b\u3002<\/li>\n<li>Stream Reasoning: \u6d41\u5f0f\u63a8\u7406&#xff0c;\u4e00\u4e2a\u63a5\u4e00\u4e2a\u7684\u751f\u6210\u7ed3\u679c\u3002\u4e0e\u4e4b\u5bf9\u5e94\u7684\u662f\u975e\u6d41\u5f0f&#xff0c;\u751f\u6210\u6240\u6709\u7ed3\u679c\u540e\u4e00\u6b21\u6027\u8fd4\u56de\u3002<\/li>\n<li>Structured Output: \u7ed3\u6784\u5316\u8f93\u51fa&#xff0c;\u4f7f\u5f97\u8f93\u5165\u6309\u7167\u4e00\u5b9a\u7684\u7ed3\u6784\u7ea6\u675f&#xff0c;\u5982&#xff0c;json\u683c\u5f0f&#xff0c;\u53c2\u770b<sup>[48]<\/sup>\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"2026-02-033s4i02fbn2t.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>RPC(Remote Procedure Call Protocol)&#xff1a;\u8fdc\u7a0b\u8fc7\u7a0b\u8c03\u7528\u534f\u8bae,\u662f\u4e00\u79cd\u901a\u8fc7\u7f51\u7edc\u4ece\u8fdc\u7a0b\u8ba1\u7b97\u673a\u7a0b\u5e8f\u4e0a\u8bf7\u6c42\u670d\u52a1,\u800c\u4e0d\u9700\u8981\u4e86\u89e3\u5e95\u5c42\u7f51\u7edc\u6280\u672f\u7684\u534f\u8bae&#xff61;<\/li>\n<li>Prometheus&#xff1a;\u666e\u7f57\u7c73\u4fee\u65af\u91c7\u96c6\u5de5\u5177&#xff0c;\u662f\u4e00\u6b3e\u57fa\u4e8e\u65f6\u5e8f\u6570\u636e\u5e93\u7684\u5f00\u6e90\u76d1\u63a7\u544a\u8b66\u7cfb\u7edf\u3002<\/li>\n<li>ORCA(Open Request Cost Aggregation)&#xff1a;\u5b83\u5b9a\u4e49\u4e86\u670d\u52a1\u7aef\u70b9&#xff08;\u5982Envoy\u4ee3\u7406&#xff09;\u5982\u4f55\u5411\u5916\u62a5\u544a\u5176\u5b9e\u65f6\u8d1f\u8f7d\u6307\u6807&#xff0c;\u4ee5\u4fbf\u4e0a\u6e38\u7684\u8d1f\u8f7d\u5747\u8861\u5668\u505a\u51fa\u66f4\u667a\u80fd\u7684\u6d41\u91cf\u5206\u53d1\u51b3\u7b56&#xff0c;\u53c2\u80031<sup>[49]<\/sup>\u30012<sup>[50]<\/sup>\u3002<\/li>\n<li>FlashMLA&#xff08;Multi-Head Latent Attention&#xff09;&#xff0c;deepseek\u63a8\u51fa\u7684\u6781\u901f\u7248\u9488\u5bf9\u63a8\u7406\u89e3\u7801\u7684MLA&#xff0c;\u53c2\u8003<sup>[51]<\/sup>\u3002<img decoding=\"async\" src=\"2026-02-031ocehmmpjc2.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<\/ul>\n<h2>5 \u63a8\u7406\u4e2d\u5e38\u7528python\u5e93<\/h2>\n<p>triton&#xff1a;\u4e00\u79cd\u6df1\u5ea6\u5b66\u4e60\u7684\u7f16\u8bd1\u5668\u548c\u52a0\u901f\u5e93&#xff0c;\u6709\u8bb8\u591a\u9ad8\u6027\u80fdGPU\u64cd\u4f5c&#xff08;kernel&#xff09;&#xff0c;\u53c2\u80031<sup>[52]<\/sup>\u30012<sup>[53]<\/sup>\u3002<\/p>\n<p><img decoding=\"async\" src=\"2026-02-033tab01d2xtp.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<ul>\n<li>ray&#xff1a; \u673a\u5668\u5b66\u4e60\u6846\u67b6\u5e93&#xff0c;\u5177\u5907\u7075\u6d3b\u7684\u8d44\u6e90\u8c03\u5ea6\u4e0e\u7ba1\u7406&#xff0c;\u6700\u521d\u8bbe\u8ba1\u4e3b\u8981\u662f\u652f\u6301\u5f3a\u5316\u5b66\u4e60\u3002\u5728\u63a8\u7406\u4e2d\u4e3b\u8981\u7528\u4e8e\u652f\u6491\u591a\u673a\u8d44\u6e90\u5e76\u884c\u8fd0\u7b97&#xff0c;\u53c2\u8003<sup>[54]<\/sup>\u3002<\/li>\n<li>lmcache&#xff1a;\u4e00\u6b3eKV cache\u5b58\u50a8\/\u4f20\u8f93\u7684\u5e93&#xff0c;\u81f4\u529b\u4e8e\u964d\u4f4eTTFT\u6307\u6807&#xff0c;\u53c2\u8003<sup>[55]<\/sup>\u3002<\/li>\n<li>nixl(NVIDIA Inference Xfer Library ): NVIDIA\u63a8\u51fa\u7684\u4e00\u6b3e\u9488\u5bf9\u63a8\u7406\u7684\u4f20\u8f93\u5e93&#xff0c;\u53c2\u8003<sup>[56]<\/sup>\u3002<\/li>\n<li>mooncake&#xff1a;\u4e00\u6b3eKV cache\u5206\u5e03\u5f0f\u5b58\u50a8\/\u4f20\u8f93\u7684\u5e93&#xff0c;\u91cc\u9762\u7684TransferEngine\u5e38\u7528\u4e8ecache\u4f20\u8f93&#xff0c;\u53c2\u8003<sup>[57]<\/sup>\u3002<\/li>\n<li>pytorch&#xff1a;\u901a\u7528\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\u3002<\/li>\n<li>punica&#xff1a;\u652f\u6301multi-lora\u529f\u80fd\u7684\u5e93&#xff0c;\u7279\u70b9\u662f\u63d0\u4f9b\u7edf\u4e00\u7684\u670d\u52a1\u63a5\u53e3&#xff0c;\u53c2\u80031<sup>[58]<\/sup>,2<sup>[59]<\/sup>\u3002<\/li>\n<li>asyncio&#xff1a;\u7528\u4e8e\u7f16\u5199\u5355\u7ebf\u7a0b\u7684\u5e76\u53d1\u4ee3\u7801\u3002\u5b83\u57fa\u4e8e\u4e8b\u4ef6\u5faa\u73af&#xff08;event loop&#xff09;\u6765\u5904\u7406\u5f02\u6b65\u4efb\u52a1&#xff0c;\u5141\u8bb8\u7a0b\u5e8f\u5728\u7b49\u5f85 I\/O \u64cd\u4f5c&#xff08;\u5982\u7f51\u7edc\u8bf7\u6c42\u3001\u6587\u4ef6\u8bfb\u5199&#xff09;\u65f6\u5207\u6362\u5230\u5176\u4ed6\u4efb\u52a1\u3002\u63a8\u7406\u4e2d&#xff0c;\u7ecf\u5e38\u7528\u5b83\u8fdb\u884ccpu\u5f02\u6b65\u64cd\u4f5c\u3002 <img decoding=\"async\" src=\"2026-02-0331dkeegii3u.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<li>uvloop&#xff1a;\u66ff\u4ee3asyncio&#xff0c;\u6027\u80fd\u66f4\u9ad8\u3002<\/li>\n<li>tqdm&#xff1a;\u662f\u4e00\u4e2a\u7528\u4e8e\u5feb\u901f\u521b\u5efa\u8fdb\u5ea6\u6761\u7684\u7b2c\u4e09\u65b9\u5e93&#xff0c;\u540d\u79f0\u6e90\u81ea\u963f\u62c9\u4f2f\u8bed \u201ctaqaddum\u201d&#xff08;\u610f\u4e3a \u201c\u8fdb\u5c55\u201d&#xff09;\u3002<\/li>\n<li>zmq&#xff1a;\u662fZeroMQ\u7684\u7f29\u5199&#xff0c;\u5168\u79f0\u4e3azero message queue\u3002\u662f\u4e00\u4e2a\u9ad8\u6027\u80fd\u3001\u53ef\u6269\u5c55\u7684\u6d88\u606f\u4f20\u9012\u5e93&#xff0c;\u7528\u4e8e\u6784\u5efa\u5206\u5e03\u5f0f\u548c\u5e76\u53d1\u5e94\u7528\u7a0b\u5e8f<sup>[60]<\/sup>\u3002<\/li>\n<li>fastapi&#xff1a;\u521b\u5efaAPI server\u5e38\u7528\u7684\u5de5\u5177&#xff0c;\u5176\u4e2dAPIRouter \u662f\u4e00\u4e2a\u975e\u5e38\u6709\u7528\u7684\u5de5\u5177&#xff0c;\u5b83\u53ef\u4ee5\u5e2e\u52a9\u4f60\u66f4\u597d\u5730\u7ec4\u7ec7\u548c\u7ba1\u7406\u8def\u7531&#xff0c;\u63d0\u9ad8\u4ee3\u7801\u7684\u53ef\u8bfb\u6027\u548c\u53ef\u7ef4\u62a4\u6027\u3002<\/li>\n<\/ul>\n<p>\u5199\u5728\u6700\u540e&#xff0c;\u5165\u95e8\u7684\u4e00\u4e2a\u5efa\u8bae&#xff1a;\u521d\u5b66\u5927\u6a21\u578b\u63a8\u7406\u65f6\u4f1a\u9047\u5230\u8bb8\u591a\u65b0\u6982\u5ff5&#xff0c;\u4e0d\u61c2\u65f6\u4f1a\u501f\u52a9\u641c\u7d22\u5f15\u64ce\u3001\u5927\u6a21\u578b\u6765\u5bfb\u627e\u7b54\u6848&#xff0c;\u4f46\u6709\u4e9b\u7ed3\u679c\u4e0d\u4e00\u5b9a\u51c6\u786e&#xff0c;\u6240\u4ee5\u5f53\u9047\u5230\u6587\u4e2d\u5185\u5bb9\u4e0e\u4f60\u6240\u4e86\u89e3\u6982\u5ff5\u4e0d\u4e00\u81f4\u65f6&#xff0c;\u5efa\u8bae\u5148\u53c2\u770b\u5bf9\u5e94\u7684\u5f15\u7528\u6587\u7ae0\u3002<\/p>\n<hr \/>\n<p>\u53c2\u8003 [1]https:\/\/www.anyscale.com\/blog\/continuous-batching-llm-inference [2]Efficient Memory Management for Large Language Model Serving with PagedAttention https:\/\/arxiv.org\/abs\/2309.06180 [3]SGLang: Efficient Execution of Structured Language Model Programs https:\/\/arxiv.org\/pdf\/2312.07104 [4]https:\/\/zhuanlan.zhihu.com\/p\/1889243870430201414 [5]https:\/\/zhuanlan.zhihu.com\/p\/1952393747112367273 [6]https:\/\/zhuanlan.zhihu.com\/p\/1896927732027335111 [7]https:\/\/arxiv.org\/pdf\/2308.16369 [8]https:\/\/arxiv.org\/pdf\/2403.02310 [9]https:\/\/docs.vllm.com.cn\/en\/latest\/features\/quantization\/index.html [10]https:\/\/docs.sglang.com.cn\/backend\/quantization.html [11]https:\/\/arxiv.org\/pdf\/2106.09685 [12]https:\/\/huggingface.co\/blog\/multi-lora-serving [13]https:\/\/platform.openai.com\/docs\/guides\/function-calling [14]https:\/\/vllm.hyper.ai\/docs\/features\/tool_calling\/ [15]https:\/\/zhuanlan.zhihu.com\/p\/1937449564509545940 [16]https:\/\/docs.vllm.ai\/en\/latest\/design\/dbo\/ [17]https:\/\/lmsys.org\/blog\/2025-05-05-large-scale-ep\/#two-batch-overlap [18[https:\/\/arxiv.org\/abs\/2205.14135 [19]https:\/\/arxiv.org\/abs\/2307.08691 [20]https:\/\/crfm.stanford.edu\/2023\/10\/12\/flashdecoding.html [21]https:\/\/flashinfer.ai\/2024\/02\/02\/introduce-flashinfer.html [22]https:\/\/zhuanlan.zhihu.com\/p\/1969419528065773811 [23]https:\/\/zhuanlan.zhihu.com\/p\/1927317160889386326 [24]https:\/\/zhuanlan.zhihu.com\/p\/29963005584 [25]\u53c2\u8003GenAI-Perf\u5b9a\u4e49 [26]vllm\/vllm\/v1\/metrics\/stats.py [27]https:\/\/developer.nvidia.cn\/blog\/llm-benchmarking-fundamental-concepts\/ [28]https:\/\/zhuanlan.zhihu.com\/p\/20401860293 [29]https:\/\/vllm.hyper.ai\/docs\/ [30]https:\/\/github.com\/vllm-project\/vllm [31]https:\/\/arxiv.org\/abs\/2312.07104 [32]https:\/\/github.com\/sgl-project\/sglang [33]https:\/\/developer.nvidia.com\/dynamo [34]https:\/\/www.zhihu.com\/question\/15465759171\/answer\/129570965681 [35]https:\/\/developer.nvidia.cn\/dynamo [36]https:\/\/arxiv.org\/html\/2504.03648v1 [37]https:\/\/aibrix.readthedocs.io\/latest\/ [38]https:\/\/zhuanlan.zhihu.com\/p\/1916529253169734444 [39]https:\/\/zhuanlan.zhihu.com\/p\/1937556222371946860 [40]https:\/\/vllm.hyper.ai\/docs\/inference-and-serving\/engine_args [41]https:\/\/docs.sglang.com.cn\/backend\/server_arguments.html [42]https:\/\/zhuanlan.zhihu.com\/p\/1900957007575511876 [43]https:\/\/zhuanlan.zhihu.com\/p\/1954128446398633139 [44]https:\/\/zhuanlan.zhihu.com\/p\/1909265969823580330 [45]https:\/\/arxiv.org\/pdf\/2201.11903 [46]https:\/\/arxiv.org\/abs\/2210.17323 [47]https:\/\/arxiv.org\/abs\/2306.00978 [48]https:\/\/vllm.hyper.ai\/docs\/getting-started\/examples\/offline-inference\/structured_outputs\/ [49]https:\/\/docs.cloud.google.com\/load-balancing\/docs\/https\/applb-custom-metrics?hl&#061;zh-cn [50]https:\/\/docs.google.com\/document\/d\/1C1ybMmDKJIVlrbOLbywhu9iRYo4rilR-cT50OTtOFTs\/edit?tab&#061;t.0 [51]https:\/\/zhuanlan.zhihu.com\/p\/27976368445 [52]https:\/\/github.com\/triton-lang\/triton [53]https:\/\/www.eecs.harvard.edu\/~htk\/publication\/2019-mapl-tillet-kung-cox.pdf [54]https:\/\/docs.ray.io\/en\/latest\/index.html [55]https:\/\/github.com\/LMCache\/LMCache [56]https:\/\/github.com\/ai-dynamo\/nixl [57]https:\/\/github.com\/kvcache-ai\/Mooncake [58]https:\/\/arxiv.org\/abs\/2310.18547 [59]https:\/\/github.com\/punica-ai\/punica [60]https:\/\/zeromq.org\/languages\/python\/<\/p>\n<p>InfraTech\u7533\u660e&#xff1a;\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d&#xff01;<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u4f5c\u8005&#xff1a;kaiyuan \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0a&#xff09; \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0b) \u672c\u6587\u603b\u7ed3\u4e86LLM\u63a8\u7406\u76f8\u5173\u7684\u4e00\u4e9b\u57fa\u7840\u77e5\u8bc6&#xff0c;\u5305\u62ec\u7ecf\u5e38\u7528\u5230\u7684\u4e13\u4e1a\u672f\u8bed&#xff08;\u7f29\u5199\u3001\u540d\u8bcd&#xff09;\u3001\u5173\u952e\u7279\u6027\u3001\u5f15\u64ce\u53c2\u6570\u7b49&#xff0c;\u5e2e\u52a9\u8bfb\u8005\u5bf9\u8be5\u65b9\u5411\u6709\u4e2a\u521d\u6b65\u4e86\u89e3\u3002 1 \u5173\u952e\u7279\u6027 Continus-batching&#xff1a;\u6301\u7eed\u5730\u5f80GPU\u4e2d\u9001\u5165\u8bf7\u6c42\u6570\u636e&#xff0c;\u800c\u4e0d\u662f\u79bb\u6563\u7684\u8fdb\u884c\u6570<\/p>\n","protected":false},"author":2,"featured_media":71208,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[132],"topic":[],"class_list":["post-71228","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-132"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/71228.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u672c\u6587\u4f5c\u8005&#xff1a;kaiyuan \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0a&#xff09; \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0b) \u672c\u6587\u603b\u7ed3\u4e86LLM\u63a8\u7406\u76f8\u5173\u7684\u4e00\u4e9b\u57fa\u7840\u77e5\u8bc6&#xff0c;\u5305\u62ec\u7ecf\u5e38\u7528\u5230\u7684\u4e13\u4e1a\u672f\u8bed&#xff08;\u7f29\u5199\u3001\u540d\u8bcd&#xff09;\u3001\u5173\u952e\u7279\u6027\u3001\u5f15\u64ce\u53c2\u6570\u7b49&#xff0c;\u5e2e\u52a9\u8bfb\u8005\u5bf9\u8be5\u65b9\u5411\u6709\u4e2a\u521d\u6b65\u4e86\u89e3\u3002 1 \u5173\u952e\u7279\u6027 Continus-batching&#xff1a;\u6301\u7eed\u5730\u5f80GPU\u4e2d\u9001\u5165\u8bf7\u6c42\u6570\u636e&#xff0c;\u800c\u4e0d\u662f\u79bb\u6563\u7684\u8fdb\u884c\u6570\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/71228.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-02-03T04:25:18+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aa42182.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"8 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/71228.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/71228.html\",\"name\":\"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-02-03T04:25:18+00:00\",\"dateModified\":\"2026-02-03T04:25:18+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/71228.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/71228.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/71228.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/71228.html","og_locale":"zh_CN","og_type":"article","og_title":"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u672c\u6587\u4f5c\u8005&#xff1a;kaiyuan \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0a&#xff09; \u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66&#xff1a;\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc(\u4e0b) \u672c\u6587\u603b\u7ed3\u4e86LLM\u63a8\u7406\u76f8\u5173\u7684\u4e00\u4e9b\u57fa\u7840\u77e5\u8bc6&#xff0c;\u5305\u62ec\u7ecf\u5e38\u7528\u5230\u7684\u4e13\u4e1a\u672f\u8bed&#xff08;\u7f29\u5199\u3001\u540d\u8bcd&#xff09;\u3001\u5173\u952e\u7279\u6027\u3001\u5f15\u64ce\u53c2\u6570\u7b49&#xff0c;\u5e2e\u52a9\u8bfb\u8005\u5bf9\u8be5\u65b9\u5411\u6709\u4e2a\u521d\u6b65\u4e86\u89e3\u3002 1 \u5173\u952e\u7279\u6027 Continus-batching&#xff1a;\u6301\u7eed\u5730\u5f80GPU\u4e2d\u9001\u5165\u8bf7\u6c42\u6570\u636e&#xff0c;\u800c\u4e0d\u662f\u79bb\u6563\u7684\u8fdb\u884c\u6570","og_url":"https:\/\/www.wsisp.com\/helps\/71228.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-02-03T04:25:18+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/02\/20260203042514-698178aa42182.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"8 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/71228.html","url":"https:\/\/www.wsisp.com\/helps\/71228.html","name":"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-02-03T04:25:18+00:00","dateModified":"2026-02-03T04:25:18+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/71228.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/71228.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/71228.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u5927\u6a21\u578b\u63a8\u7406\u5fc5\u5b66\uff1a\u4e13\u4e1a\u672f\u8bed\u4e0e\u6982\u5ff5\u5f15\u5bfc"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/71228","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=71228"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/71228\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/71208"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=71228"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=71228"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=71228"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=71228"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}