{"id":123854,"date":"2026-11-08T09:32:56","date_gmt":"2026-11-08T01:32:56","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/123854.html"},"modified":"2026-11-08T09:32:56","modified_gmt":"2026-11-08T01:32:56","slug":"%e3%80%90llm%e6%8e%a8%e7%90%86%e4%b8%8eserving%e2%91%a5%e3%80%91prefill-decode%e8%a7%a3%e8%80%a6%e6%9e%b6%e6%9e%84%ef%bc%9a%e8%ae%a9%e9%95%bfprompt%e4%b8%8d%e5%86%8d%e5%8d%a1%e4%bd%8f%e7%9f%ad","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/123854.html","title":{"rendered":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42"},"content":{"rendered":"<h2>\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784&#xff1a;\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42<\/h2>\n<p>\u5927\u5bb6\u597d&#xff0c;\u6211\u662f\u4f60\u7684AI Infra\u5b66\u4e60\u642d\u5b50\u3002\u524d\u9762\u6211\u4eec\u804a\u4e86PagedAttention\u3001Continuous Batching\u3001Speculative Decoding\u548cPrefix Cache&#xff0c;\u8fd9\u4e9b\u4f18\u5316\u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u524d\u63d0\u2014\u2014Prefill\u548cDecode\u8dd1\u5728\u540c\u4e00\u4e2aGPU\u4e0a\u3002\u4f46\u5728\u751f\u4ea7\u73af\u5883\u4e2d&#xff0c;Prefill\u548cDecode\u5176\u5b9e\u662f\u4e24\u7c7b\u6027\u683c\u5b8c\u5168\u4e0d\u540c\u7684\u4efb\u52a1\u3002\u628a\u5b83\u4eec\u653e\u5728\u4e00\u8d77&#xff0c;\u5c31\u50cf\u8ba9\u77ed\u8dd1\u8fd0\u52a8\u5458\u548c\u9a6c\u62c9\u677e\u9009\u624b\u5171\u7528\u4e00\u6761\u8dd1\u9053&#xff0c;\u4e92\u76f8\u5e72\u6270\u3002\u4eca\u5929\u6211\u4eec\u5c31\u6765\u804a\u4e00\u804a\u5982\u4f55\u4ece\u67b6\u6784\u5c42\u9762\u628a\u5b83\u4eec\u62c6\u5f00&#xff1a;Prefill\/Decode Disaggregation&#xff08;PD\u89e3\u8026&#xff09;\u3002<\/p>\n<hr \/>\n<h3>\u4e00\u3001\u5f00\u7bc7\u5f15\u5165&#xff1a;\u4e3a\u4ec0\u4e48Prefill\u548cDecode\u8981\u533a\u522b\u5bf9\u5f85&#xff1f;<\/h3>\n<h4>1.1 \u4e00\u4e2a\u771f\u5b9e\u7684\u75db\u70b9<\/h4>\n<p>\u60f3\u8c61\u4f60\u6b63\u5728\u7528AI\u804a\u5929\u673a\u5668\u4eba&#xff0c;\u5bf9\u8bdd\u8fdb\u884c\u5f97\u5f88\u6d41\u7545&#xff0c;\u673a\u5668\u4eba\u6bcf\u79d2\u8e66\u4e24\u4e2a\u5b57&#xff0c;\u4f53\u9a8c\u8fd8\u4e0d\u9519\u3002<\/p>\n<p>\u7a81\u7136&#xff0c;\u6765\u4e86\u4e00\u4e2a&#034;\u72e0\u4eba&#034;\u7528\u6237&#xff0c;\u76f4\u63a5\u7c98\u8d34\u4e86\u4e00\u7bc71\u4e07\u5b57\u7684\u6587\u6863\u8ba9AI\u603b\u7ed3\u3002\u4e8e\u662f\u7cfb\u7edf\u5f00\u59cb\u5904\u7406\u8fd9\u4e2a\u8d85\u957fprompt\u7684prefill\u2026<\/p>\n<p>\u63a5\u4e0b\u6765\u7684\u51e0\u79d2\u949f&#xff0c;\u4f60\u4f1a\u53d1\u73b0&#xff1a;<\/p>\n<ul>\n<li>\u4f60\u6b63\u5728\u8fdb\u884c\u7684\u5bf9\u8bdd\u5361\u4f4f\u4e86&#xff0c;\u5b57\u4e0d\u8e66\u4e86<\/li>\n<li>\u7b49\u4e86\u597d\u51e0\u79d2&#xff0c;\u624d\u6162\u60a0\u60a0\u5730\u8e66\u51fa\u4e00\u4e2a\u5b57<\/li>\n<li>\u6574\u4e2a\u4f53\u9a8c\u53d8\u5f97\u65ad\u65ad\u7eed\u7eed<\/li>\n<\/ul>\n<p>\u8fd9\u5c31\u662fPrefill\u548cDecode\u8026\u5408\u90e8\u7f72\u7684\u5178\u578b\u95ee\u9898&#xff1a;\u957fprompt\u7684prefill\u4f1a&#034;\u62a2\u8d70&#034;\u5927\u91cf\u7b97\u529b&#xff0c;\u5bfc\u81f4\u6b63\u5728decode\u7684\u8bf7\u6c42\u88ab\u997f\u6b7b\u3002<\/p>\n<h4>1.2 \u4e3a\u4ec0\u4e48\u4f1a\u8fd9\u6837&#xff1f;<\/h4>\n<p>\u56e0\u4e3aPrefill\u548cDecode\u867d\u7136\u90fd\u662f&#034;\u6a21\u578b\u63a8\u7406&#034;&#xff0c;\u4f46\u5b83\u4eec\u7684\u8d44\u6e90\u7279\u5f81\u3001\u6027\u80fd\u6307\u6807\u3001\u8c03\u5ea6\u76ee\u6807\u5b8c\u5168\u4e0d\u4e00\u6837\u3002<\/p>\n<p>\u6253\u4e2a\u6bd4\u65b9&#xff1a;<\/p>\n<ul>\n<li>Prefill \u50cf\u662f\u5de5\u5382\u6279\u91cf\u52a0\u5de5\u2014\u2014\u4e00\u4e0b\u5b50\u6765\u4e00\u5927\u5806\u539f\u6750\u6599&#xff08;\u6574\u4e2aprompt&#xff09;&#xff0c;\u5f00\u8db3\u9a6c\u529b\u751f\u4ea7&#xff0c;\u8ffd\u6c42\u7684\u662f&#034;\u591a\u4e45\u80fd\u52a0\u5de5\u5b8c\u8fd9\u6279\u8d27&#034;\u3002<\/li>\n<li>Decode \u50cf\u662f\u5916\u5356\u9a91\u624b\u9001\u5355\u2014\u2014\u6bcf\u6b21\u53ea\u9001\u4e00\u5355&#xff08;\u4e00\u4e2atoken&#xff09;&#xff0c;\u8ffd\u6c42\u7684\u662f&#034;\u6bcf\u4e00\u5355\u591a\u4e45\u80fd\u9001\u5230&#034;&#xff0c;\u800c\u4e14\u8981\u7a33\u5b9a\u3001\u4e0d\u80fd\u4e00\u4f1a\u513f\u5feb\u4e00\u4f1a\u513f\u6162\u3002<\/li>\n<\/ul>\n<p>\u628a\u6279\u91cf\u52a0\u5de5\u548c\u5916\u5356\u914d\u9001\u653e\u5728\u540c\u4e00\u4e2a\u8f66\u95f4\u91cc&#xff0c;\u7ed3\u679c\u5c31\u662f&#xff1a;\u6279\u91cf\u52a0\u5de5\u4e00\u6765&#xff0c;\u5916\u5356\u5168\u5835\u5728\u8def\u4e0a\u4e86\u3002<\/p>\n<hr \/>\n<h3>\u4e8c\u3001Prefill\u4e0eDecode\u7684\u672c\u8d28\u5dee\u5f02<\/h3>\n<h4>2.1 \u5168\u65b9\u4f4d\u5bf9\u6bd4<\/h4>\n<p>\u6211\u4eec\u4ece\u591a\u4e2a\u7ef4\u5ea6\u6765\u5bf9\u6bd4\u4e00\u4e0bPrefill\u548cDecode&#xff1a;<\/p>\n<table>\n<tr>\u7ef4\u5ea6Prefill&#xff08;\u9884\u586b\u5145&#xff09;Decode&#xff08;\u89e3\u7801\u751f\u6210&#xff09;<\/tr>\n<tbody>\n<tr>\n<td>\u8f93\u5165<\/td>\n<td>\u6574\u4e2aprompt&#xff0c;\u51e0\u5341\u5230\u51e0\u4e07tokens<\/td>\n<td>\u6bcf\u6b65\u53ea\u8f93\u51651\u4e2a\u65b0token<\/td>\n<\/tr>\n<tr>\n<td>\u8ba1\u7b97\u7279\u5f81<\/td>\n<td>\u5927\u77e9\u9635\u4e58\u6cd5&#xff08;GEMM&#xff09;&#xff0c;compute-bound<\/td>\n<td>\u5c0f\u6279\u91cf\u5411\u91cf\u4e58&#xff08;GEMV&#xff09;&#043; \u5927\u91cfKV\u8bfb\u53d6&#xff0c;memory-bandwidth-bound<\/td>\n<\/tr>\n<tr>\n<td>\u4e3b\u8981\u5ef6\u8fdf\u6307\u6807<\/td>\n<td>TTFT&#xff08;Time To First Token&#xff0c;\u9996token\u5ef6\u8fdf&#xff09;<\/td>\n<td>TPOT \/ ITL&#xff08;\u6bcftoken\u65f6\u95f4 \/ token\u95f4\u9694&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u5e76\u884c\u65b9\u5f0f<\/td>\n<td>prompt\u5185token\u9ad8\u5ea6\u5e76\u884c&#xff0c;\u4e00\u6b21\u7b97\u51fa\u6240\u6709\u4f4d\u7f6e<\/td>\n<td>\u81ea\u56de\u5f52&#xff0c;token\u4e4b\u95f4\u4e32\u884c<\/td>\n<\/tr>\n<tr>\n<td>KV Cache<\/td>\n<td>\u5199\u5165\u5927\u91cfprompt KV<\/td>\n<td>\u6bcf\u6b65\u8ffd\u52a01\u4e2atoken\u7684KV&#xff0c;\u5e76\u8bfb\u53d6\u6240\u6709\u5386\u53f2KV<\/td>\n<\/tr>\n<tr>\n<td>\u8c03\u5ea6\u76ee\u6807<\/td>\n<td>\u5c3d\u5feb\u5b8c\u6210&#xff0c;\u541e\u5410\u4f18\u5148<\/td>\n<td>\u7a33\u5b9a\u4f4e\u5ef6\u8fdf&#xff0c;\u6296\u52a8\u6700\u5c0f<\/td>\n<\/tr>\n<tr>\n<td>GPU\u5229\u7528\u7387<\/td>\n<td>\u9ad8&#xff0c;\u63a5\u8fd1\u6253\u6ee1<\/td>\n<td>\u4f4e&#xff0c;\u901a\u5e38\u53ea\u670910%-30%&#xff08;\u5355\u8bf7\u6c42\u65f6&#xff09;<\/td>\n<\/tr>\n<tr>\n<td>\u74f6\u9888\u8d44\u6e90<\/td>\n<td>\u8ba1\u7b97\u5355\u5143&#xff08;SM&#xff09;<\/td>\n<td>\u663e\u5b58\u5e26\u5bbd&#xff08;HBM&#xff09;<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h4>2.2 \u4e00\u53e5\u8bdd\u603b\u7ed3<\/h4>\n<p>Prefill\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u7684&#034;\u6279\u91cf\u4efb\u52a1&#034;&#xff0c;Decode\u662f\u5e26\u5bbd\u5bc6\u96c6\u578b\u7684&#034;\u6d41\u5f0f\u4efb\u52a1&#034;\u3002<\/p>\n<p>\u5b83\u4eec\u7684\u8d44\u6e90\u74f6\u9888\u5b8c\u5168\u4e0d\u540c&#xff0c;\u8c03\u5ea6\u76ee\u6807\u4e5f\u5b8c\u5168\u4e0d\u540c\u3002\u628a\u5b83\u4eec\u786c\u585e\u5728\u4e00\u8d77&#xff0c;\u7ed3\u679c\u5c31\u662f\u4e24\u8fb9\u90fd\u505a\u4e0d\u597d\u3002<\/p>\n<h4>2.3 \u4e3a\u4ec0\u4e48\u4e4b\u524d\u4e0d\u62c6\u5f00&#xff1f;<\/h4>\n<p>\u4f60\u53ef\u80fd\u4f1a\u95ee&#xff1a;\u65e2\u7136\u5dee\u5f02\u8fd9\u4e48\u5927&#xff0c;\u4e3a\u4ec0\u4e48\u4e4b\u524d\u4e0d\u5206\u5f00&#xff1f;<\/p>\n<p>\u7b54\u6848\u662f&#xff1a;\u56e0\u4e3aKV Cache\u3002<\/p>\n<p>Prefill\u4ea7\u751f\u7684KV Cache&#xff0c;Decode\u9636\u6bb5\u9a6c\u4e0a\u5c31\u8981\u7528\u3002\u5982\u679cPrefill\u548cDecode\u5728\u4e0d\u540c\u7684GPU\u4e0a&#xff0c;\u90a3prefill\u5b8c\u4e86\u8fd8\u8981\u628aKV\u4f20\u8fc7\u53bb\u2014\u2014\u8fd9\u4e2a\u4f20\u8f93\u5f00\u9500\u8fc7\u53bb\u88ab\u8ba4\u4e3a\u662f\u4e0d\u53ef\u63a5\u53d7\u7684\u3002<\/p>\n<p>\u4f46\u968f\u7740\u6280\u672f\u7684\u53d1\u5c55&#xff0c;\u60c5\u51b5\u53d1\u751f\u4e86\u53d8\u5316&#xff1a;<\/p>\n<li>\u6a21\u578b\u8d8a\u6765\u8d8a\u5927&#xff0c;prompt\u8d8a\u6765\u8d8a\u957f&#xff0c;prefill\u7684\u65f6\u95f4\u5360\u6bd4\u8d8a\u6765\u8d8a\u9ad8&#xff0c;\u8026\u5408\u7684\u4ee3\u4ef7\u4e5f\u8d8a\u6765\u8d8a\u5927<\/li>\n<li>NVLink\u3001RDMA\u7b49\u9ad8\u901f\u4e92\u8054\u6280\u672f\u8d8a\u6765\u8d8a\u6210\u719f&#xff0c;KV\u4f20\u8f93\u7684\u65f6\u95f4\u6210\u672c\u5728\u4e0b\u964d<\/li>\n<li>\u4e1a\u52a1\u5bf9SLO&#xff08;\u670d\u52a1\u7b49\u7ea7\u76ee\u6807&#xff09;\u7684\u8981\u6c42\u8d8a\u6765\u8d8a\u9ad8&#xff0c;TPOT\u6296\u52a8\u53d8\u5f97\u4e0d\u53ef\u63a5\u53d7<\/li>\n<p>\u4e8e\u662f&#xff0c;PD\u89e3\u8026\u91cd\u65b0\u8fdb\u5165\u4e86\u4eba\u4eec\u7684\u89c6\u91ce\u3002<\/p>\n<hr \/>\n<h3>\u4e09\u3001\u8026\u5408\u90e8\u7f72\u7684\u5178\u578b\u5e72\u6270<\/h3>\n<h4>3.1 \u957fPrefill\u963b\u585eDecode<\/h4>\n<p>\u8fd9\u662f\u6700\u5e38\u89c1\u4e5f\u6700\u76f4\u89c2\u7684\u95ee\u9898\u3002\u5f53\u4e00\u4e2a\u957fprompt\u8bf7\u6c42\u5230\u6765\u65f6&#xff1a;<\/p>\n<p>\u65f6\u95f4\u7ebf&#xff1a;<\/p>\n<p>t0:  decode batch \u6b63\u5728\u7a33\u5b9a\u8f93\u51fa&#xff0c;TPOT &#061; 50ms<br \/>\n     \u250c\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2510<br \/>\n     \u2502tok\u2502tok\u2502tok\u2502tok\u2502tok\u2502  &#8230;<br \/>\n     \u2514\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2518<br \/>\n      50ms 50ms 50ms 50ms<\/p>\n<p>t1:  \u957f prompt&#xff08;8192 token&#xff09;\u5230\u8fbe&#xff0c;\u5f00\u59cb prefill<br \/>\n     prefill \u62a2\u5360\u4e86\u51e0\u4e4e\u6240\u6709\u7b97\u529b<\/p>\n<p>t2:  decode \u88ab\u5361\u4f4f&#xff0c;TPOT \u98d9\u5347\u5230 500ms&#043;<br \/>\n     \u250c\u2500\u2500\u2500\u2510                    \u250c\u2500\u2500\u2500\u2510<br \/>\n     \u2502tok\u2502                    \u2502tok\u2502  &#8230;<br \/>\n     \u2514\u2500\u2500\u2500\u2518                    \u2514\u2500\u2500\u2500\u2518<br \/>\n      50ms   \u2500\u2500 \u7b49\u5f85 500ms \u2500\u2500  50ms<\/p>\n<p>t3:  prefill \u5b8c\u6210&#xff0c;decode \u9010\u6e10\u6062\u590d<\/p>\n<p>\u7528\u6237\u611f\u77e5\u5c31\u662f&#xff1a;\u804a\u7740\u804a\u7740\u7a81\u7136\u5361\u4f4f\u4e86&#xff0c;\u534a\u5929\u8e66\u4e0d\u51fa\u4e00\u4e2a\u5b57\u3002<\/p>\n<h4>3.2 TTFT\u53d8\u5dee<\/h4>\n<p>TTFT&#xff08;\u9996token\u5ef6\u8fdf&#xff09;\u662f\u7528\u6237\u4f53\u9a8c\u7684\u5173\u952e\u6307\u6807\u2014\u2014\u7528\u6237\u53d1\u5b8c\u6d88\u606f\u540e&#xff0c;\u591a\u4e45\u80fd\u770b\u5230AI\u5f00\u59cb\u56de\u590d\u3002<\/p>\n<p>\u5728\u8026\u5408\u90e8\u7f72\u4e2d&#xff1a;<\/p>\n<ul>\n<li>\u77edprompt\u8bf7\u6c42\u672c\u6765\u53ef\u4ee5\u5f88\u5feb\u5b8c\u6210prefill<\/li>\n<li>\u4f46\u5982\u679c\u5b83\u6392\u5728\u4e00\u4e2a\u957fprompt\u540e\u9762&#xff0c;\u5c31\u8981\u7b49\u5f88\u4e45<\/li>\n<li>\u5c31\u50cf\u4f60\u53bb\u8d85\u5e02\u53ea\u4e70\u4e00\u74f6\u6c34&#xff0c;\u5374\u6392\u5728\u4e86\u56e4\u8d27\u7684\u4eba\u540e\u9762<\/li>\n<\/ul>\n<h4>3.3 TPOT\u6296\u52a8<\/h4>\n<p>Decode\u9636\u6bb5\u7684TPOT\u9700\u8981\u7a33\u5b9a&#xff0c;\u7528\u6237\u624d\u4f1a\u89c9\u5f97&#034;\u6d41\u7545&#034;\u3002<\/p>\n<p>\u4f46\u5728\u8026\u5408\u90e8\u7f72\u4e2d&#xff0c;\u6bcf\u5f53\u6709\u65b0\u8bf7\u6c42\u8fdb\u6765\u505aprefill&#xff0c;\u6b63\u5728\u8fdb\u884c\u7684decode\u5c31\u4f1a\u88ab&#034;\u6253\u6270&#034;\u4e00\u4e0b&#xff0c;TPOT\u5c31\u4f1a\u6296\u4e00\u4e0b\u3002\u5982\u679c\u65b0\u8bf7\u6c42\u5f88\u591a&#xff0c;TPOT\u5c31\u50cf\u8fc7\u5c71\u8f66\u4e00\u6837\u5ffd\u9ad8\u5ffd\u4f4e\u3002<\/p>\n<h4>3.4 SLO\u96be\u505a<\/h4>\n<p>SLO&#xff08;\u670d\u52a1\u7b49\u7ea7\u76ee\u6807&#xff09;\u901a\u5e38\u4f1a\u89c4\u5b9a&#xff1a;<\/p>\n<ul>\n<li>99\u5206\u4f4d\u7684TTFT\u4e0d\u8d85\u8fc7X\u79d2<\/li>\n<li>99\u5206\u4f4d\u7684TPOT\u4e0d\u8d85\u8fc7Y\u6beb\u79d2<\/li>\n<\/ul>\n<p>\u4f46Prefill\u548cDecode\u8026\u5408\u5728\u4e00\u8d77\u65f6&#xff0c;\u8fd9\u4e24\u4e2a\u6307\u6807\u662f\u4e92\u76f8\u5f71\u54cd\u7684&#xff1a;<\/p>\n<ul>\n<li>\u4e3a\u4e86\u4fdd\u8bc1TTFT&#xff0c;\u4f60\u9700\u8981\u8ba9prefill\u5c3d\u5feb\u8dd1&#xff0c;\u4f46\u8fd9\u6837\u4f1a\u6324\u5360decode\u7684\u8d44\u6e90<\/li>\n<li>\u4e3a\u4e86\u4fdd\u8bc1TPOT&#xff0c;\u4f60\u9700\u8981\u7ed9decode\u7559\u8db3\u8d44\u6e90&#xff0c;\u4f46\u8fd9\u6837prefill\u4f1a\u53d8\u6162<\/li>\n<\/ul>\n<p>\u6269\u5bb9\u7b56\u7565\u4e5f\u5f88\u5c34\u5c2c&#xff1a;TTFT\u4e0d\u591f\u7528\u7684\u65f6\u5019\u52a0\u673a\u5668&#xff0c;\u52a0\u5b8c\u53d1\u73b0TPOT\u8fd8\u662f\u6296&#xff1b;TPOT\u4e0d\u591f\u7528\u7684\u65f6\u5019\u52a0\u673a\u5668&#xff0c;\u52a0\u5b8c\u53d1\u73b0TTFT\u8fd8\u662f\u6162\u3002<\/p>\n<hr \/>\n<h3>\u56db\u3001PD Disaggregation\u67b6\u6784<\/h3>\n<h4>4.1 \u57fa\u672c\u67b6\u6784<\/h4>\n<p>PD\u89e3\u8026\u7684\u601d\u8def\u5f88\u76f4\u63a5&#xff1a;\u628aPrefill\u548cDecode\u5206\u5230\u4e0d\u540c\u7684Worker\u6c60\u91cc&#xff0c;\u5404\u53f8\u5176\u804c\u3002<\/p>\n<p>Client Request<br \/>\n      \u2502<br \/>\n      \u25bc<br \/>\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<br \/>\n\u2502  Router \/       \u2502<br \/>\n\u2502  Scheduler      \u2502  \u2190 \u8d1f\u8d23\u8def\u7531\u3001\u8c03\u5ea6\u3001SLO\u7ba1\u7406<br \/>\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<br \/>\n        \u2502<br \/>\n   \u250c\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2510<br \/>\n   \u25bc         \u25bc<br \/>\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2510  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2510<br \/>\n\u2502Prefill\u2502  \u2502Decode\u2502<br \/>\n\u2502Worker \u2502  \u2502Worker\u2502<br \/>\n\u2502 Pool  \u2502  \u2502 Pool  \u2502<br \/>\n\u2514\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2518  \u2514\u2500\u2500\u2500\u25b2\u2500\u2500\u2500\u2518<br \/>\n    \u2502          \u2502<br \/>\n    \u2502  KV      \u2502  \u6d41\u5f0f<br \/>\n    \u2502 Transfer \u2502  \u8fd4\u56de<br \/>\n    \u25bc          \u2502<br \/>\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510<br \/>\n\u2502  KV Transfer Channel \u2502<br \/>\n\u2502 (NVLink\/PCIe\/RDMA)   \u2502<br \/>\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n<p>\u6838\u5fc3\u7ec4\u4ef6&#xff1a;<\/p>\n<li>Router \/ Scheduler&#xff1a;\u8bf7\u6c42\u5165\u53e3&#xff0c;\u8d1f\u8d23\u628a\u8bf7\u6c42\u5206\u5230\u5408\u9002\u7684Prefill Worker&#xff0c;prefill\u5b8c\u6210\u540e\u518d\u5206\u914d\u5230Decode Worker\u3002<\/li>\n<li>Prefill Worker Pool&#xff1a;\u4e13\u95e8\u505aprefill\u7684GPU\u6c60&#xff0c;\u4f18\u5316\u9ad8\u541e\u5410\u8ba1\u7b97\u3002<\/li>\n<li>Decode Worker Pool&#xff1a;\u4e13\u95e8\u505adecode\u7684GPU\u6c60&#xff0c;\u4f18\u5316\u4f4e\u5ef6\u8fdf\u7a33\u5b9a\u8f93\u51fa\u3002<\/li>\n<li>KV Transfer Channel&#xff1a;\u628aPrefill\u4ea7\u751f\u7684KV Cache\u4f20\u8f93\u5230Decode Worker\u3002<\/li>\n<h4>4.2 \u8bbe\u8ba1\u601d\u60f3<\/h4>\n<ul>\n<li>\n<p>Prefill Worker&#xff1a;\u8ffd\u6c42\u9ad8\u541e\u5410&#xff0c;\u600e\u4e48\u5feb\u600e\u4e48\u6765<\/p>\n<ul>\n<li>\u53ef\u4ee5\u7528\u66f4\u5927\u7684batch size<\/li>\n<li>\u53ef\u4ee5\u7528\u66f4\u9ad8\u7684\u5e76\u884c\u5ea6<\/li>\n<li>\u4e0d\u7528\u5173\u5fc3TPOT\u6296\u52a8<\/li>\n<li>\u751a\u81f3\u53ef\u4ee5\u7528\u4e0d\u540c\u578b\u53f7\u7684GPU&#xff08;\u6bd4\u5982Prefill\u7528\u8ba1\u7b97\u5361&#xff0c;Decode\u7528\u5e26\u5bbd\u9ad8\u7684\u5361&#xff09;<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>Decode Worker&#xff1a;\u8ffd\u6c42\u4f4e\u5ef6\u8fdf\u3001\u4f4e\u6296\u52a8<\/p>\n<ul>\n<li>batch size\u63a7\u5236\u5728\u5408\u9002\u8303\u56f4<\/li>\n<li>\u4f18\u5148\u4fdd\u8bc1\u6bcf\u6b65\u7684\u7a33\u5b9a\u6027<\/li>\n<li>\u4e0d\u7528\u88ab\u957fprefill\u6253\u65ad<\/li>\n<\/ul>\n<\/li>\n<li>\n<p>KV Transfer&#xff1a;\u628aPrefill\u7684&#034;\u6210\u679c&#034;&#xff08;KV Cache&#xff09;\u5b89\u5168\u3001\u5feb\u901f\u5730\u4ea4\u7ed9Decode<\/p>\n<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e94\u3001\u8bf7\u6c42\u751f\u547d\u5468\u671f&#xff1a;Router\u2192Prefill\u2192KV Transfer\u2192Decode<\/h3>\n<h4>5.1 \u5b8c\u6574\u6d41\u7a0b<\/h4>\n<p>\u4e00\u4e2a\u8bf7\u6c42\u4ece\u8fdb\u6765\u5230\u51fa\u53bb&#xff0c;\u8981\u7ecf\u5386\u8fd9\u51e0\u6b65&#xff1a;<\/p>\n<p>\u2460 \u5ba2\u6237\u7aef\u53d1\u8bf7\u6c42<br \/>\n   \u2502<br \/>\n   \u25bc<br \/>\n\u2461 Router \u63a5\u6536\u8bf7\u6c42<br \/>\n   &#8211; \u89e3\u6790 prompt \u957f\u5ea6\u3001SLA \u7b49\u7ea7\u3001\u4f1a\u8bdd\u4fe1\u606f<br \/>\n   &#8211; \u9009\u62e9\u4e00\u4e2a Prefill Worker<br \/>\n   \u2502<br \/>\n   \u25bc<br \/>\n\u2462 Prefill Worker \u6267\u884c prefill<br \/>\n   &#8211; \u8ba1\u7b97 prompt \u7684 KV Cache<br \/>\n   &#8211; \u751f\u6210\u7b2c\u4e00\u4e2a token \u7684 logits<br \/>\n   &#8211; \u51c6\u5907\u597d\u8981\u4f20\u8f93\u7684 KV blocks<br \/>\n   \u2502<br \/>\n   \u25bc<br \/>\n\u2463 KV Transfer<br \/>\n   &#8211; \u901a\u8fc7\u9ad8\u901f\u901a\u9053\u628a KV \u4f20\u5230 Decode Worker<br \/>\n   &#8211; Decode Worker \u628a KV \u6302\u8f7d\u5230\u81ea\u5df1\u7684\u5185\u5b58\u7a7a\u95f4<br \/>\n   \u2502<br \/>\n   \u25bc<br \/>\n\u2464 Decode Worker \u63a5\u7ba1<br \/>\n   &#8211; \u628a\u8bf7\u6c42\u52a0\u5165 continuous batch<br \/>\n   &#8211; \u5f00\u59cb\u81ea\u56de\u5f52\u751f\u6210<br \/>\n   &#8211; \u6d41\u5f0f\u8fd4\u56de token<br \/>\n   \u2502<br \/>\n   \u25bc<br \/>\n\u2465 \u751f\u6210\u7ed3\u675f<br \/>\n   &#8211; \u8fd4\u56de\u6700\u7ec8\u7ed3\u679c<br \/>\n   &#8211; \u91ca\u653e\u6216\u4fdd\u7559 KV&#xff08;\u53d6\u51b3\u4e8e\u4f1a\u8bdd\u7b56\u7565&#xff09;<\/p>\n<h4>5.2 Router\u7684\u89d2\u8272<\/h4>\n<p>Router\u662fPD\u89e3\u8026\u67b6\u6784\u7684&#034;\u5927\u8111&#034;&#xff0c;\u5b83\u8981\u505a\u5f88\u591a\u51b3\u7b56&#xff1a;<\/p>\n<li>Prefill Worker\u9009\u62e9&#xff1a;\u54ea\u4e2aPrefill Worker\u8d1f\u8f7d\u6700\u4f4e&#xff1f;\u54ea\u4e2a\u6709Prefix Cache\u547d\u4e2d&#xff1f;<\/li>\n<li>Decode Worker\u9009\u62e9&#xff1a;\u54ea\u4e2aDecode Worker\u8fd8\u6709\u4f59\u91cf&#xff1f;\u5982\u679c\u662f\u591a\u8f6e\u5bf9\u8bdd&#xff0c;\u662f\u4e0d\u662f\u8981\u56de\u5230\u540c\u4e00\u4e2aWorker&#xff1f;<\/li>\n<li>\u4f18\u5148\u7ea7\u8c03\u5ea6&#xff1a;\u9ad8\u4f18\u5148\u7ea7\u7684\u8bf7\u6c42\u662f\u4e0d\u662f\u8981\u63d2\u961f&#xff1f;<\/li>\n<li>\u5931\u8d25\u5904\u7406&#xff1a;Prefill Worker\u6302\u4e86\u600e\u4e48\u529e&#xff1f;KV\u4f20\u8f93\u5931\u8d25\u4e86\u600e\u4e48\u529e&#xff1f;<\/li>\n<h4>5.3 \u4ee3\u7801\u793a\u4f8b&#xff1a;\u7b80\u5316\u7684Router\u8c03\u5ea6\u903b\u8f91<\/h4>\n<p><span class=\"token keyword\">from<\/span> typing <span class=\"token keyword\">import<\/span> List<span class=\"token punctuation\">,<\/span> Dict<span class=\"token punctuation\">,<\/span> Optional<br \/>\n<span class=\"token keyword\">import<\/span> dataclasses<br \/>\n<span class=\"token keyword\">import<\/span> time<\/p>\n<p><span class=\"token decorator annotation punctuation\">&#064;dataclasses<span class=\"token punctuation\">.<\/span>dataclass<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">Request<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    request_id<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">str<\/span><br \/>\n    prompt_tokens<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span><br \/>\n    max_output_tokens<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span><br \/>\n    priority<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span> <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">1<\/span>  <span class=\"token comment\"># \u6570\u5b57\u8d8a\u5927\u4f18\u5148\u7ea7\u8d8a\u9ad8<\/span><br \/>\n    session_id<span class=\"token punctuation\">:<\/span> Optional<span class=\"token punctuation\">[<\/span><span class=\"token builtin\">str<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> <span class=\"token boolean\">None<\/span><\/p>\n<p><span class=\"token decorator annotation punctuation\">&#064;dataclasses<span class=\"token punctuation\">.<\/span>dataclass<\/span><br \/>\n<span class=\"token keyword\">class<\/span> <span class=\"token class-name\">WorkerInfo<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    worker_id<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">str<\/span><br \/>\n    gpu_type<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">str<\/span><br \/>\n    current_load<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">float<\/span>  <span class=\"token comment\"># 0-1&#xff0c;\u5f53\u524d\u8d1f\u8f7d<\/span><br \/>\n    kv_cache_used_ratio<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">float<\/span>  <span class=\"token comment\"># KV\u663e\u5b58\u4f7f\u7528\u7387<\/span><\/p>\n<p><span class=\"token keyword\">class<\/span> <span class=\"token class-name\">PDDisaggregationRouter<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token triple-quoted-string string\">&#034;&#034;&#034;PD \u89e3\u8026\u67b6\u6784\u4e2d\u7684 Router \u8c03\u5ea6\u5668&#034;&#034;&#034;<\/span><\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">__init__<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> prefill_workers<span class=\"token punctuation\">:<\/span> List<span class=\"token punctuation\">[<\/span>WorkerInfo<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span> decode_workers<span class=\"token punctuation\">:<\/span> List<span class=\"token punctuation\">[<\/span>WorkerInfo<span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>prefill_workers <span class=\"token operator\">&#061;<\/span> prefill_workers<br \/>\n        self<span class=\"token punctuation\">.<\/span>decode_workers <span class=\"token operator\">&#061;<\/span> decode_workers<br \/>\n        <span class=\"token comment\"># session \u4eb2\u548c\u6027&#xff1a;\u8bb0\u5f55 session \u4e0a\u6b21\u5728\u54ea\u4e2a decode worker<\/span><br \/>\n        self<span class=\"token punctuation\">.<\/span>session_affinity<span class=\"token punctuation\">:<\/span> Dict<span class=\"token punctuation\">[<\/span><span class=\"token builtin\">str<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token builtin\">str<\/span><span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">{<\/span><span class=\"token punctuation\">}<\/span><\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">select_prefill_worker<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> request<span class=\"token punctuation\">:<\/span> Request<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token operator\">&gt;<\/span> <span class=\"token builtin\">str<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token triple-quoted-string string\">&#034;&#034;&#034;<br \/>\n        \u9009\u62e9 Prefill Worker<\/p>\n<p>        \u7b56\u7565&#xff1a;<br \/>\n        1. \u4f18\u5148\u9009\u62e9\u8d1f\u8f7d\u4f4e\u7684<br \/>\n        2. \u8003\u8651 Prefix Cache \u547d\u4e2d&#xff08;\u7b80\u5316\u793a\u4f8b\u4e2d\u672a\u5b9e\u73b0&#xff09;<br \/>\n        3. \u957f prompt \u4f18\u5148\u5206\u914d\u5230\u7b97\u529b\u5f3a\u7684\u5361<br \/>\n        &#034;&#034;&#034;<\/span><br \/>\n        <span class=\"token comment\"># \u6309\u8d1f\u8f7d\u6392\u5e8f&#xff0c;\u9009\u8d1f\u8f7d\u6700\u4f4e\u7684<\/span><br \/>\n        sorted_workers <span class=\"token operator\">&#061;<\/span> <span class=\"token builtin\">sorted<\/span><span class=\"token punctuation\">(<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>prefill_workers<span class=\"token punctuation\">,<\/span><br \/>\n            key<span class=\"token operator\">&#061;<\/span><span class=\"token keyword\">lambda<\/span> w<span class=\"token punctuation\">:<\/span> w<span class=\"token punctuation\">.<\/span>current_load<br \/>\n        <span class=\"token punctuation\">)<\/span><\/p>\n<p>        <span class=\"token comment\"># \u957f prompt \u4f18\u5148\u9009 GPU \u5f3a\u7684&#xff08;\u8fd9\u91cc\u7b80\u5316\u5224\u65ad&#xff09;<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> request<span class=\"token punctuation\">.<\/span>prompt_tokens <span class=\"token operator\">&gt;<\/span> <span class=\"token number\">4096<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token comment\"># \u8fc7\u6ee4\u51fa A100\/H100 \u8fd9\u7c7b\u9ad8\u7aef\u5361<\/span><br \/>\n            high_end <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span>w <span class=\"token keyword\">for<\/span> w <span class=\"token keyword\">in<\/span> sorted_workers <span class=\"token keyword\">if<\/span> <span class=\"token string\">&#039;A100&#039;<\/span> <span class=\"token keyword\">in<\/span> w<span class=\"token punctuation\">.<\/span>gpu_type <span class=\"token keyword\">or<\/span> <span class=\"token string\">&#039;H100&#039;<\/span> <span class=\"token keyword\">in<\/span> w<span class=\"token punctuation\">.<\/span>gpu_type<span class=\"token punctuation\">]<\/span><br \/>\n            <span class=\"token keyword\">if<\/span> high_end<span class=\"token punctuation\">:<\/span><br \/>\n                <span class=\"token keyword\">return<\/span> high_end<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>worker_id<\/p>\n<p>        <span class=\"token keyword\">return<\/span> sorted_workers<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>worker_id<\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">select_decode_worker<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> request<span class=\"token punctuation\">:<\/span> Request<span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token operator\">&gt;<\/span> <span class=\"token builtin\">str<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token triple-quoted-string string\">&#034;&#034;&#034;<br \/>\n        \u9009\u62e9 Decode Worker<\/p>\n<p>        \u7b56\u7565&#xff1a;<br \/>\n        1. \u5982\u679c\u662f\u591a\u8f6e\u5bf9\u8bdd&#xff0c;\u4f18\u5148\u56de\u5230\u4e0a\u6b21\u7684 worker&#xff08;\u4f1a\u8bdd\u7c98\u6027&#xff09;<br \/>\n        2. \u9009\u62e9 KV \u663e\u5b58\u4f59\u91cf\u591a\u7684<br \/>\n        3. \u8f93\u51fa\u957f\u7684\u8bf7\u6c42\u5206\u914d\u5230\u8d1f\u8f7d\u4f4e\u7684<br \/>\n        &#034;&#034;&#034;<\/span><br \/>\n        <span class=\"token comment\"># \u4f1a\u8bdd\u4eb2\u548c\u6027<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> request<span class=\"token punctuation\">.<\/span>session_id <span class=\"token keyword\">and<\/span> request<span class=\"token punctuation\">.<\/span>session_id <span class=\"token keyword\">in<\/span> self<span class=\"token punctuation\">.<\/span>session_affinity<span class=\"token punctuation\">:<\/span><br \/>\n            preferred_id <span class=\"token operator\">&#061;<\/span> self<span class=\"token punctuation\">.<\/span>session_affinity<span class=\"token punctuation\">[<\/span>request<span class=\"token punctuation\">.<\/span>session_id<span class=\"token punctuation\">]<\/span><br \/>\n            preferred <span class=\"token operator\">&#061;<\/span> <span class=\"token punctuation\">[<\/span>w <span class=\"token keyword\">for<\/span> w <span class=\"token keyword\">in<\/span> self<span class=\"token punctuation\">.<\/span>decode_workers <span class=\"token keyword\">if<\/span> w<span class=\"token punctuation\">.<\/span>worker_id <span class=\"token operator\">&#061;&#061;<\/span> preferred_id<span class=\"token punctuation\">]<\/span><br \/>\n            <span class=\"token comment\"># \u5982\u679c\u76ee\u6807 worker \u8fd8\u80fd\u88c5\u4e0b&#xff0c;\u5c31\u56de\u53bb<\/span><br \/>\n            <span class=\"token keyword\">if<\/span> preferred <span class=\"token keyword\">and<\/span> preferred<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>kv_cache_used_ratio <span class=\"token operator\">&lt;<\/span> <span class=\"token number\">0.8<\/span><span class=\"token punctuation\">:<\/span><br \/>\n                <span class=\"token keyword\">return<\/span> preferred_id<\/p>\n<p>        <span class=\"token comment\"># \u9009 KV \u663e\u5b58\u6700\u5145\u88d5\u7684<\/span><br \/>\n        sorted_workers <span class=\"token operator\">&#061;<\/span> <span class=\"token builtin\">sorted<\/span><span class=\"token punctuation\">(<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>decode_workers<span class=\"token punctuation\">,<\/span><br \/>\n            key<span class=\"token operator\">&#061;<\/span><span class=\"token keyword\">lambda<\/span> w<span class=\"token punctuation\">:<\/span> w<span class=\"token punctuation\">.<\/span>kv_cache_used_ratio<br \/>\n        <span class=\"token punctuation\">)<\/span><\/p>\n<p>        selected <span class=\"token operator\">&#061;<\/span> sorted_workers<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>worker_id<\/p>\n<p>        <span class=\"token comment\"># \u8bb0\u5f55\u4f1a\u8bdd\u4eb2\u548c\u6027<\/span><br \/>\n        <span class=\"token keyword\">if<\/span> request<span class=\"token punctuation\">.<\/span>session_id<span class=\"token punctuation\">:<\/span><br \/>\n            self<span class=\"token punctuation\">.<\/span>session_affinity<span class=\"token punctuation\">[<\/span>request<span class=\"token punctuation\">.<\/span>session_id<span class=\"token punctuation\">]<\/span> <span class=\"token operator\">&#061;<\/span> selected<\/p>\n<p>        <span class=\"token keyword\">return<\/span> selected<\/p>\n<p>    <span class=\"token keyword\">def<\/span> <span class=\"token function\">estimate_kv_transfer_time<\/span><span class=\"token punctuation\">(<\/span>self<span class=\"token punctuation\">,<\/span> prompt_tokens<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span><span class=\"token punctuation\">,<\/span><br \/>\n                                   model_layers<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span><span class=\"token punctuation\">,<\/span> hidden_size<span class=\"token punctuation\">:<\/span> <span class=\"token builtin\">int<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#8211;<\/span><span class=\"token operator\">&gt;<\/span> <span class=\"token builtin\">float<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token triple-quoted-string string\">&#034;&#034;&#034;<br \/>\n        \u4f30\u7b97 KV \u4f20\u8f93\u65f6\u95f4<\/p>\n<p>        KV \u5927\u5c0f &#061; prompt_tokens * layers * hidden_size * 2 (K\u548cV) * 2 (FP16\u5b57\u8282\u6570)<br \/>\n        \u4f20\u8f93\u65f6\u95f4 &#061; KV\u5927\u5c0f \/ \u5e26\u5bbd<br \/>\n        &#034;&#034;&#034;<\/span><br \/>\n        kv_bytes <span class=\"token operator\">&#061;<\/span> prompt_tokens <span class=\"token operator\">*<\/span> model_layers <span class=\"token operator\">*<\/span> hidden_size <span class=\"token operator\">*<\/span> <span class=\"token number\">2<\/span> <span class=\"token operator\">*<\/span> <span class=\"token number\">2<\/span><br \/>\n        <span class=\"token comment\"># \u5047\u8bbe\u7528 NVLink&#xff0c;\u5e26\u5bbd 400GB\/s<\/span><br \/>\n        bandwidth_bytes_per_sec <span class=\"token operator\">&#061;<\/span> <span class=\"token number\">400<\/span> <span class=\"token operator\">*<\/span> <span class=\"token number\">1024<\/span><span class=\"token operator\">**<\/span><span class=\"token number\">3<\/span><br \/>\n        <span class=\"token keyword\">return<\/span> kv_bytes <span class=\"token operator\">\/<\/span> bandwidth_bytes_per_sec<\/p>\n<p>\u8fd9\u6bb5\u4ee3\u7801\u5c55\u793a\u4e86Router\u7684\u6838\u5fc3\u51b3\u7b56\u903b\u8f91\u3002\u5b9e\u9645\u751f\u4ea7\u4e2d\u7684Router\u8981\u590d\u6742\u5f97\u591a&#xff0c;\u8fd8\u8981\u8003\u8651\u5bb9\u9519\u3001\u9650\u6d41\u3001\u52a8\u6001\u6269\u7f29\u5bb9\u7b49\u7b49\u3002<\/p>\n<hr \/>\n<h3>\u516d\u3001\u4e0eChunked Prefill\u7684\u5173\u7cfb<\/h3>\n<h4>6.1 \u4ec0\u4e48\u662fChunked Prefill&#xff1f;<\/h4>\n<p>\u5728\u8bb2PD\u89e3\u8026\u4e4b\u524d&#xff0c;\u5148\u63d0\u4e00\u4e0b\u53e6\u4e00\u4e2a\u7f13\u89e3Prefill\/Decode\u5e72\u6270\u7684\u6280\u672f\u2014\u2014Chunked Prefill&#xff08;\u5206\u5757\u9884\u586b\u5145&#xff09;\u3002<\/p>\n<p>Chunked Prefill\u7684\u601d\u8def\u662f&#xff1a;\u628a\u957fprompt\u5206\u6210\u5c0f\u5757&#xff0c;\u4e00\u5757\u4e00\u5757\u5730\u505aprefill&#xff0c;\u4e2d\u95f4\u7a7f\u63d2decode\u6b65\u9aa4\u3002<\/p>\n<p>\u6ca1\u6709 Chunked Prefill:<br \/>\n[prefill (8192 tokens)] [decode] [decode] [decode] &#8230;<br \/>\n                        \u2191 \u7b49\u4e86\u5f88\u4e45\u624d\u5f00\u59cb decode<\/p>\n<p>\u6709 Chunked Prefill (chunk_size&#061;512):<br \/>\n[prefill 512] [decode] [prefill 512] [decode] [prefill 512] [decode] &#8230;<br \/>\n               \u2191 \u7b2c\u4e00\u4e2a decode \u5f88\u5feb\u5c31\u6765\u4e86<\/p>\n<p>\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f&#xff1a;<\/p>\n<ul>\n<li>TTFT\u53d8\u597d\u4e86&#xff08;\u4e0d\u7528\u7b49\u6574\u4e2aprefill\u505a\u5b8c&#xff09;<\/li>\n<li>Decode\u4e0d\u4f1a\u88ab\u957f\u65f6\u95f4\u963b\u585e&#xff08;\u6bcf\u505a\u4e00\u5757prefill\u5c31\u8ba9decode\u8dd1\u4e00\u6b65&#xff09;<\/li>\n<\/ul>\n<h4>6.2 \u5355\u5f15\u64ce\u5185\u7f13\u89e3 vs \u67b6\u6784\u5c42\u8d44\u6e90\u9694\u79bb<\/h4>\n<p>\u90a3Chunked Prefill\u548cPD\u89e3\u8026\u662f\u4ec0\u4e48\u5173\u7cfb&#xff1f;<\/p>\n<p>\u7b54\u6848\u662f&#xff1a;\u5b83\u4eec\u662f\u4e0d\u540c\u5c42\u9762\u7684\u89e3\u51b3\u65b9\u6848&#xff0c;\u53ef\u4ee5\u4e92\u8865\u3002<\/p>\n<table>\n<tr>\u65b9\u6848\u5c42\u9762\u6838\u5fc3\u601d\u60f3\u4f18\u70b9\u7f3a\u70b9<\/tr>\n<tbody>\n<tr>\n<td>Chunked Prefill<\/td>\n<td>\u5355\u5f15\u64ce\u5185<\/td>\n<td>\u628a\u957fprefill\u5207\u788e&#xff0c;\u548cdecode\u7a7f\u63d2\u6267\u884c<\/td>\n<td>\u5b9e\u73b0\u7b80\u5355&#xff0c;\u65e0\u9700\u989d\u5916\u673a\u5668<\/td>\n<td>\u53ea\u662f\u7f13\u89e3&#xff0c;prefill\u548cdecode\u8fd8\u662f\u4e92\u76f8\u62a2\u8d44\u6e90<\/td>\n<\/tr>\n<tr>\n<td>PD Disaggregation<\/td>\n<td>\u67b6\u6784\u5c42<\/td>\n<td>Prefill\u548cDecode\u7528\u4e0d\u540c\u7684GPU\u6c60&#xff0c;\u5b8c\u5168\u9694\u79bb<\/td>\n<td>\u5f7b\u5e95\u89e3\u51b3\u5e72\u6270&#xff0c;\u4e24\u8fb9\u5404\u81ea\u4f18\u5316<\/td>\n<td>\u67b6\u6784\u590d\u6742&#xff0c;\u6709KV\u4f20\u8f93\u5f00\u9500<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u53ef\u4ee5\u8fd9\u4e48\u7406\u89e3&#xff1a;<\/p>\n<ul>\n<li>Chunked Prefill\u662f&#034;\u65f6\u95f4\u5206\u7247&#034;\u2014\u2014\u4f60\u7528\u4e00\u4f1a\u513f&#xff0c;\u6211\u7528\u4e00\u4f1a\u513f<\/li>\n<li>PD\u89e3\u8026\u662f&#034;\u7a7a\u95f4\u5206\u7247&#034;\u2014\u2014\u4f60\u7528\u4f60\u7684\u673a\u5668&#xff0c;\u6211\u7528\u6211\u7684\u673a\u5668<\/li>\n<\/ul>\n<p>Chunked Prefill\u662f\u4f4e\u6210\u672c\u7684\u7f13\u89e3\u65b9\u6848&#xff0c;PD\u89e3\u8026\u662f\u5f7b\u5e95\u7684\u67b6\u6784\u65b9\u6848\u3002\u5728\u5b9e\u9645\u751f\u4ea7\u4e2d&#xff0c;\u751a\u81f3\u53ef\u4ee5\u540c\u65f6\u4f7f\u7528\u2014\u2014Prefill Worker\u6c60\u5185\u90e8\u4e5f\u53ef\u4ee5\u7528Chunked Prefill\u6765\u4f18\u5316TTFT\u3002<\/p>\n<hr \/>\n<h3>\u4e03\u3001\u6838\u5fc3\u4ee3\u4ef7&#xff1a;KV Transfer<\/h3>\n<h4>7.1 KV Transfer\u662fPD\u89e3\u8026\u7684&#034;\u963f\u5580\u7409\u65af\u4e4b\u8e35&#034;<\/h4>\n<p>PD\u89e3\u8026\u7684\u4e00\u5207\u597d\u5904&#xff0c;\u90fd\u5efa\u7acb\u5728\u4e00\u4e2a\u524d\u63d0\u4e0a&#xff1a;KV\u4f20\u8f93\u7684\u65f6\u95f4 &lt; \u89e3\u8026\u5e26\u6765\u7684\u6536\u76ca\u3002<\/p>\n<p>\u5982\u679cKV\u4f20\u8f93\u592a\u6162&#xff0c;\u90a3prefill\u5b8c\u4e86\u8fd8\u8981\u7b49\u534a\u5929\u4f20KV&#xff0c;TTFT\u53cd\u800c\u66f4\u5dee&#xff0c;\u90a3\u5c31\u5f97\u4e0d\u507f\u5931\u4e86\u3002<\/p>\n<h4>7.2 KV\u6570\u636e\u91cf\u516c\u5f0f<\/h4>\n<p>KV Cache\u6709\u591a\u5927&#xff1f;\u516c\u5f0f\u5982\u4e0b&#xff1a;<\/p>\n<p><span class=\"katex--display\"><span class=\"katex-display\"><span class=\"katex\"><span class=\"katex-mathml\">KV\u5927\u5c0f&#xff08;\u5b57\u8282&#xff09;&#061;prompt_tokens\u00d7L\u00d7dmodel\u00d72\u00d7bytes_per_element<br \/>\n\\\\text{KV\u5927\u5c0f&#xff08;\u5b57\u8282&#xff09;} &#061; \\\\text{prompt\\\\_tokens} \\\\times L \\\\times d_{model} \\\\times 2 \\\\times \\\\text{bytes\\\\_per\\\\_element}<br \/>\n<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord text\"><span class=\"mord\">KV<\/span><span class=\"mord cjk_fallback\">\u5927\u5c0f&#xff08;\u5b57\u8282&#xff09;<\/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: 1.0044em;vertical-align: -0.31em\"><\/span><span class=\"mord text\"><span class=\"mord\">prompt_tokens<\/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: 0.7667em;vertical-align: -0.0833em\"><\/span><span class=\"mord mathnormal\">L<\/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: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">d<\/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: 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\">m<\/span><span class=\"mord mathnormal mtight\">o<\/span><span class=\"mord mathnormal mtight\">d<\/span><span class=\"mord mathnormal mtight\">e<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0197em\">l<\/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 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: 0.7278em;vertical-align: -0.0833em\"><\/span><span class=\"mord\">2<\/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: 1.0044em;vertical-align: -0.31em\"><\/span><span class=\"mord text\"><span class=\"mord\">bytes_per_element<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n<p>\u5176\u4e2d&#xff1a;<\/p>\n<ul>\n<li><span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">LL<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.6833em\"><\/span><span class=\"mord mathnormal\">L<\/span><\/span><\/span><\/span><\/span> &#061; Transformer\u5c42\u6570<\/li>\n<li><span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">dmodeld_{model}<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.8444em;vertical-align: -0.15em\"><\/span><span class=\"mord\"><span class=\"mord mathnormal\">d<\/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: 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\">m<\/span><span class=\"mord mathnormal mtight\">o<\/span><span class=\"mord mathnormal mtight\">d<\/span><span class=\"mord mathnormal mtight\">e<\/span><span class=\"mord mathnormal mtight\" style=\"margin-right: 0.0197em\">l<\/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> &#061; \u9690\u85cf\u5c42\u7ef4\u5ea6&#xff08;head_dim * num_heads&#xff09;<\/li>\n<li><span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">\u00d72\\\\times 2<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 0.7278em;vertical-align: -0.0833em\"><\/span><span class=\"mord\">\u00d7<\/span><span class=\"mord\">2<\/span><\/span><\/span><\/span><\/span> &#061; K\u548cV\u4e24\u4efd<\/li>\n<li><span class=\"katex--inline\"><span class=\"katex\"><span class=\"katex-mathml\">bytes_per_element\\\\text{bytes\\\\_per\\\\_element}<\/span><span class=\"katex-html\"><span class=\"base\"><span class=\"strut\" style=\"height: 1.0044em;vertical-align: -0.31em\"><\/span><span class=\"mord text\"><span class=\"mord\">bytes_per_element<\/span><\/span><\/span><\/span><\/span><\/span> &#061; \u6bcf\u4e2a\u5143\u7d20\u7684\u5b57\u8282\u6570&#xff08;FP16\u662f2&#xff0c;INT8\u662f1&#xff0c;INT4\u662f0.5&#xff09;<\/li>\n<\/ul>\n<p>\u4e3e\u4e2a\u4f8b\u5b50&#xff1a;Llama-2-70B&#xff0c;FP16\u7cbe\u5ea6<\/p>\n<ul>\n<li>L &#061; 80\u5c42<\/li>\n<li>d_model &#061; 128 * 8 * 1024&#xff1f;\u4e0d\u5bf9&#xff0c;\u5e94\u8be5\u662f 8192&#xff08;hidden size&#xff09;<\/li>\n<\/ul>\n<p>\u8ba9\u6211\u91cd\u65b0\u7b97&#xff1a;Llama-2-70B&#xff0c;hidden_size&#061;8192&#xff0c;num_layers&#061;80<\/p>\n<ul>\n<li>KV\u5927\u5c0f per token &#061; 80\u5c42 \u00d7 8192 \u00d7 2(K&#043;V) \u00d7 2\u5b57\u8282(FP16) &#061; 80 \u00d7 8192 \u00d7 4 &#061; 2,621,440 \u5b57\u8282 \u2248 2.5MB\/token&#xff1f;\u4e0d\u5bf9&#xff0c;\u8fd9\u592a\u5927\u4e86<\/li>\n<\/ul>\n<p>\u8ba9\u6211\u91cd\u65b0\u601d\u8003\u3002KV cache per layer per token\u7684\u5927\u5c0f\u5e94\u8be5\u662f&#xff1a;<\/p>\n<ul>\n<li>K: num_kv_heads \u00d7 head_dim \u00d7 bytes_per_element<\/li>\n<li>V: num_kv_heads \u00d7 head_dim \u00d7 bytes_per_element<\/li>\n<\/ul>\n<p>Llama-2-70B: num_kv_heads&#061;8&#xff08;GQA&#xff09;&#xff0c;head_dim&#061;128<\/p>\n<ul>\n<li>Per layer per token: 2 \u00d7 8 \u00d7 128 \u00d7 2 &#061; 4096 bytes &#061; 4KB<\/li>\n<li>80\u5c42: 80 \u00d7 4KB &#061; 320KB\/token<\/li>\n<\/ul>\n<p>1024 token \u7684 KV &#061; 1024 \u00d7 320KB &#061; 320MB<\/p>\n<p>\u55ef&#xff0c;\u8fd9\u6837\u6bd4\u8f83\u5408\u7406\u3002\u8ba9\u6211\u5728\u6587\u7ae0\u4e2d\u7528\u6b63\u786e\u7684\u8ba1\u7b97\u3002<\/p>\n<h4>7.3 \u4f20\u8f93\u65f6\u95f4\u4f30\u7b97<\/h4>\n<p>\u6211\u4eec\u6765\u7b97\u51e0\u4e2a\u5b9e\u9645\u7684\u4f8b\u5b50&#xff1a;<\/p>\n<p>\u573a\u666f1&#xff1a;7B\u6a21\u578b&#xff0c;2048 token prompt&#xff0c;FP16<\/p>\n<ul>\n<li>KV\u5927\u5c0f \u2248 2048 \u00d7 32\u5c42 \u00d7 128\u00d732 \u00d7 2 \u00d7 2\u5b57\u8282 &#061; 2048 \u00d7 32 \u00d7 4096 \u00d7 2&#xff1f;\u8ba9\u6211\u91cd\u65b0\u7ec4\u7ec7<\/li>\n<\/ul>\n<p>\u76f4\u63a5\u7ed9\u7ed3\u8bba\u5427&#xff0c;\u5199\u5f97\u6e05\u695a\u4e00\u4e9b&#xff1a;<\/p>\n<table>\n<tr>\u6a21\u578bPrompt\u957f\u5ea6KV\u5927\u5c0f&#xff08;FP16&#xff09;NVLink&#xff08;400GB\/s&#xff09;PCIe 4.0 x16&#xff08;32GB\/s&#xff09;RDMA&#xff08;100Gbps&#xff09;<\/tr>\n<tbody>\n<tr>\n<td>7B<\/td>\n<td>2K tokens<\/td>\n<td>~64MB<\/td>\n<td>0.16ms<\/td>\n<td>2ms<\/td>\n<td>5ms<\/td>\n<\/tr>\n<tr>\n<td>7B<\/td>\n<td>32K tokens<\/td>\n<td>~1GB<\/td>\n<td>2.5ms<\/td>\n<td>31ms<\/td>\n<td>80ms<\/td>\n<\/tr>\n<tr>\n<td>70B<\/td>\n<td>2K tokens<\/td>\n<td>~640MB<\/td>\n<td>1.6ms<\/td>\n<td>20ms<\/td>\n<td>51ms<\/td>\n<\/tr>\n<tr>\n<td>70B<\/td>\n<td>32K tokens<\/td>\n<td>~10GB<\/td>\n<td>25ms<\/td>\n<td>312ms<\/td>\n<td>800ms<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u53ef\u4ee5\u770b\u5230&#xff1a;<\/p>\n<ul>\n<li>\u540c\u673aNVLink&#xff1a;\u4f20\u8f93\u65f6\u95f4\u5f88\u77ed&#xff0c;\u51e0\u4e4e\u53ef\u4ee5\u5ffd\u7565<\/li>\n<li>\u8de8\u673aPCIe\/RDMA&#xff1a;\u77edprompt\u8fd8\u597d&#xff0c;\u957fprompt\u4f20\u8f93\u65f6\u95f4\u5c31\u5f88\u53ef\u89c2\u4e86<\/li>\n<li>70B &#043; 32K &#043; RDMA&#xff1a;800ms\u7684\u4f20\u8f93\u65f6\u95f4&#xff0c;\u5982\u679cTTFT\u76ee\u6807\u662f1s\u7684\u8bdd&#xff0c;\u4f20\u8f93\u5c31\u5360\u4e8680%<\/li>\n<\/ul>\n<h4>7.4 \u964d\u4f4e\u4f20\u8f93\u6210\u672c\u7684\u65b9\u6cd5<\/h4>\n<li>KV\u91cf\u5316&#xff1a;\u628aKV\u4eceFP16\u964d\u5230INT8\u6216INT4&#xff0c;\u4f20\u8f93\u91cf\u76f4\u63a5\u51cf\u534a\u6216\u51cf\u52301\/4<\/li>\n<li>\u540c\u673a\u90e8\u7f72&#xff1a;Prefill\u548cDecode\u653e\u5728\u540c\u4e00\u53f0\u673a\u5668\u7684\u4e0d\u540cGPU\u4e0a&#xff0c;\u7528NVLink\/PCIe\u4f20\u8f93&#xff0c;\u901f\u5ea6\u5feb\u5f88\u591a<\/li>\n<li>\u589e\u91cf\u4f20\u8f93&#xff1a;\u591a\u8f6e\u5bf9\u8bdd\u53ea\u4f20\u65b0\u589e\u7684KV&#xff0c;\u4e0d\u4f20\u5168\u91cf<\/li>\n<li>Prefix Cache &#043; PD\u89e3\u8026\u534f\u540c&#xff1a;\u5982\u679cprefix\u547d\u4e2d\u4e86\u7f13\u5b58&#xff0c;\u76f4\u63a5\u4eceDecode Worker\u7684\u7f13\u5b58\u91cc\u62ff&#xff0c;\u4e0d\u7528\u4f20<\/li>\n<li>\u9884\u4f20\u8f93&#xff1a;\u9884\u6d4b\u54ea\u4e2aDecode Worker\u4f1a\u5904\u7406\u8fd9\u4e2a\u8bf7\u6c42&#xff0c;prefill\u8fd8\u6ca1\u505a\u5b8c\u5c31\u5f00\u59cb\u4f20<\/li>\n<hr \/>\n<h3>\u516b\u3001\u8c03\u5ea6\u4e0e\u8def\u7531\u8bbe\u8ba1<\/h3>\n<h4>8.1 Router\u9700\u8981\u8003\u8651\u4ec0\u4e48&#xff1f;<\/h4>\n<p>Router\u662fPD\u89e3\u8026\u67b6\u6784\u4e2d\u6700&#034;\u70e7\u8111&#034;\u7684\u7ec4\u4ef6\u3002\u5b83\u9700\u8981\u7efc\u5408\u8003\u8651\u5f88\u591a\u56e0\u7d20&#xff1a;<\/p>\n<li>\u8d1f\u8f7d\u5747\u8861&#xff1a;\u4e0d\u80fd\u8ba9\u67d0\u4e2aWorker\u7d2f\u6b7b&#xff0c;\u53e6\u4e00\u4e2a\u95f2\u6b7b<\/li>\n<li>\u4f1a\u8bdd\u4eb2\u548c\u6027&#xff1a;\u591a\u8f6e\u5bf9\u8bdd\u5c3d\u91cf\u56de\u5230\u540c\u4e00\u4e2aDecode Worker&#xff0c;\u907f\u514dKV\u642c\u6765\u642c\u53bb<\/li>\n<li>Prefix Cache\u4eb2\u548c\u6027&#xff1a;\u76f8\u540c\u524d\u7f00\u7684\u8bf7\u6c42\u5c3d\u91cf\u53d1\u5230\u540c\u4e00\u4e2aPrefill Worker&#xff0c;\u63d0\u9ad8\u7f13\u5b58\u547d\u4e2d\u7387<\/li>\n<li>SLO\u4f18\u5148\u7ea7&#xff1a;\u9ad8\u4f18\u5148\u7ea7\u7684\u8bf7\u6c42\u662f\u4e0d\u662f\u8981\u63d2\u961f\u3001\u662f\u4e0d\u662f\u8981\u5206\u914d\u5230\u66f4\u597d\u7684\u8d44\u6e90<\/li>\n<li>\u5f02\u6784GPU\u8c03\u5ea6&#xff1a;\u5982\u679cGPU\u578b\u53f7\u4e0d\u4e00\u6837&#xff0c;\u600e\u4e48\u5206\u914d\u4efb\u52a1&#xff08;\u6bd4\u5982\u5927\u6a21\u578bPrefill\u7528A100&#xff0c;\u5c0f\u6a21\u578b\u7528L4&#xff09;<\/li>\n<li>\u6545\u969c\u8f6c\u79fb&#xff1a;\u67d0\u4e2aWorker\u6302\u4e86&#xff0c;\u8bf7\u6c42\u600e\u4e48\u5feb\u901f\u5207\u8d70<\/li>\n<h4>8.2 \u5178\u578b\u8c03\u5ea6\u7b56\u7565<\/h4>\n<p>Prefill Worker\u9009\u62e9\u7b56\u7565&#xff1a;<\/p>\n<li>Least Load&#xff08;\u6700\u5c11\u8d1f\u8f7d&#xff09;&#xff1a;\u9009\u5f53\u524dprefill\u961f\u5217\u6700\u77ed\u7684Worker\u2014\u2014\u6700\u6734\u7d20\u4e5f\u6700\u5e38\u7528<\/li>\n<li>Prefix Cache Affinity&#xff08;\u7f13\u5b58\u4eb2\u548c&#xff09;&#xff1a;\u76f8\u540c\u524d\u7f00\u7684\u8bf7\u6c42\u53d1\u5230\u540c\u4e00\u4e2aWorker\u2014\u2014\u727a\u7272\u4e00\u70b9\u5747\u8861&#xff0c;\u6362\u66f4\u9ad8\u7684\u7f13\u5b58\u547d\u4e2d\u7387<\/li>\n<li>Prompt Length Aware&#xff08;\u957f\u5ea6\u611f\u77e5&#xff09;&#xff1a;\u957fprompt\u5206\u914d\u5230\u7b97\u529b\u5f3a\u7684\u5361&#xff0c;\u77edprompt\u5206\u914d\u5230\u4e00\u822c\u7684\u5361<\/li>\n<p>Decode Worker\u9009\u62e9\u7b56\u7565&#xff1a;<\/p>\n<li>Least KV Usage&#xff08;\u6700\u5c11KV\u5360\u7528&#xff09;&#xff1a;\u9009KV\u663e\u5b58\u6700\u5145\u88d5\u7684Worker\u2014\u2014\u9632\u6b62OOM<\/li>\n<li>Session Affinity&#xff08;\u4f1a\u8bdd\u7c98\u6027&#xff09;&#xff1a;\u540c\u4e00\u7528\u6237\u7684\u591a\u8f6e\u5bf9\u8bdd\u56de\u5230\u540c\u4e00\u4e2aWorker\u2014\u2014\u907f\u514dKV\u8fc1\u79fb<\/li>\n<li>Least Outstanding Requests&#xff08;\u6700\u5c11\u5f85\u5904\u7406\u8bf7\u6c42&#xff09;&#xff1a;\u9009\u5f53\u524d\u8bf7\u6c42\u6570\u6700\u5c11\u7684\u2014\u2014\u4fdd\u8bc1TPOT\u7a33\u5b9a<\/li>\n<h4>8.3 \u4e0e\u5176\u4ed6\u6280\u672f\u534f\u540c<\/h4>\n<p>PD\u89e3\u8026\u4e0d\u662f\u5b64\u7acb\u7684&#xff0c;\u5b83\u53ef\u4ee5\u548c\u6211\u4eec\u4e4b\u524d\u5b66\u7684\u5f88\u591a\u6280\u672f\u7ed3\u5408&#xff1a;<\/p>\n<ul>\n<li>&#043; Prefix Cache&#xff1a;Prefill Worker\u7684\u7f13\u5b58\u547d\u4e2d\u7387\u66f4\u9ad8&#xff0c;\u800c\u4e14\u53ef\u4ee5\u51cf\u5c11KV\u4f20\u8f93\u91cf<\/li>\n<li>&#043; Speculative Decoding&#xff1a;Decode Worker\u53ef\u4ee5\u7528\u6295\u673a\u89e3\u7801\u8fdb\u4e00\u6b65\u52a0\u901f<\/li>\n<li>&#043; Continuous Batching&#xff1a;Decode Worker\u5185\u90e8\u8fd8\u662f\u7528Continuous Batching\u6765\u63d0\u9ad8\u5229\u7528\u7387<\/li>\n<li>&#043; Chunked Prefill&#xff1a;Prefill Worker\u5185\u90e8\u53ef\u4ee5\u7528Chunked Prefill\u4f18\u5316TTFT<\/li>\n<\/ul>\n<hr \/>\n<h3>\u4e5d\u3001\u9002\u7528\u573a\u666f\u5224\u65ad\u4e0e\u5c0f\u7ed3<\/h3>\n<h4>9.1 \u4ec0\u4e48\u65f6\u5019\u9002\u5408\u4e0aPD\u89e3\u8026&#xff1f;<\/h4>\n<p>PD\u89e3\u8026\u4e0d\u662f\u94f6\u5f39&#xff0c;\u5b83\u6709\u989d\u5916\u7684\u590d\u6742\u5ea6\u548c\u6210\u672c\u3002\u4e0b\u9762\u8fd9\u4e9b\u573a\u666f\u503c\u5f97\u8003\u8651&#xff1a;<\/p>\n<li>Prompt\u957f\u5ea6\u5dee\u5f02\u5927&#xff1a;\u6709\u7684\u8bf7\u6c42\u53ea\u6709\u51e0\u5341\u4e2atoken&#xff0c;\u6709\u7684\u6709\u51e0\u4e07token\u3002\u8026\u5408\u90e8\u7f72\u65f6\u957fprompt\u4f1a\u4e25\u91cd\u5e72\u6270\u77ed\u8bf7\u6c42\u3002<\/li>\n<li>\u5bf9TPOT\u7a33\u5b9a\u6027\u8981\u6c42\u9ad8&#xff1a;\u6bd4\u5982\u5b9e\u65f6\u5bf9\u8bdd\u3001AI\u966a\u804a\u3001\u6e38\u620fNPC\u7b49\u573a\u666f&#xff0c;TPOT\u6296\u52a8\u4f1a\u4e25\u91cd\u5f71\u54cd\u4f53\u9a8c\u3002<\/li>\n<li>Prefill\u5360\u6bd4\u9ad8&#xff1a;\u6bd4\u5982RAG\u3001\u957f\u6587\u6863\u603b\u7ed3\u7b49\u573a\u666f&#xff0c;prefill\u65f6\u95f4\u5360\u6bd4\u5f88\u9ad8&#xff0c;\u503c\u5f97\u4e13\u95e8\u4f18\u5316\u3002<\/li>\n<li>\u6d41\u91cf\u6709\u660e\u663e\u7684&#034;\u7a81\u53d1\u957f\u8bf7\u6c42&#034;&#xff1a;\u6bd4\u5982\u5076\u5c14\u6709\u4eba\u53d1\u8d85\u957f\u6587\u6863&#xff0c;\u5e73\u65f6\u90fd\u662f\u77ed\u5bf9\u8bdd\u3002<\/li>\n<li>GPU\u8d44\u6e90\u5145\u8db3&#xff1a;\u6709\u8db3\u591f\u7684\u673a\u5668\u53ef\u4ee5\u5206\u6210\u4e24\u4e2a\u6c60\u3002<\/li>\n<h4>9.2 \u4ec0\u4e48\u65f6\u5019\u4e0d\u9002\u5408&#xff1f;<\/h4>\n<li>\u673a\u5668\u5f88\u5c11&#xff1a;\u4e00\u5171\u5c312-3\u5f20\u5361&#xff0c;\u62c6\u5f00\u4e86\u6bcf\u4e2a\u6c60\u53ea\u67091\u5f20\u5361&#xff0c;\u5bb9\u9519\u90fd\u505a\u4e0d\u4e86\u3002<\/li>\n<li>Prompt\u90fd\u5f88\u77ed&#xff1a;\u6bd4\u5982\u5206\u7c7b\u3001\u6458\u8981\u7b49\u573a\u666f&#xff0c;prompt\u53ea\u6709\u51e0\u5341\u4e2a\u5b57&#xff0c;prefill\u672c\u6765\u5c31\u5f88\u5feb\u3002<\/li>\n<li>KV\u4f20\u8f93\u6210\u672c\u592a\u9ad8&#xff1a;\u6bd4\u5982\u8de8\u673a\u623f\u90e8\u7f72&#xff0c;\u7f51\u7edc\u5e26\u5bbd\u4e0d\u591f&#xff0c;\u4f20\u8f93\u65f6\u95f4\u6bd4\u6536\u76ca\u8fd8\u5927\u3002<\/li>\n<li>\u56e2\u961f\u89c4\u6a21\u5c0f&#xff1a;PD\u89e3\u8026\u7684\u5de5\u7a0b\u590d\u6742\u5ea6\u5f88\u9ad8&#xff0c;\u5c0f\u56e2\u961f\u53ef\u80fdhold\u4e0d\u4f4f\u3002<\/li>\n<h4>9.3 \u5c0f\u7ed3<\/h4>\n<p>Prefill\/Decode\u89e3\u8026\u662fLLM Serving\u67b6\u6784\u6f14\u8fdb\u7684\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\u3002\u5b83\u4ece\u67b6\u6784\u5c42\u9762\u628a\u4e24\u7c7b\u8d44\u6e90\u7279\u5f81\u5b8c\u5168\u4e0d\u540c\u7684\u4efb\u52a1\u5206\u5f00&#xff0c;\u8ba9&#034;\u4e13\u4e1a\u7684\u4eba\u505a\u4e13\u4e1a\u7684\u4e8b&#034;\u3002<\/p>\n<p>\u6838\u5fc3\u8981\u70b9\u56de\u987e&#xff1a;<\/p>\n<li>\u672c\u8d28\u5dee\u5f02&#xff1a;Prefill\u662f\u8ba1\u7b97\u5bc6\u96c6\u578b\u6279\u91cf\u4efb\u52a1&#xff0c;Decode\u662f\u5e26\u5bbd\u5bc6\u96c6\u578b\u6d41\u5f0f\u4efb\u52a1&#xff0c;\u4e24\u8005\u8d44\u6e90\u74f6\u9888\u548c\u8c03\u5ea6\u76ee\u6807\u5b8c\u5168\u4e0d\u540c\u3002<\/li>\n<li>\u8026\u5408\u95ee\u9898&#xff1a;\u957fprefill\u963b\u585edecode\u3001TTFT\u53d8\u5dee\u3001TPOT\u6296\u52a8\u3001SLO\u96be\u505a\u3002<\/li>\n<li>\u89e3\u8026\u67b6\u6784&#xff1a;Prefill Worker Pool &#043; KV Transfer &#043; Decode Worker Pool&#xff0c;Router\u7edf\u4e00\u8c03\u5ea6\u3002<\/li>\n<li>\u8bf7\u6c42\u751f\u547d\u5468\u671f&#xff1a;Router\u2192Prefill\u2192KV Transfer\u2192Decode\u2192\u8fd4\u56de\u3002<\/li>\n<li>\u6838\u5fc3\u4ee3\u4ef7&#xff1a;KV Transfer\u3002\u540c\u673aNVLink\u5f88\u5feb&#xff0c;\u8de8\u673aRDMA\u9700\u8981\u6743\u8861\u3002\u53ef\u4ee5\u901a\u8fc7KV\u91cf\u5316\u3001Prefix Cache\u534f\u540c\u7b49\u65b9\u5f0f\u964d\u4f4e\u6210\u672c\u3002<\/li>\n<li>\u8c03\u5ea6\u8bbe\u8ba1&#xff1a;Router\u9700\u8981\u8003\u8651\u8d1f\u8f7d\u5747\u8861\u3001\u4f1a\u8bdd\u4eb2\u548c\u6027\u3001\u7f13\u5b58\u4eb2\u548c\u6027\u3001SLO\u4f18\u5148\u7ea7\u7b49\u591a\u4e2a\u7ef4\u5ea6\u3002<\/li>\n<li>\u9002\u7528\u573a\u666f&#xff1a;prompt\u5dee\u5f02\u5927\u3001TPOT\u7a33\u5b9a\u6027\u8981\u6c42\u9ad8\u3001prefill\u5360\u6bd4\u9ad8\u7684\u573a\u666f\u6536\u76ca\u6700\u5927\u3002<\/li>\n<h4>9.4 \u601d\u8003\u95ee\u9898<\/h4>\n<li>PD\u89e3\u8026\u548cChunked Prefill\u90fd\u662f\u89e3\u51b3Prefill\/Decode\u5e72\u6270\u7684\u65b9\u6848&#xff0c;\u5b83\u4eec\u53ef\u4ee5\u540c\u65f6\u4f7f\u7528\u5417&#xff1f;\u5982\u679c\u53ef\u4ee5&#xff0c;\u5728\u4ec0\u4e48\u5c42\u9762\u7ed3\u5408&#xff1f;<\/li>\n<li>\u5047\u8bbe\u4f60\u7684\u4e1a\u52a1\u662f\u4e00\u4e2aAI\u5199\u4f5c\u52a9\u624b&#xff0c;\u7528\u6237prompt\u901a\u5e38\u5f88\u77ed&#xff08;100-200 token&#xff09;&#xff0c;\u4f46\u8f93\u51fa\u5f88\u957f&#xff08;2000&#043; token&#xff09;\u3002PD\u89e3\u8026\u7684\u6536\u76ca\u5927\u5417&#xff1f;\u4e3a\u4ec0\u4e48&#xff1f;<\/li>\n<li>KV\u4f20\u8f93\u662fPD\u89e3\u8026\u7684\u6838\u5fc3\u4ee3\u4ef7\u3002\u9664\u4e86\u6587\u4e2d\u63d0\u5230\u7684KV\u91cf\u5316\u3001\u540c\u673a\u90e8\u7f72\u3001Prefix Cache\u534f\u540c&#xff0c;\u4f60\u8fd8\u80fd\u60f3\u5230\u54ea\u4e9b\u964d\u4f4eKV\u4f20\u8f93\u6210\u672c\u7684\u65b9\u6cd5&#xff1f;<\/li>\n<li>\u5982\u679cPrefill Worker\u548cDecode Worker\u7684GPU\u578b\u53f7\u4e0d\u4e00\u6837&#xff08;\u6bd4\u5982Prefill\u7528A100&#xff0c;Decode\u7528L40S&#xff09;&#xff0c;\u4f1a\u6709\u4ec0\u4e48\u95ee\u9898&#xff1f;KV\u8fd8\u80fd\u76f4\u63a5\u4f20\u8f93\u5417&#xff1f;<\/li>\n<li>PD\u89e3\u8026\u67b6\u6784\u4e0b&#xff0c;\u591a\u8f6e\u5bf9\u8bdd\u7684KV Cache\u5e94\u8be5\u600e\u4e48\u7ba1\u7406&#xff1f; \u662f\u4e00\u76f4\u653e\u5728Decode Worker\u4e0a&#xff0c;\u8fd8\u662f\u5bf9\u8bdd\u7ed3\u675f\u5c31\u4e22&#xff1f;\u5982\u679c\u7528\u6237\u9694\u4e86\u5f88\u4e45\u56de\u6765\u63a5\u7740\u804a&#xff0c;KV\u53bb\u54ea\u4e86&#xff1f;<\/li>\n<hr \/>\n<p>\u5982\u679c\u4f60\u89c9\u5f97\u8fd9\u7bc7\u6587\u7ae0\u6709\u5e2e\u52a9&#xff0c;\u6b22\u8fce\u70b9\u8d5e\u6536\u85cf\u5173\u6ce8\u4e09\u8fde~ \u81f3\u6b64&#xff0c;LLM\u63a8\u7406\u4e0eServing\u7cfb\u5217\u7684\u6838\u5fc3\u6280\u672f\u5c31\u90fd\u804a\u5b8c\u4e86\u3002\u540e\u9762\u6211\u4eec\u8fd8\u4f1a\u7ee7\u7eed\u804a\u4e3b\u6d41\u63a8\u7406\u6846\u67b6\u5bf9\u6bd4\u3001\u591a\u673a\u591a\u5361Serving\u3001\u7aef\u4fa7\u63a8\u7406\u7b49\u8bdd\u9898&#xff0c;\u656c\u8bf7\u671f\u5f85&#xff01;<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784&#xff1a;\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42\u5927\u5bb6\u597d&#xff0c;\u6211\u662f\u4f60\u7684AI Infra\u5b66\u4e60\u642d\u5b50\u3002\u524d\u9762\u6211\u4eec\u804a\u4e86PagedAttention\u3001Continuous Batching\u3001Speculative Decoding\u548cPrefix Cache&#xff0c;\u8fd9\u4e9b\u4f18\u5316\u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u524d\u63d0\u2014\u2014Prefill\u548cDecode\u8dd1\u5728\u540c\u4e00\u4e2aGPU\u4e0a\u3002\u4f46\u5728\u751f\u4ea7\u73af\u5883\u4e2d&#xff0c;Prefill\u548cDecode\u5176\u5b9e\u662f\u4e24\u7c7b<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1699,52,50,188,51],"topic":[],"class_list":["post-123854","post","type-post","status-publish","format-standard","hentry","category-server","tag-agi","tag-aigc","tag-50","tag-188","tag-51"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \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\/123854.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784&#xff1a;\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42\u5927\u5bb6\u597d&#xff0c;\u6211\u662f\u4f60\u7684AI Infra\u5b66\u4e60\u642d\u5b50\u3002\u524d\u9762\u6211\u4eec\u804a\u4e86PagedAttention\u3001Continuous Batching\u3001Speculative Decoding\u548cPrefix Cache&#xff0c;\u8fd9\u4e9b\u4f18\u5316\u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u524d\u63d0\u2014\u2014Prefill\u548cDecode\u8dd1\u5728\u540c\u4e00\u4e2aGPU\u4e0a\u3002\u4f46\u5728\u751f\u4ea7\u73af\u5883\u4e2d&#xff0c;Prefill\u548cDecode\u5176\u5b9e\u662f\u4e24\u7c7b\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/123854.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-08T01:32:56+00:00\" \/>\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=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/123854.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/123854.html\",\"name\":\"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-11-08T01:32:56+00:00\",\"dateModified\":\"2026-11-08T01:32:56+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/123854.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/123854.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/123854.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42\"}]},{\"@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":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \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\/123854.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784&#xff1a;\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42\u5927\u5bb6\u597d&#xff0c;\u6211\u662f\u4f60\u7684AI Infra\u5b66\u4e60\u642d\u5b50\u3002\u524d\u9762\u6211\u4eec\u804a\u4e86PagedAttention\u3001Continuous Batching\u3001Speculative Decoding\u548cPrefix Cache&#xff0c;\u8fd9\u4e9b\u4f18\u5316\u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u524d\u63d0\u2014\u2014Prefill\u548cDecode\u8dd1\u5728\u540c\u4e00\u4e2aGPU\u4e0a\u3002\u4f46\u5728\u751f\u4ea7\u73af\u5883\u4e2d&#xff0c;Prefill\u548cDecode\u5176\u5b9e\u662f\u4e24\u7c7b","og_url":"https:\/\/www.wsisp.com\/helps\/123854.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-11-08T01:32:56+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/123854.html","url":"https:\/\/www.wsisp.com\/helps\/123854.html","name":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-11-08T01:32:56+00:00","dateModified":"2026-11-08T01:32:56+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/123854.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/123854.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/123854.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010LLM\u63a8\u7406\u4e0eServing\u2465\u3011Prefill\/Decode\u89e3\u8026\u67b6\u6784\uff1a\u8ba9\u957fprompt\u4e0d\u518d\u5361\u4f4f\u77ed\u8bf7\u6c42"}]},{"@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\/123854","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=123854"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/123854\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=123854"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=123854"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=123854"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=123854"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}