{"id":35668,"date":"2024-12-29T10:00:11","date_gmt":"2024-12-29T02:00:11","guid":{"rendered":"https:\/\/17aitech.com\/?p=35668"},"modified":"2024-12-29T10:00:11","modified_gmt":"2024-12-29T02:00:11","slug":"%e6%98%93%e7%94%a8%e6%80%a7%e5%af%b9%e9%bd%90-vllm%ef%bc%8c%e6%8e%a8%e7%90%86%e6%95%88%e7%8e%87%e6%8f%90%e5%8d%87%e8%b6%85200%ef%bc%8c%e8%bf%99%e6%ac%be%e5%9b%bd%e4%ba%a7%e5%8a%a0%e9%80%9f%e6%a1%86","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35668","title":{"rendered":"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-18\" target=\"_blank\">\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f<\/a><\/p>\n<div id=\"ez-toc-container\" class=\"ez-toc-v2_0_85 ez-toc-wrap-left-text counter-hierarchy ez-toc-counter ez-toc-light-blue ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">\u6587\u7ae0\u76ee\u5f55<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999;color:#999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999;color:#999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 ' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/17aitech.com\/?p=35668\/#%E4%B8%80%E3%80%81%E8%A1%8C%E4%B8%9A%E8%83%8C%E6%99%AF\" >\u4e00\u3001\u884c\u4e1a\u80cc\u666f<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/17aitech.com\/?p=35668\/#%E4%BA%8C%E3%80%81%E4%B8%9A%E5%86%85%E6%8E%A8%E7%90%86%E5%BC%95%E6%93%8E%E6%96%B9%E6%A1%88\" >\u4e8c\u3001\u4e1a\u5185\u63a8\u7406\u5f15\u64ce\u65b9\u6848<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/17aitech.com\/?p=35668\/#%E4%B8%89%E3%80%81TACO-LLM_%E5%A6%82%E4%BD%95%E5%AE%9E%E7%8E%B0%E9%99%8D%E6%9C%AC%E5%A2%9E%E6%95%88\" >\u4e09\u3001TACO-LLM \u5982\u4f55\u5b9e\u73b0\u964d\u672c\u589e\u6548?<\/a><ul class='ez-toc-list-level-3' ><li class='ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/17aitech.com\/?p=35668\/#1%E3%80%81%E4%BD%BF%E7%94%A8%E5%B9%B6%E8%A1%8C%E8%A7%A3%E7%A0%81%E8%BF%9B%E8%A1%8C_Generation_%E4%BC%98%E5%8C%96\" >1\u3001\u4f7f\u7528\u5e76\u884c\u89e3\u7801\u8fdb\u884c Generation \u4f18\u5316<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/17aitech.com\/?p=35668\/#2%E3%80%81%E4%BD%BF%E7%94%A8_Prefix_Cache_%E6%8A%80%E6%9C%AF%E9%99%8D%E4%BD%8E_TTFT\" >2\u3001\u4f7f\u7528 Prefix Cache \u6280\u672f\u964d\u4f4e TTFT<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/17aitech.com\/?p=35668\/#3%E3%80%81%E9%95%BF%E5%BA%8F%E5%88%97%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96%E4%B8%AD%E7%9A%84_TurboAttention\" >3\u3001\u957f\u5e8f\u5217\u63a8\u7406\u4f18\u5316\u4e2d\u7684 TurboAttention<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-3'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/17aitech.com\/?p=35668\/#4%E3%80%81%E9%80%9A%E8%BF%87_LLM_%E9%87%8F%E5%8C%96%E6%8A%80%E6%9C%AF%E9%99%8D%E4%BD%8E%E6%8E%A8%E7%90%86%E6%88%90%E6%9C%AC\" >4\u3001\u901a\u8fc7 LLM \u91cf\u5316\u6280\u672f\u964d\u4f4e\u63a8\u7406\u6210\u672c<\/a><\/li><\/ul><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/17aitech.com\/?p=35668\/#%E5%9B%9B%E3%80%81TACO-LLM_%E5%AE%9E%E9%99%85%E6%95%88%E6%9E%9C%E4%B8%8E%E5%BA%94%E7%94%A8%E6%A1%88%E4%BE%8B\" >\u56db\u3001TACO-LLM \u5b9e\u9645\u6548\u679c\u4e0e\u5e94\u7528\u6848\u4f8b<\/a><\/li><\/ul><\/nav><\/div>\n<h2><span class=\"ez-toc-section\" id=\"%E4%B8%80%E3%80%81%E8%A1%8C%E4%B8%9A%E8%83%8C%E6%99%AF\"><\/span><strong>\u4e00\u3001\u884c\u4e1a\u80cc\u666f<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>2022 \u5e74 10 \u6708\uff0cChatGPT \u7684\u95ee\u4e16\u5f15\u7206\u4e86\u4ee5\u5927\u8bed\u8a00\u6a21\u578b\u4e3a\u4ee3\u8868\u7684\u7684 AI \u6d6a\u6f6e\uff0c\u5168\u7403\u79d1\u6280\u4f01\u4e1a\u7eb7\u7eb7\u52a0\u5165\u5927\u8bed\u8a00\u6a21\u578b\u7684\u519b\u5907\u7ade\u8d5b\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6570\u91cf\u3001\u53c2\u6570\u89c4\u6a21\u53ca\u8ba1\u7b97\u9700\u6c42\u5448\u6307\u6570\u7ea7\u63d0\u5347\u3002<\/p>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Model\uff0c\u7b80\u79f0 LLM \u5927\u6a21\u578b\uff09\u6307\u4f7f\u7528\u5927\u91cf\u6587\u672c\u6570\u636e\u8bad\u7ec3\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u53ef\u4ee5\u751f\u6210\u81ea\u7136\u8bed\u8a00\u6587\u672c\u6216\u7406\u89e3\u8bed\u8a00\u6587\u672c\u7684\u542b\u4e49\u3002\u5927\u6a21\u578b\u901a\u5e38\u5305\u542b\u767e\u4ebf\u81f3\u4e07\u4ebf\u4e2a\u53c2\u6570\uff0c\u8bad\u7ec3\u65f6\u9700\u8981\u5904\u7406\u6570\u4e07\u4ebf\u4e2a Token\uff0c\u8fd9\u5bf9\u663e\u5361\u7b49\u7b97\u529b\u63d0\u51fa\u4e86\u6781\u9ad8\u7684\u8981\u6c42\uff0c\u4e5f\u5e26\u6765\u4e86\u80fd\u6e90\u6d88\u8017\u7684\u6fc0\u589e\u3002<\/p>\n<p>\u636e\u65af\u5766\u798f\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u6240\u53d1\u5e03\u7684\u300a2023 \u5e74 AI \u6307\u6570\u62a5\u544a\u300b\uff0c\u5927\u8bed\u8a00\u6a21\u578b GPT-3 \u4e00\u6b21\u8bad\u7ec3\u7684\u8017\u7535\u91cf\u4e3a 1287 \u5146\u74e6\u65f6\uff0c\u76f8\u5f53\u4e8e\u6392\u653e\u4e86 552 \u5428\u4e8c\u6c27\u5316\u78b3\u3002\u968f\u7740 AI \u7684\u8fdb\u4e00\u6b65\u666e\u53ca\uff0c\u9884\u6d4b\u5230 2025 \u5e74\uff0cAI \u76f8\u5173\u4e1a\u52a1\u5728\u5168\u7403\u6570\u636e\u4e2d\u5fc3\u7528\u7535\u91cf\u4e2d\u7684\u5360\u6bd4\u5c06\u4ece 2% \u589e\u52a0\u81f3 10%\u3002\u5230 2030 \u5e74\uff0c\u667a\u80fd\u8ba1\u7b97\u5e74\u8017\u7535\u91cf\u5c06\u8fbe\u5230 5000 \u4ebf\u5343\u74e6\u65f6\uff0c\u5360\u5168\u7403\u53d1\u7535\u603b\u91cf\u7684 5%\u3002<\/p>\n<p>\u9664\u4e86\u7b97\u529b\u4e0e\u80fd\u8017\uff0c\u5f53\u5927\u6a21\u578b\u8fdb\u5165\u884c\u4e1a\u843d\u5730\u9636\u6bb5\uff0c\u5b9a\u5236\u5316\u548c\u8fd0\u8425\u6210\u672c\u4e5f\u6210\u4e3a\u65b0\u7684\u6838\u5fc3\u77db\u76fe\u3002\u4ee5\u6700\u65b0\u53d1\u5e03\u7684 Llama 3.1 405B \u4e3a\u4f8b\uff0c\u9700\u8981 450GB \u663e\u5b58\uff1b0.6B pixart \u5728 A800 \u4e0a\u751f\u6210 4096 px \u7684\u56fe\u7247\u9700\u8981 3 \u5206\u949f\uff0c\u5bf9\u4e1a\u52a1\u63d0\u51fa\u4e86\u6781\u9ad8\u7684\u8981\u6c42\u3002<\/p>\n<p>\u5982\u4f55\u5728\u66f4\u591a\u7684\u4e1a\u52a1\u4e0a\u5e94\u7528\u5927\u6a21\u578b\uff0c\u540c\u65f6\u964d\u4f4e\u6210\u672c\uff0c\u63d0\u9ad8\u6548\u7387\uff0c\u6210\u4e3a\u884c\u4e1a\u666e\u904d\u9700\u8981\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E4%BA%8C%E3%80%81%E4%B8%9A%E5%86%85%E6%8E%A8%E7%90%86%E5%BC%95%E6%93%8E%E6%96%B9%E6%A1%88\"><\/span><strong>\u4e8c\u3001\u4e1a\u5185\u63a8\u7406\u5f15\u64ce\u65b9\u6848<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u7684\u548c\u7528\u6237\u7684\u4ea4\u4e92\u8fc7\u7a0b\u4e2d\uff0c\u63a8\u7406\u6846\u67b6\u662f AI \u7684\u6838\u5fc3\u5f15\u64ce\uff0c\u8d1f\u8d23\u63a5\u6536\u7528\u6237\u7684\u8bf7\u6c42\uff0c\u5e76\u4e14\u5c06\u5176\u8fdb\u884c\u5904\u7406\u548c\u56de\u5e94\u3002\u6574\u4e2a AI \u884c\u4e1a\u90fd\u5728\u63a2\u7d22\u5982\u4f55\u9ad8\u6548\u5229\u7528\u8ba1\u7b97\u8d44\u6e90\uff0c\u5e76\u884c\u5904\u7406\u66f4\u591a\u7684\u63a8\u7406\u8bf7\u6c42\uff0c\u4ece\u800c\u9488\u5bf9\u73b0\u6709\u7684\u63a8\u7406\u6784\u67b6\u505a\u4f18\u5316\uff0c\u63a8\u51fa\u65b0\u7684\u5f02\u6784\u7b97\u529b\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<p>vLLM \u662f\u4f2f\u514b\u5229\u5927\u5b66\u7ec4\u7ec7\u5f00\u6e90\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u9ad8\u901f\u63a8\u7406\u6846\u67b6\uff0c\u4f7f\u7528 PagedAttention \u9ad8\u6548\u7ba1\u7406\u6ce8\u610f\u529b\u952e\u548c\u503c\u5185\u5b58\uff0c\u652f\u6301\u8fde\u7eed\u6279\u5904\u7406\u548c\u5feb\u901f\u6a21\u578b\u6267\u884c\uff0c\u901a\u8fc7\u5f15\u5165\u64cd\u4f5c\u7cfb\u7edf\u7684\u865a\u62df\u5185\u5b58\u5206\u9875\u601d\u60f3\uff0c\u63d0\u9ad8\u8bed\u8a00\u6a21\u578b\u670d\u52a1\u5728\u5b9e\u65f6\u573a\u666f\u4e0b\u7684\u541e\u5410\u4e0e\u5185\u5b58\u4f7f\u7528\u6548\u7387\u3002<\/p>\n<p>\u9664 vLLM \u5916\uff0c\u4f17\u591a\u5927\u6a21\u578b\u4e0a\u4e0b\u6e38\u5382\u5546\u4e5f\u7eb7\u7eb7\u7ed9\u51fa\u4e86\u81ea\u5df1\u7684\u65b9\u6848\uff1a<\/p>\n<p>Text Generation Inference\uff08TGI\uff09\u662f Hugging Face \u63a8\u51fa\u7684\u652f\u6301 Hugging Face Inference API \u548c Hugging Chat \u4e0a\u7684 LLM \u63a8\u7406\u7684\u5de5\u5177\uff0c\u65e8\u5728\u652f\u6301\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u4f18\u5316\u63a8\u7406\u3002<\/p>\n<p>TensorRT-LLM \u662f\u7531 NVIDIA \u63a8\u51fa\u7684\u5728 TensorRT \u63a8\u7406\u5f15\u64ce\u57fa\u7840\u4e0a\u9488\u5bf9 Transformer \u7c7b\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u7684\u5de5\u5177\uff0c\u652f\u6301\u591a\u79cd\u4f18\u5316\u6280\u672f\uff0c\u5982 kernel \u878d\u5408\u3001\u77e9\u9635\u4e58\u4f18\u5316\u3001\u91cf\u5316\u611f\u77e5\u8bad\u7ec3\u7b49\uff0c\u4ee5\u63d0\u5347\u63a8\u7406\u6027\u80fd\u3002<\/p>\n<p>DeepSpeed \u662f\u7531\u5fae\u8f6f\u5f00\u53d1\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u5de5\u5177\uff0c\u65e8\u5728\u652f\u6301\u66f4\u5927\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u5e76\u63d0\u4f9b\u4e86\u66f4\u591a\u7684\u4f18\u5316\u7b56\u7565\u548c\u5de5\u5177\uff0c\u5982 zero\u3001offload \u7b49\u3002\u652f\u6301\u591a\u79cd\u5e76\u884c\u7b56\u7565\uff0c\u5982\u6570\u636e\u5e76\u884c\u3001\u6a21\u578b\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\u4ee5\u53ca\u5b83\u4eec\u7684\u7ec4\u5408\uff083D \u5e76\u884c\uff09\uff0c\u53ef\u4ee5\u5728\u591a\u4e2a\u7ef4\u5ea6\u4e0a\u4f18\u5316\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u3002<\/p>\n<p>LightLLM \u662f\u4e00\u4e2a\u57fa\u4e8e Python \u7684 LLM \u63a8\u7406\u548c\u670d\u52a1\u6846\u67b6\uff0c\u4ee5\u8f7b\u91cf\u7ea7\u8bbe\u8ba1\u3001\u6613\u4e8e\u6269\u5c55\u548c\u9ad8\u901f\u6027\u80fd\u800c\u95fb\u540d\u3002LightLLM \u5229\u7528\u8bb8\u591a\u5907\u53d7\u597d\u8bc4\u7684\u5f00\u6e90\u5b9e\u73b0\u4f18\u52bf\uff0c\u5305\u62ec Faster Transformer\u3001TGI\u3001vLLM \u548c Flash Attention \u7b49\u3002<\/p>\n<p>\u8fd9\u4e9b\u6846\u67b6\u6709\u7740\u4e0d\u540c\u7684\u6280\u672f\u7279\u70b9\uff0c\u5177\u4f53\u6027\u80fd\u548c\u4f18\u52bf\u4e5f\u4f1a\u56e0\u4e0d\u540c\u7684\u5e94\u7528\u573a\u666f\u3001\u6a21\u578b\u914d\u7f6e\u548c\u786c\u4ef6\u73af\u5883\u800c\u6709\u6240\u5dee\u5f02\uff0c\u4f46\u4f9d\u7136\u6ca1\u6709\u89e3\u51b3\u6838\u5fc3\u7684\u6210\u672c\u95ee\u9898\u3002\u4e3a\u6b64\uff0c\u817e\u8baf\u63a8\u51fa\u4e86 TACO-LLM \u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u6846\u67b6\uff0c\u4e3a\u5b9a\u5236\u5316\u3001\u81ea\u5efa\u3001\u4e0a\u4e91\u3001\u79c1\u6709\u5316\u63d0\u4f9b\u5b8c\u6574\u90e8\u7f72\u65b9\u6848\u548c\u6781\u81f4\u6027\u4ef7\u6bd4\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E4%B8%89%E3%80%81TACO-LLM_%E5%A6%82%E4%BD%95%E5%AE%9E%E7%8E%B0%E9%99%8D%E6%9C%AC%E5%A2%9E%E6%95%88\"><\/span><strong>\u4e09\u3001TACO-LLM \u5982\u4f55\u5b9e\u73b0\u964d\u672c\u589e\u6548?<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>TACO-LLM \uff08TencentCloud Accelerated Computing Optimization LLM\uff09\u662f\u57fa\u4e8e\u817e\u8baf\u4e91\u5f02\u6784\u8ba1\u7b97\u4ea7\u54c1\u63a8\u51fa\u7684\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u52a0\u901f\u5f15\u64ce\uff0c\u901a\u8fc7\u5145\u5206\u5229\u7528\u8ba1\u7b97\u8d44\u6e90\u7684\u5e76\u884c\u8ba1\u7b97\u80fd\u529b\uff0c\u53ef\u540c\u65f6\u5904\u7406\u66f4\u591a\u7684\u7528\u6237\u8bf7\u6c42\uff0c\u63d0\u9ad8\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u6548\u80fd\uff0c\u4e3a\u5ba2\u6237\u63d0\u4f9b\u517c\u987e\u9ad8\u541e\u5410\u548c\u4f4e\u65f6\u5ef6\u7684\u4f18\u5316\u65b9\u6848\uff0c\u5e2e\u52a9\u5ba2\u6237\u5b9e\u73b0\u964d\u672c\u589e\u6548\u3002<\/p>\n<p>\u9488\u5bf9\u5404\u4e2a\u5e94\u7528\u573a\u666f\uff0cTaco-LLM \u7684\u4f18\u5316\u5927\u81f4\u5206\u4e3a\u56db\u7c7b\uff1a<strong>Generation \u4f18\u5316\u3001Prefill \u4f18\u5316\u3001\u957f\u5e8f\u5217\u4f18\u5316\u4e0e\u9ad8\u6027\u80fd\u91cf\u5316\u7b97\u5b50\u3002<\/strong>\u63a5\u4e0b\u6765\u5bf9\u8fd9\u56db\u4e2a\u573a\u666f\u8fdb\u884c\u8be6\u7ec6\u4ecb\u7ecd\uff1a<\/p>\n<h3><span class=\"ez-toc-section\" id=\"1%E3%80%81%E4%BD%BF%E7%94%A8%E5%B9%B6%E8%A1%8C%E8%A7%A3%E7%A0%81%E8%BF%9B%E8%A1%8C_Generation_%E4%BC%98%E5%8C%96\"><\/span><strong>1\u3001\u4f7f\u7528\u5e76\u884c\u89e3\u7801\u8fdb\u884c Generation \u4f18\u5316<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Generation \u4f18\u5316\u662f\u81ea\u56de\u5f52\u5f0f LLM \u6700\u91cd\u8981\u7684\u4f18\u5316\u4e4b\u4e00\uff0c\u51e0\u4e4e\u8986\u76d6\u6240\u6709 LLM \u5e94\u7528\u573a\u666f\u3002\u4f8b\u5982\uff1a\u6587\u6848\u521b\u4f5c\u3001\u667a\u80fd\u5ba2\u670d\u3001\u804a\u5929\u673a\u5668\u4eba\u3001\u4ee3\u7801\u751f\u6210\u3001\u54a8\u8be2\u7cfb\u7edf\u3001AI \u52a9\u624b\u7b49\uff1b \u8fd9\u91cc Taco-LLM \u7528\u5230\u5e76\u884c\u89e3\u7801\uff0c\u9ad8\u6027\u80fd\u7b97\u5b50\u7b49\u6280\u672f\u3002Taco-LLM \u4f7f\u7528\u5e76\u884c\u89e3\u7801\u7684\u4e3b\u8981\u4f18\u52bf\u6709\uff1a<\/p>\n<p><strong>1\u3001\u5e76<\/strong>\u884c\u89e3\u7801\u7a81\u7834\u81ea Transformer-Decoder \u67b6\u6784\u4e0b\u7684\u56de\u5f52\u9650\u5236\uff0c\u7f13\u89e3 Generation \u8fc7\u7a0bbandwidth bound \u95ee\u9898\u3002<\/p>\n<p><strong>2\u3001<\/strong>\u4e0e\u5355\u7eaf\u589e\u52a0 Generation \u8fc7\u7a0b\u7684 batch-size \u76f8\u6bd4\uff0c\u5e76\u884c\u89e3\u7801\u662f\u901a\u8fc7\u964d\u4f4e\u6bcf\u4e00\u6761\u8bf7\u6c42\u7684\u5ef6\u65f6\u6765\u589e\u52a0\u541e\u5410\uff0c\u53ef\u4ee5\u4fdd\u8bc1\u66f4\u4f4e\u7684 TPOT\u3002<\/p>\n<p><strong>3\u3001<\/strong>\u4e0e\u589e\u52a0 memory bandwidth \u7684\u5f02\u6784\u65b9\u6848\u76f8\u6bd4\uff0c\u5e76\u884c\u89e3\u7801\u4e0d\u9700\u8981\u5f02\u6784\u5316\u96c6\u7fa4\uff0c\u90e8\u7f72\u6210\u672c\u66f4\u4f4e\uff0c\u7cfb\u7edf\u66f4\u52a0\u7b80\u5355\u6613\u7ef4\u62a4\u3002<\/p>\n<p>Taco-LLM \u5728\u5e76\u884c\u89e3\u7801\u4e0a\u7684\u4e3b\u8981\u5c1d\u8bd5\u662f\u81ea\u9884\u6d4b\u65b9\u6848\uff0c\u4e00\u662f\u5b83\u89e3\u51b3\u4e86\u5c0f\u6a21\u578b\u4ece\u54ea\u91cc\u6765\u7684\u95ee\u9898\uff0c\u53ea\u9700\u8981\u4f7f\u7528\u5927\u6a21\u578b\u7684\u90e8\u5206\u5c42\u6216\u8005\u91cf\u5316\u540e\u7684\u5927\u6a21\u578b\u5f53\u4f5c\u5c0f\u6a21\u578b\u5373\u53ef\uff0c\u7528\u6237\u4e0d\u7528\u611f\u77e5 draft model \u7684\u5b58\u5728\uff0c\u4e8c\u662f\u8be5\u65b9\u6848\u76f8\u6bd4\u5176\u4ed6\u65b9\u6848\u6709\u8f83\u9ad8\u7684\u547d\u4e2d\u7387\uff0c\u8f83\u5c0f\u7684\u5197\u4f59\u8ba1\u7b97\uff0c\u4e3b\u8981\u7528\u4e8e 70B+ \u7684\u5927\u6a21\u578b\u63a8\u7406\u52a0\u901f\u3002\u9664\u81ea\u9884\u6d4b\u65b9\u6848\u4e4b\u5916\uff0cTaco-LLM \u4e5f\u652f\u6301 <strong>RawLookaheadCache \u548c TurboLookaheadCache\u00a0<\/strong>\u4e24\u79cd cache \u65b9\u6848\uff0c\u51cf\u5c11\u5197\u4f59\u8ba1\u7b97\uff0c\u63d0\u9ad8\u6027\u80fd\u4e0e\u6574\u4f53\u547d\u4e2d\u7387\u3002<\/p>\n<h3><span class=\"ez-toc-section\" id=\"2%E3%80%81%E4%BD%BF%E7%94%A8_Prefix_Cache_%E6%8A%80%E6%9C%AF%E9%99%8D%E4%BD%8E_TTFT\"><\/span><strong>2\u3001\u4f7f\u7528 Prefix Cache \u6280\u672f\u964d\u4f4e TTFT<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>Prefill \u4f18\u5316\u7684\u4e3b\u8981\u76ee\u6807\u662f\u964d\u4f4e TTFT\uff0c\u4f18\u5316\u7528\u6237\u4f7f\u7528\u4f53\u9a8c\uff0c\u8fd9\u91cc\u5e38\u7528\u7684\u4f18\u5316\u662f\u591a\u5361\u5e76\u884c\uff0c\u4f8b\u5982 TP \u548c SP\uff0c\u6765\u964d\u4f4e TTFT\uff0cTaco-LLM \u5728\u6b64\u57fa\u7840\u4e0a\u4f7f\u7528 GPU &amp; CPU \u7ed3\u5408\u591a\u7ea7\u7f13\u5b58\u7684 Prefix Cache \u6280\u672f\uff0c\u8ba9\u4e00\u90e8\u5206\u7684 prompt token \u901a\u8fc7\u67e5\u627e\u5386\u53f2\u7684 kv-cache \u83b7\u5f97\uff0c\u800c\u4e0d\u7528\u53c2\u4e0e Prefill \u9636\u6bb5\u7684\u8ba1\u7b97\uff0c\u51cf\u5c11\u8ba1\u7b97\u91cf\uff0c\u4ece\u800c\u964d\u4f4e TTFT\u3002\u8fd9\u9879\u6280\u672f\u5728\u4ee3\u7801\u52a9\u624b\u573a\u666f\u5c24\u5176\u6709\u6548\u3002<\/p>\n<p>\u4e3a\u4e86\u8282\u7701 Prefill \u7684\u8fd0\u884c\u65f6\u95f4\uff0c\u5c06\u5386\u53f2\u7684\u8bf7\u6c42\u7684 prefill cache \u6309\u7167\u524d\u7f00\u6811\u7684\u65b9\u5f0f\u4fdd\u5b58\u5728 GPU &amp; CPU \u4e2d\uff0c\u4ece\u800c\u5c06 Prefill \u7684\u8ba1\u7b97\u8fc7\u7a0b\u8f6c\u5316\u4e3a KV-Cache \u7684\u67e5\u8be2\u8fc7\u7a0b\uff0c\u5e76\u5c06\u6ca1\u6709\u547d\u4e2d\u7684 Tokens \u53c2\u5165 Prefill \u7684\u8ba1\u7b97\uff0c\u4ece\u800c\u964d\u4f4e\u8ba1\u7b97\u5f00\u9500\uff0c\u964d\u4f4e TTFT\u3002\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png\"><\/a><\/p>\n<h3><span class=\"ez-toc-section\" id=\"3%E3%80%81%E9%95%BF%E5%BA%8F%E5%88%97%E6%8E%A8%E7%90%86%E4%BC%98%E5%8C%96%E4%B8%AD%E7%9A%84_TurboAttention\"><\/span><strong>3\u3001\u957f\u5e8f\u5217\u63a8\u7406\u4f18\u5316\u4e2d\u7684 TurboAttention<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>\u957f\u5e8f\u5217\u4f18\u5316\u5206\u4e3a Prefill \u9636\u6bb5\u7684\u957f\u5e8f\u5217\uff0c\u4f8b\u5982\u6587\u672c\u6458\u8981\uff0c\u4fe1\u606f\u68c0\u7d22\u7b49\uff0c\u548c Generation \u9636\u6bb5\u7684\u957f\u5e8f\u5217\uff0c\u4f8b\u5982\u957f\u6587\u672c\u521b\u4f5c\u7b49\u3002\u524d\u8005\u7528\u5230 Prefix Cache \u548c\u591a\u5361\u5e76\u884c\u63a8\u7406\u6280\u672f\uff0c\u5bf9\u4e8e\u540e\u8005\u6211\u4eec\u81ea\u7814\u4e86 Turbo Attention \u7cfb\u5217\u7b97\u5b50\u548c\u4e00\u4e9b\u4f18\u5316\u540e\u7684\u91cf\u5316\u7b97\u5b50\u3002<\/p>\n<p>Taco-LLM \u957f\u5e8f\u5217\u80fd\u529b\u4e3b\u8981\u4f53\u73b0\u5728 TurboAttention\u3001Prefix Cache\u3001\u548c\u5e8f\u5217\u5e76\u884c\u4e0a\uff0c\u8fd9\u91cc\u4e3b\u8981\u4ecb\u7ecd TurboAttention\u3002TurboAttention \u7ed3\u5408\u4e86 Page \u7ba1\u7406\u673a\u5236\u548cFlash\u673a\u5236\uff0c\u4e13\u95e8\u4e3a\u957f\u5e8f\u5217\u4e0bLookahead\u5b9e\u73b0\u7684Kernel\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff1a<\/p>\n<h3><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-455cddbb4a9edcc3bc2e2f7ca9a250fc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-455cddbb4a9edcc3bc2e2f7ca9a250fc.png\"><\/a><\/h3>\n<h3><span class=\"ez-toc-section\" id=\"4%E3%80%81%E9%80%9A%E8%BF%87_LLM_%E9%87%8F%E5%8C%96%E6%8A%80%E6%9C%AF%E9%99%8D%E4%BD%8E%E6%8E%A8%E7%90%86%E6%88%90%E6%9C%AC\"><\/span><strong>4\u3001\u901a\u8fc7 LLM \u91cf\u5316\u6280\u672f\u964d\u4f4e\u63a8\u7406\u6210\u672c<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h3>\n<p>\u968f\u7740 LLM \u6a21\u578b\u53c2\u6570\u8fc5\u901f\u589e\u52a0\uff0cLLM \u7684\u63a8\u7406\u5ef6\u65f6\u548c\u63a8\u7406\u6210\u672c\u4e5f\u6025\u5267\u4e0a\u5347\u3002LLM \u91cf\u5316\u6280\u672f\u6210\u4e3a\u4f18\u5316 LLM \u63a8\u7406\u6027\u80fd\uff0c\u964d\u4f4e\u63a8\u7406\u6210\u672c\u7684\u4e00\u79cd\u91cd\u8981\u624b\u6bb5\u3002\u9ad8\u6027\u80fd\u91cf\u5316\u7b97\u5b50\u5bf9\u4e8e\u4e00\u4e9b\u5bf9\u7cbe\u5ea6\u8981\u6c42\u4e0d\u9ad8\u7684\u573a\u666f\uff0c\u50cf\u6587\u672c\u5206\u7c7b\uff0c\u6587\u672c\u5f02\u5e38\u68c0\u6d4b\uff0c\u6587\u672c\u6da6\u8272\u7b49\uff0c\u91cf\u5316\u5f80\u5f80\u6709\u8f83\u597d\u7684\u6548\u679c\uff0c\u53ef\u4ee5\u6709\u6548\u51cf\u5c11 GPU \u5185\u5b58\u5360\u7528\uff0c\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n<p>\u4e3a\u4e86\u8fbe\u5230\u91cf\u5316\u9884\u671f\u7684\u76ee\u6807\uff0c\u9ad8\u6548\u7684\u91cf\u5316\u7b97\u5b50\u5b9e\u73b0\u662f\u5fc5\u4e0d\u53ef\u5c11\u7684\uff0cTaco-LLM \u9488\u5bf9 GEMM \u548c Attention \u4e24\u7c7b\u7b97\u5b50\uff0c\u5f00\u53d1\u4e86<strong>\u6743\u91cd\u8ba1\u7b97\u611f\u77e5\u91cd\u6392\u3001\u4efb\u52a1\u8c03\u5ea6\u548c\u540c\u6b65\u7b56\u7565\u3001\u5feb\u901f\u53cd\u91cf\u5316\u3001Integer Scale \u6280\u672f<\/strong>\u7b49\u591a\u4e2a\u9ad8\u6548\u7684\u91cf\u5316\u7b97\u5b50\u6280\u672f\u3002<\/p>\n<p>\u603b\u7684\u6765\u8bf4\uff0cTaco-LLM \u7684\u4f18\u5316\u5305\u62ec\uff1a\u901a\u8fc7\u5e76\u884c\u89e3\u7801\u8fdb\u884c\u6295\u673a\u91c7\u6837\u548c LookaheadCache\uff1bPrefix Cache \u7684 GPU &amp; CPU \u7684\u591a\u7ea7\u7f13\u5b58\u6280\u672f\u4ee5\u53ca\u5185\u5b58\u7ba1\u7406\u6280\u672f\uff1b\u957f\u5e8f\u5217\u4e3b\u8981\u5305\u62ec\u81ea\u7814\u7684 TurboAttention \u7cfb\u5217\u7b97\u5b50\u3001Prefix Cache \u548c\u5e8f\u5217\u5e76\u884c\u7b49\uff1b\u9ad8\u6027\u80fd\u91cf\u5316\u7b97\u5b50\u5305\u62ec Taco-LLM \u5bf9 W4A8\uff0cW4A16\uff0cW8A16\uff0cW8A8 \u7b49\u91cf\u5316\u7b97\u5b50\u7684\u9ad8\u6548\u5b9e\u73b0\u3002<\/p>\n<h2><span class=\"ez-toc-section\" id=\"%E5%9B%9B%E3%80%81TACO-LLM_%E5%AE%9E%E9%99%85%E6%95%88%E6%9E%9C%E4%B8%8E%E5%BA%94%E7%94%A8%E6%A1%88%E4%BE%8B\"><\/span><strong>\u56db\u3001TACO-LLM \u5b9e\u9645\u6548\u679c\u4e0e\u5e94\u7528\u6848\u4f8b<\/strong><span class=\"ez-toc-section-end\"><\/span><\/h2>\n<p>\u901a\u8fc7\u5404\u9879\u4f18\u5316\u6280\u672f\uff0cTACO- LLM \u5728\u6027\u80fd\u4ee5\u53ca\u6210\u672c\uff0c\u76f8\u8f83\u4e8e\u73b0\u6709\u7684\u5f00\u6e90\u548c\u5382\u5546\u6846\u67b6\uff0c\u6709\u4e86\u660e\u663e\u7684\u4f18\u52bf\uff0c\u4e14\u5728\u6613\u7528\u6027\u4e0a\u5b8c\u5168\u5bf9\u9f50 vLLM\u3002<\/p>\n<p>TACO \u7684\u6027\u80fd\u4f18\u52bf\u76f4\u89c2\u4f53\u73b0\u5728\u5e2e\u52a9\u5927\u6a21\u578b\u670d\u52a1\u63d0\u9ad8\u541e\u5410\u548c\u964d\u4f4e\u5ef6\u8fdf\u3002\u5728 TACO \u52a0\u901f\u5f15\u64ce\u7684\u652f\u6301\u4e0b\uff0c\u76f8\u540c\u7684\u786c\u4ef6\u8bbe\u5907\u4e0a\u80fd\u591f\u5904\u7406\u7684 Token \u6570\u91cf\u663e\u8457\u589e\u52a0\uff0c\u539f\u672c\u6bcf\u79d2\u80fd\u5904\u7406 100 \u4e2a Token\uff0cTACO \u52a0\u6301\u540e\u6bcf\u79d2\u80fd\u5904\u7406<strong>\u00a0200\u00a0<\/strong>\u751a\u81f3<strong>\u00a0300\u00a0<\/strong>\u4e2aToken\u3002\u541e\u5410\u7684\u63d0\u5347\u6ca1\u6709\u4ee5\u727a\u7272\u5ef6\u8fdf\u4e3a\u4ee3\u4ef7\uff0c\u76f8\u53cd\u6bcf\u4e2a Token \u7684\u5e73\u5747\u5904\u7406\u65f6\u95f4\u5927\u5927\u964d\u4f4e\uff0c\u4ee3\u8868\u54cd\u5e94\u6548\u7387\u548c\u7528\u6237\u4f53\u9a8c\u7684\u63d0\u5347\uff0cLLM \u90e8\u7f72\u7684\u6210\u672c\u4e5f\u968f\u4e4b\u5927\u5e45\u964d\u4f4e\u3002<\/p>\n<p>\u4ee5 Llama &#8211; 3.1 70B \u6a21\u578b\u4e3a\u4f8b\uff0c\u5728\u4f7f\u7528 4 \u5f20 Ampere \u5b9e\u4f8b\uff0c\u8f93\u5165\u5e8f\u5217\u957f\u5ea6\u4e3a 1 K \u5de6\u53f3\uff0c\u8f93\u51fa \u4e3a 400 \u5de6\u53f3\uff0cBs = 1\uff0c2\uff0c4\uff0c8 \u7684\u6d4b\u8bd5\u573a\u666f\u4e2d\uff0c\u76f8\u8f83\u4e8e\u4e1a\u5185\u4e3b\u6d41\u7684 vLLM\uff0cTACO &#8211; LLM \u7684\u541e\u5410\u6027\u80fd\u76f8\u5bf9\u4e8e\u793e\u533a SOTA <strong>\u63d0\u5347 1.8\uff5e2.5 \u500d<\/strong>\uff1b\u8425\u8fd0\u6210\u672c<strong>\u964d\u4f4e 44\uff5e64%<\/strong>\uff0c\u4e14\u4f7f\u7528\u65b9\u5f0f\u3001\u8c03\u7528\u63a5\u53e3\u4fdd\u6301\u4e00\u81f4\uff0c\u652f\u6301\u65e0\u7f1d\u5207\u6362\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7586b59b28feae712838cb18c1994582.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7586b59b28feae712838cb18c1994582.png\"><\/a><\/p>\n<p><\/p>\n<p>\u540c\u6837\u4ee5 Llama-3.1 70B \u4e3a\u4f8b\uff0c\u4f7f\u7528 TACO-LLM \u90e8\u7f72\u7684<strong>\u6210\u672c\u4f4e\u81f3 &lt;$0.5\/1M tokens<\/strong>\uff0c\u76f8\u6bd4\u76f4\u63a5\u8c03\u7528 MaaS API \u7684<strong>\u6210\u672c\u8282\u7ea6\u8d85\u8fc7 60%+<\/strong>\uff0c\u4e14\u4f7f\u7528\u65b9\u5f0f\u3001\u8c03\u7528\u63a5\u53e3\u4fdd\u6301\u4e00\u81f4\uff0c\u652f\u6301\u65e0\u7f1d\u5207\u6362\u3002TACO &#8211; LLM \u5353\u8d8a\u7684\u80fd\u6548\u6bd4\u663e\u8457\u964d\u4f4e\u4e86 LLM \u4e1a\u52a1\u6210\u672c\uff0c\u5728\u5e7f\u6cdb\u7684\u5b9e\u9645\u573a\u666f\u4e2d\u5b9e\u73b0\u4e86\u964d\u672c\u589e\u6548\uff1a<\/p>\n<p>\u5728\u5fae\u4fe1\u67d0\u6587\u672c\u5904\u7406\u4e1a\u52a1\u4e2d\uff0cTACO &#8211; LLM \u541e\u5410\u6027\u80fd\u76f8\u5bf9\u4e8e\u7ade\u54c1<strong>\u63d0\u5347 2.8 \u500d<\/strong>\uff0c\u8425\u8fd0<strong>\u6210\u672c\u964d\u4f4e 64%<\/strong>\uff0c\u8d85\u65f6\u5931\u8d25<strong>\u964d\u4f4e\u7ea6 95%<\/strong>\uff0c\u8fdb\u4e00\u6b65\u6269\u5927\u652f\u6301\u6587\u672c\u7684\u6700\u5927\u957f\u5ea6\u3002<\/p>\n<p>\u5728\u67d0\u5934\u90e8\u89c6\u9891\u5e73\u53f0\u4e1a\u52a1\u4e2d\uff0c\u5ba2\u6237\u5e0c\u671b\u5728\u81ea\u5efa\u9ad8\u7aef\u5b9e\u4f8b\u4e0a\u90e8\u7f72\u63a8\u7406\u670d\u52a1\uff0c\u8981\u6c42\u76f8\u5bf9\u5382\u5546\u5b98\u65b9\u63a8\u7406\u6846\u67b6<strong>\u6027\u80fd\u63d0\u5347 50% \u4ee5\u4e0a<\/strong>\u3002\u6700\u7ec8\uff0cTACO-LLM \u5728\u4e0d\u540c bs \u4e0b\u76f8\u5bf9\u7ade\u54c1<strong>\u6027\u80fd\u63d0\u5347 1.7~2.5 \u500d<\/strong>\u3002<\/p>\n<p>\u5728\u987a\u4e30\u67d0\u4e1a\u52a1\u4e2d\uff0cTACO &#8211; LLM \u5728\u77ed\u8f93\u51fa\u573a\u666f\uff0c\u5728\u4e0d\u540c bs \u4e0b\uff0c<strong>\u52a0\u901f\u5e45\u5ea6\u4e3a 2~3 \u500d<\/strong>\uff1b\u5728\u957f\u8f93\u51fa\u573a\u666f\uff0c\u5728\u4e0d\u540c bs \u4e0b\uff0c<strong>\u52a0\u901f\u5e45\u5ea6\u4e3a 1.4 \u5230 1.99 \u500d<\/strong>\u3002<\/p>\n<p>TACO &#8211; LLM \u7684\u51fa\u73b0\u6253\u7834\u4e86\u4ee5\u5f80\u9ad8\u6602\u6210\u672c\u5bf9\u4eba\u4eec\u4f7f\u7528 AI \u7684\u9650\u5236\uff0c\u4e0d\u4ec5\u80fd\u591f\u6ee1\u8db3\u7528\u6237\u5bf9\u9ad8\u541e\u5410\u548c\u4f4e\u65f6\u5ef6\u7684\u9700\u6c42\uff0c\u8fd8\u80fd\u5e2e\u52a9\u4f01\u4e1a\u964d\u672c\u589e\u6548\uff0c\u4e3a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5e94\u7528\u63d0\u4f9b\u4e86\u4e00\u79cd\u66f4\u9ad8\u6548\u3001\u66f4\u7ecf\u6d4e\u7684\u89e3\u51b3\u65b9\u6848\uff0c<\/p>\n<p>\u5728\u672a\u6765\uff0c\u968f\u7740\u6280\u672f\u7684\u4e0d\u65ad\u8fed\u4ee3\uff0cTACO &#8211; LLM \u6709\u671b\u5728\u66f4\u591a\u9886\u57df\u5f97\u5230\u5e7f\u6cdb\u5e94\u7528\uff0c\u63a8\u52a8\u884c\u4e1a\u7684\u53d1\u5c55\u548c\u521b\u65b0\uff0c\u8ba9 AI \u771f\u6b63\u8d70\u8fdb\u4e86\u4eba\u4eec\u7684\u751f\u6d3b\uff0c\u6210\u4e3a\u4eba\u4eec\u5b66\u4e60\u548c\u751f\u6d3b\u7684\u5f97\u529b\u52a9\u624b\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-18\" target=\"_blank\">\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68,65],"class_list":["post-35668","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-65"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f - \u4e00\u8d77AI\u6280\u672f<\/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:\/\/17aitech.com\/?p=35668\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f\",\"datePublished\":\"2024-12-29T02:00:11+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668\"},\"wordCount\":223,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/09\\\/frc-fbde48f64b1f537994521880c2398662.png\",\"keywords\":[\"\u5927\u6a21\u578b\",\"\u6df1\u5ea6\u5b66\u4e60\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35668#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=35668\",\"name\":\"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/09\\\/frc-fbde48f64b1f537994521880c2398662.png\",\"datePublished\":\"2024-12-29T02:00:11+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35668\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/09\\\/frc-fbde48f64b1f537994521880c2398662.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/09\\\/frc-fbde48f64b1f537994521880c2398662.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35668#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f - \u4e00\u8d77AI\u6280\u672f","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:\/\/17aitech.com\/?p=35668","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=35668#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=35668"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f","datePublished":"2024-12-29T02:00:11+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=35668"},"wordCount":223,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=35668#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png","keywords":["\u5927\u6a21\u578b","\u6df1\u5ea6\u5b66\u4e60"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=35668#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35668","url":"https:\/\/17aitech.com\/?p=35668","name":"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35668#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35668#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png","datePublished":"2024-12-29T02:00:11+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35668#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35668"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35668#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-fbde48f64b1f537994521880c2398662.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35668#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u6613\u7528\u6027\u5bf9\u9f50 vLLM\uff0c\u63a8\u7406\u6548\u7387\u63d0\u5347\u8d85200%\uff0c\u8fd9\u6b3e\u56fd\u4ea7\u52a0\u901f\u6846\u67b6\u4ec0\u4e48\u6765\u5934\uff1f"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35668","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=35668"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35668\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35668"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35668"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35668"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}