{"id":24943,"date":"2025-03-07T11:14:02","date_gmt":"2025-03-07T03:14:02","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=24943"},"modified":"2025-03-07T11:14:02","modified_gmt":"2025-03-07T03:14:02","slug":"%e5%bd%93%e5%bc%80%e6%ba%90%e5%88%9b%e6%96%b0%e9%81%87%e4%b8%8a%e6%8e%a8%e7%90%86%e9%9d%a9%e5%91%bd%ef%bc%9asglang%e5%a6%82%e4%bd%95%e7%82%bc%e5%b0%b1deepseek%e6%9c%80%e5%bc%ba%e5%bc%80%e6%ba%90","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/03\/07\/24943\/","title":{"rendered":"\u5f53\u5f00\u6e90\u521b\u65b0\u9047\u4e0a\u63a8\u7406\u9769\u547d\uff1aSGLang\u5982\u4f55\u70bc\u5c31DeepSeek\u6700\u5f3a\u5f00\u6e90\u63a8\u7406\u5f15\u64ce\uff1f"},"content":{"rendered":"<div><img data-dominant-color=\"c6bfbe\" data-has-transparency=\"false\" style=\"--dominant-color: #c6bfbe;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-24945\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/03\/f2ffb5308eab2758d2e4427377812930bd4d3f-300x167-1.jpg\" width=\"300\" height=\"167\" alt=\"\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/03\/f2ffb5308eab2758d2e4427377812930bd4d3f-300x167-1.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/03\/f2ffb5308eab2758d2e4427377812930bd4d3f-300x167-1-150x84.jpg 150w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/div>\n<div><\/div>\n<div class=\"article-desc\">\u4ece DeepSeek \u6a21\u578b\u53d1\u5e03\u5f53\u5929\u4fbf\u5b9e\u73b0\u6700\u4f73\u9002\u914d\uff0c\u5230\u957f\u671f\u7a33\u5c45 SOTA \u6027\u80fd\u699c\u9996\uff0cSGLang \u7684\u8fdb\u5316\u8f68\u8ff9\u63ed\u793a\u4e86\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\u7684\u786c\u6838\u751f\u5b58\u6cd5\u5219\uff1a\u7528\u5de5\u7a0b\u521b\u65b0\uff0c\u653b\u514b\u5f00\u53d1\u8005\u6700\u68d8\u624b\u7684\u6027\u80fd\u74f6\u9888\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<p>2025 \u5e74\u5f00\u5e74\uff0cDeepSeek R1 \u548c V3 \u91cd\u78c5\u53d1\u5e03\uff0c\u5176\u8d85\u5f3a\u7684\u8bed\u8a00\u5efa\u6a21\u4e0e\u63a8\u7406\u80fd\u529b\uff0c\u5f15\u7206\u4e86\u5168\u7403 AI \u793e\u533a\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u4e00\u4e2a\u9690\u85cf\u5728\u8d85\u5927\u89c4\u6a21\u6a21\u578b\u8eab\u540e\u7684\u6280\u672f\u547d\u9898\u6d6e\u51fa\u6c34\u9762\uff1a<strong>\u5982\u4f55\u8ba9\u5343\u4ebf\u53c2\u6570\u8d85\u5927\u89c4\u6a21 AI \u6a21\u578b\u771f\u6b63\u8fbe\u5230\u5546\u4e1a\u7ea7\u63a8\u7406\u901f\u5ea6\uff1f<\/strong>\u8fd9\u4e00\u95ee\u9898\u7684\u7b54\u6848\uff0c<strong>\u9690\u85cf\u5728\u63a8\u7406\u5f15\u64ce SGLang \u7684\u4ee3\u7801\u4ed3\u5e93\u4e2d<\/strong>\u3002\u8be5\u9879\u76ee\u7531 LMSYS Org \u53d1\u8d77\uff0c\u5e76\u53d7\u5230 xAI\u3001NVIDIA\u3001AMD \u7b49\u5de8\u5934\u7684\u9752\u7750\uff0c\u6b63\u5728\u901a\u8fc7\u591a\u9879\u5173\u952e\u6280\u672f\u7a81\u7834\uff0c\u91cd\u65b0\u5b9a\u4e49 LLM \u63a8\u7406\u7684\u6548\u7387\u8fb9\u754c\u3002<\/p>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202503\/07\/b7943019125253422625753e891c712a6a95ea.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u4ece DeepSeek \u6a21\u578b\u53d1\u5e03\u5f53\u5929\u4fbf\u5b9e\u73b0\u6700\u4f73\u9002\u914d\uff0c\u5230\u957f\u671f\u7a33\u5c45 SOTA \u6027\u80fd\u699c\u9996\uff0cSGLang \u7684\u8fdb\u5316\u8f68\u8ff9\u63ed\u793a\u4e86\u4e00\u4e2a\u5f00\u6e90\u9879\u76ee\u7684\u786c\u6838\u751f\u5b58\u6cd5\u5219\uff1a<strong>\u7528\u5de5\u7a0b\u521b\u65b0\uff0c\u653b\u514b\u5f00\u53d1\u8005\u6700\u68d8\u624b\u7684\u6027\u80fd\u74f6\u9888\u3002<\/strong><\/p>\n<p>\u901a\u8fc7\u9886\u5148\u7684 Multi-head Latent Attention Optimzation\u3001Data Parallelism Router\u3001Eagle Speculative Decoding \u7b49\u7b49\u6280\u672f\u65b9\u6848\uff0cSGLang \u957f\u671f\u4fdd\u6301\u5f00\u6e90\u6a21\u578b\u9876\u5c16\u7684\u63a8\u7406\u901f\u5ea6\u548c\u541e\u5410\u91cf\u3002<\/p>\n<p>\u4f46\u662f\uff0cSGLang \u7684\u5f81\u7a0b\u7edd\u4e0d\u6b62\u6b65\u4e8e\u6b64\u3002\u5f53 Agent \u7684\u5de5\u7a0b\u5e08\u4eec\u7528\u5176\u90e8\u7f72\u667a\u80fd\u4f53\u65f6\uff0c\u5f53\u5f00\u53d1\u8005\u5728 NVIDIA Triton \u5185\u6838\u4e2d\u878d\u5165\u5176\u4f18\u5316\u7b56\u7565\u65f6\uff0c\u5f53\u5168\u4e16\u754c\u7684\u7814\u7a76\u8005\u9ad8\u5f3a\u5ea6\u4f7f\u7528 DeepSeek \u672c\u5730\u90e8\u7f72\u65f6\uff0c\u8fd9\u4e2a\u9879\u76ee\u7684\u771f\u6b63\u4ef7\u503c\u6b63\u5728\u663e\u73b0\uff1a<strong>\u5b83\u4e0d\u4ec5\u662f\u957f\u671f\u9886\u5148\u7684\u63a8\u7406\u5f15\u64ce\uff0c\u66f4\u662f\u5f00\u6e90\u793e\u533a\u96c6\u4f53\u667a\u6167\u7684\u7ed3\u6676<\/strong>\u3002\u672c\u6587\u5c06\u4ece\u6838\u5fc3\u6280\u672f\u7a81\u7834\u3001\u7cfb\u7edf\u7ea7\u4f18\u5316\u5230\u5f00\u53d1\u8005\u751f\u6001\uff0c\u89e3\u7801 SGLang \u72ec\u5230\u7684\u8fdb\u5316\u4e4b\u8def\u3002<\/p>\n<h3>\u4e00\u3001DeepSeek \u6a21\u578b\u6301\u7eed\u4f18\u5316\uff0c\u67b6\u6784\u9002\u914d\u7684\u5de5\u7a0b\u5b9e\u8df5<\/h3>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s2.51cto.com\/oss\/202503\/07\/f2446270657baa8a90a0345c8d2f98f8b0efce.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>image credit: DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model<\/p>\n<p>\u81ea\u4ece DeepSeek V2 \u53d1\u5e03\u4ee5\u6765\uff0cSGLang \u56e2\u961f\u9488\u5bf9 DeepSeek \u7cfb\u5217\u6a21\u578b\u7684 MLA\uff08Multi-head Latent Attention\uff09\u67b6\u6784\u8fdb\u884c\u4e86\u6df1\u5ea6\u4f18\u5316\u3002\u8fd9\u4e9b\u6280\u672f\u8986\u76d6\u4e86\u6570\u636e\u5e76\u884c\u6ce8\u610f\u529b\uff08Data Parallelism Attention\uff09\u3001\u591a\u8282\u70b9\u5f20\u91cf\u5e76\u884c\uff08Multi Node Tensor Parallelism\uff09\u4ee5\u53ca\u5757\u7ea7 FP8 \u91cf\u5316\uff08Block-wise FP8\uff09\uff0c\u4ece\u800c\u5728\u89e3\u7801\u8ba1\u7b97\u3001\u663e\u5b58\u7ba1\u7406\u548c\u591a\u8282\u70b9\u534f\u540c\u7b49\u591a\u4e2a\u73af\u8282\u5b9e\u73b0\u4e86\u7a81\u7834\u6027\u63d0\u5347\u3002<\/p>\n<p>\u5bf9\u4e8e Multi-head Latent Attention\uff08MLA\uff09\u7684\u4f18\u5316\uff0c\u56e2\u961f\u901a\u8fc7\u4f7f\u7528\u6743\u91cd\u5438\u6536\u91cd\u65b0\u6392\u5217\u8ba1\u7b97\u6b65\u9aa4\uff0c\u5728\u4fdd\u8bc1\u6a21\u578b\u8868\u8fbe\u80fd\u529b\u7684\u524d\u63d0\u4e0b\uff0c\u5e73\u8861\u4e86\u8ba1\u7b97\u4e0e\u5185\u5b58\u8bbf\u95ee\u8d1f\u8f7d\uff0c\u964d\u4f4e\u4e86\u89e3\u7801\u8fc7\u7a0b\u4e2d\u7684\u5197\u4f59\u8ba1\u7b97\uff0c\u964d\u4f4e\u4e86 MLA \u5728 Decode \u8fc7\u7a0b\u4e2d\u7684\u8ba1\u7b97\u91cf\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u9488\u5bf9 MLA \u89e3\u7801\u6838\u4ec5\u4fdd\u7559\u4e00\u4e2a KV \u5934\u7684\u8bbe\u8ba1\uff0cSGLang \u56e2\u961f\u5f00\u53d1\u4e86 Triton \u89e3\u7801\u6838\u4f18\u5316\u65b9\u6848\u3002\u8be5\u65b9\u6848\u901a\u8fc7\u5728\u540c\u4e00\u8ba1\u7b97\u5757\u5185\u540c\u65f6\u5904\u7406\u591a\u4e2a query \u5934\uff0c\u663e\u8457\u51cf\u5c11\u4e86\u5bf9 KV Cache \u7684\u5185\u5b58\u8bbf\u95ee\u9700\u6c42\uff0c\u4ece\u800c\u52a0\u901f\u4e86\u89e3\u7801\u6d41\u7a0b\u3002\u6b64\u5916\uff0c\u56e2\u961f\u7ed3\u5408 W8A8 FP8\u3001KV Cache FP8 \u91cf\u5316\u6280\u672f\uff0c\u5e76\u5f00\u53d1\u4e86 FP8 \u6279\u91cf\u77e9\u9635\u4e58\u6cd5\uff08BMM\uff09\u7b97\u5b50\uff0c\u5b9e\u73b0\u4e86 MLA \u9ad8\u6548\u7684 FP8 \u63a8\u7406\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cMLA \u4e0e Mixture-of-Experts\uff08MoE\uff09\u6a21\u5757\u5747\u5df2\u517c\u5bb9 CUDA Graph \u548c Torch.compile\uff0c\u80fd\u591f\u8fdb\u4e00\u6b65\u964d\u4f4e\u5c0f\u6279\u91cf\u63a8\u7406\u65f6\u7684\u5ef6\u8fdf\u3002\u7ecf\u8fc7\u8fd9\u4e9b\u7efc\u5408\u4f18\u5316\uff0cDeepSeek \u7cfb\u5217\u6a21\u578b\u5728\u8f93\u51fa\u541e\u5410\u7387\u65b9\u9762\u8f83\u4e0a\u4e00\u7248\u672c\u5b9e\u73b0\u4e86\u6700\u9ad8\u8fbe 7 \u500d\u7684\u52a0\u901f\u6548\u679c\u3002<\/p>\n<p>\u9762\u5bf9\u9ad8\u5e76\u53d1\u548c\u5927\u6279\u91cf\u6570\u636e\u7684\u5b9e\u9645\u5e94\u7528\u9700\u6c42\uff0c\u56e2\u961f\u8fdb\u4e00\u6b65\u5728 MLA \u6ce8\u610f\u529b\u673a\u5236\u4e2d\u5f15\u5165\u4e86\u6570\u636e\u5e76\u884c\u6ce8\u610f\u529b\u6280\u672f\u3002\u8be5\u65b9\u6848\u901a\u8fc7\u5c06\u4e0d\u540c\u7c7b\u578b\u7684 batch\uff08\u5305\u62ec prefill\u3001decode\u3001extend \u4ee5\u53ca idle \u72b6\u6001\uff09\u5206\u522b\u5206\u914d\u7ed9\u5404\u4e2a\u6570\u636e\u5e76\u884c\u5de5\u4f5c\u5355\u5143\uff0c\u4f7f\u5f97\u5404\u5355\u5143\u80fd\u591f\u72ec\u7acb\u5904\u7406\u5404\u81ea\u4efb\u52a1\u3002\u5f85\u4efb\u52a1\u5b8c\u6210\u540e\uff0c\u5728 Mixture-of-Experts\uff08MoE\uff09\u5c42\u524d\u540e\u518d\u8fdb\u884c\u5fc5\u8981\u7684\u540c\u6b65\u64cd\u4f5c\uff0c\u4ece\u800c\u663e\u8457\u964d\u4f4e\u4e86 KV Cache \u7684\u91cd\u590d\u5b58\u50a8\u8d1f\u62c5\uff0c\u4f18\u5316\u4e86\u5185\u5b58\u4f7f\u7528\uff0c\u5e76\u652f\u6301\u66f4\u5927\u6279\u91cf\u8bf7\u6c42\u7684\u9ad8\u6548\u5904\u7406\u3002\u8be5\u4f18\u5316\u4e13\u4e3a\u9ad8 QPS\uff08Queries Per Second\uff09\u573a\u666f\u8bbe\u8ba1\uff0c\u7528\u6237\u5728\u4f7f\u7528 DeepSeek \u7cfb\u5217\u6a21\u578b\u65f6\u53ef\u901a\u8fc7\u547d\u4ee4\u53c2\u6570 &#8211;enable-dp-attention \u4e00\u952e\u542f\u7528\u8fd9\u4e00\u529f\u80fd\u3002<\/p>\n<p>\u5728\u5355\u8282\u70b9\u5185\u5b58\u53d7\u9650\u7684\u60c5\u51b5\u4e0b\uff0cSGLang \u56e2\u961f\u8fd8\u63a8\u51fa\u4e86\u591a\u8282\u70b9\u5f20\u91cf\u5e76\u884c\u6280\u672f\u3002\u8be5\u65b9\u6848\u5141\u8bb8\u5c06\u8d85\u5927\u89c4\u6a21\u6a21\u578b\uff08\u5982 DeepSeek V3\uff09\u8de8\u591a\u4e2a GPU \u6216\u8282\u70b9\u8fdb\u884c\u53c2\u6570\u5206\u533a\u90e8\u7f72\uff0c\u6709\u6548\u7a81\u7834\u5355\u8282\u70b9\u5185\u5b58\u74f6\u9888\u3002\u7528\u6237\u53ef\u4ee5\u6839\u636e\u5b9e\u9645\u8d44\u6e90\u60c5\u51b5\uff0c\u5728\u96c6\u7fa4\u73af\u5883\u4e2d\u7075\u6d3b\u914d\u7f6e\u591a\u8282\u70b9\u5f20\u91cf\u5e76\u884c\uff0c\u786e\u4fdd\u6a21\u578b\u5728\u9ad8\u8d1f\u8f7d\u573a\u666f\u4e0b\u4f9d\u7136\u80fd\u4fdd\u6301\u9ad8\u6548\u63a8\u7406\u548c\u8d44\u6e90\u5229\u7528\u7387\u3002<\/p>\n<p>\u4e3a\u4e86\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u8fdb\u4e00\u6b65\u5e73\u8861\u6570\u503c\u7cbe\u5ea6\u4e0e\u8ba1\u7b97\u6548\u7387\uff0c\u56e2\u961f\u8fd8\u5f00\u53d1\u4e86\u5757\u7ea7 FP8 \u91cf\u5316\u65b9\u6848\u3002\u5728\u6fc0\u6d3b\u503c\u91cf\u5316\u65b9\u9762\uff0c\u91c7\u7528 E4M3 \u683c\u5f0f\uff0c\u5e76\u901a\u8fc7\u5bf9\u6bcf\u4e2a token \u5185 128 \u901a\u9053\u5b50\u5411\u91cf\u8fdb\u884c\u5728\u7ebf casting\uff0c\u5b9e\u73b0\u52a8\u6001\u7f29\u653e\uff0c\u4ece\u800c\u786e\u4fdd\u91cf\u5316\u540e\u6fc0\u6d3b\u503c\u7684\u6570\u503c\u7a33\u5b9a\u6027\uff1b\u800c\u5728\u6743\u91cd\u91cf\u5316\u4e0a\uff0c\u5219\u4ee5 128\u00d7128 \u5757\u4e3a\u57fa\u672c\u5355\u5143\u8fdb\u884c\u5904\u7406\uff0c\u4f7f\u5f97\u91cf\u5316\u8fc7\u7a0b\u66f4\u4e3a\u7cbe\u7ec6\uff0c\u6709\u6548\u6355\u6349\u6743\u91cd\u5206\u5e03\u7279\u6027\u3002\u8fd9\u4e00\u65b9\u6848\u5df2\u5728 DeepSeek V3 \u6a21\u578b\u4e2d\u9ed8\u8ba4\u542f\u7528\uff0c\u4e3a\u6a21\u578b\u5728\u9ad8\u6548\u63a8\u7406\u7684\u540c\u65f6\u4fdd\u6301\u8f83\u9ad8\u7cbe\u5ea6\u63d0\u4f9b\u4e86\u6709\u529b\u4fdd\u969c\u3002<\/p>\n<p>\u5728\u5982\u6b64\u6781\u81f4\u7684\u4f18\u5316\u4e4b\u4e0b\uff0cSGLang \u56e2\u961f\u5b9e\u73b0\u4e86\u4ece\u89e3\u7801\u8ba1\u7b97\u5230\u5185\u5b58\u7ba1\u7406\u3001\u4ece\u5355\u8282\u70b9\u4f18\u5316\u5230\u8de8\u8282\u70b9\u534f\u540c\u7684\u5168\u65b9\u4f4d\u63d0\u5347\u3002\u8fd9\u4e9b\u6280\u672f\u521b\u65b0\u4f7f\u5f97 SGLang \u5728 DeepSeek \u6a21\u578b\u5728\u4fdd\u6301\u9ad8\u7cbe\u5ea6\u7684\u57fa\u7840\u4e0a\uff0c\u5176\u8f93\u51fa\u541e\u5410\u7387\u6700\u9ad8\u53ef\u8fbe 7 \u500d\u63d0\u5347\uff0c\u5e76\u5728\u9ad8\u5e76\u53d1\u548c\u5927\u89c4\u6a21\u90e8\u7f72\u573a\u666f\u4e2d\u5c55\u73b0\u51fa\u5353\u8d8a\u7684\u6027\u80fd\u548c\u7075\u6d3b\u6027\u3002\u66f4\u591a\u8be6\u7ec6\u6280\u672f\u4fe1\u606f\u53ca\u4f7f\u7528\u6848\u4f8b\uff0c\u8bf7\u53c2\u9605\u5b98\u65b9 Blog \u4e0e\u76f8\u5173\u6280\u672f\u6f14\u793a\u6587\u7a3f\u3002<\/p>\n<h3>\u4e8c\u3001Zero-Overhead Batch Scheduler\uff1a\u8c03\u5ea6\u5668\u7684\u6548\u80fd\u9769\u547d<\/h3>\n<p>\u5728\u4f20\u7edf\u63a8\u7406\u5f15\u64ce\u4e2d\uff0c\u5c3d\u7ba1\u5927\u6a21\u578b\u7684\u63a8\u7406\u4e3b\u8981\u4f9d\u8d56 GPU \u8fd0\u7b97\uff0c\u4f46 CPU \u4ecd\u9700\u627f\u62c5\u6279\u8c03\u5ea6\u3001\u5185\u5b58\u5206\u914d\u3001\u524d\u7f00\u5339\u914d\u7b49\u5927\u91cf\u5de5\u4f5c\u3002\u672a\u7ecf\u5145\u5206\u4f18\u5316\u7684\u63a8\u7406\u7cfb\u7edf\u5f80\u5f80\u4f1a\u5c06\u591a\u8fbe\u4e00\u534a\u7684\u65f6\u95f4\u8017\u8d39\u5728\u8fd9\u4e9b CPU \u5f00\u9500\u4e0a\uff0c\u4e25\u91cd\u5f71\u54cd\u6574\u4f53\u6027\u80fd\u3002SGLang \u4e00\u76f4\u4ee5\u9ad8\u6548\u7684\u6279\u8c03\u5ea6\u5668\u8457\u79f0\uff0c\u800c\u5728 0.4 \u7248\u672c\u4e2d\uff0c\u56e2\u961f\u8fdb\u4e00\u6b65\u7a81\u7834\uff0c\u5b9e\u73b0\u4e86\u8fd1\u4e4e\u96f6\u5f00\u9500\u7684\u6279\u8c03\u5ea6\u5668\u3002<\/p>\n<p>\u8fd9\u4e00\u6280\u672f\u7684\u6838\u5fc3\u5728\u4e8e\u5c06 CPU \u8c03\u5ea6\u4e0e GPU \u8ba1\u7b97\u91cd\u53e0\u6267\u884c\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8c03\u5ea6\u5668\u63d0\u524d\u4e00\u6279\u8fd0\u884c\uff0c\u5728 GPU \u6267\u884c\u5f53\u524d\u4efb\u52a1\u7684\u540c\u65f6\uff0c\u4fbf\u540c\u6b65\u51c6\u5907\u597d\u4e0b\u4e00\u6279\u6240\u9700\u7684\u6240\u6709\u5143\u6570\u636e\u3002\u8fd9\u6837\u4e00\u6765\uff0cGPU \u59cb\u7ec8\u5904\u4e8e\u5fd9\u788c\u72b6\u6001\uff0c\u65e0\u9700\u7b49\u5f85 CPU \u7684\u8c03\u5ea6\u7ed3\u679c\uff0c\u6210\u529f\u9690\u85cf\u4e86\u8bf8\u5982\u5339\u914d radix cache \u7b49\u8f83\u4e3a\u6602\u8d35\u7684\u64cd\u4f5c\u7684\u5f00\u9500\u3002\u901a\u8fc7 Nsight profiling \u5de5\u5177\u7684\u6d4b\u8bd5\u663e\u793a\uff0c\u5728\u8fde\u7eed\u4e94\u4e2a\u89e3\u7801\u6279\u6b21\u4e2d\uff0cGPU \u5168\u7a0b\u4fdd\u6301\u9ad8\u8d1f\u8f7d\uff0c\u672a\u51fa\u73b0\u4efb\u4f55\u7a7a\u95f2\u65f6\u6bb5\uff08\u6ce8\uff1a\u8be5\u6d4b\u8bd5\u57fa\u4e8e Triton attention \u540e\u7aef\uff0cFlashInfer \u540e\u7aef\u5c06\u5728\u540e\u7eed\u7248\u672c\u4e2d\u8fdb\u4e00\u6b65\u4f18\u5316\uff09\u3002<\/p>\n<p>\u501f\u52a9\u8fd9\u4e00\u4f18\u5316\uff0cSGLang v0.4 \u80fd\u591f\u5145\u5206\u6316\u6398 GPU \u7684\u8ba1\u7b97\u6f5c\u529b\uff0c\u5728 batch size \u663e\u8457\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u4e86\u76f8\u8f83\u4e8e\u4e0a\u4e00\u7248\u672c\u7684\u660e\u663e\u63d0\u5347\u3002\u5c24\u5176\u5728\u5c0f\u6a21\u578b\u548c\u5927\u89c4\u6a21\u5f20\u91cf\u5e76\u884c\u573a\u666f\u4e0b\uff0c\u8fd9\u4e00\u4f18\u5316\u6548\u679c\u5c24\u4e3a\u660e\u663e\u3002\u8be5\u8fd1\u96f6\u5f00\u9500\u6279\u8c03\u5ea6\u6280\u672f\u5df2\u9ed8\u8ba4\u542f\u7528\uff0c\u7528\u6237\u65e0\u9700\u989d\u5916\u914d\u7f6e\uff0c\u5373\u53ef\u4eab\u53d7\u6027\u80fd\u4e0a\u7684\u663e\u8457\u63d0\u5347\u3002<\/p>\n<h3>\u4e09\u3001\u591a\u6a21\u6001\u652f\u6301\uff1a\u89c6\u89c9\u4e0e\u8bed\u8a00\u7684\u534f\u540c\u52a0\u901f<\/h3>\n<p>\u5728\u591a\u6a21\u6001\u5e94\u7528\u573a\u666f\u4e2d\uff0cSGLang \u6301\u7eed\u4e0e\u56fd\u5185\u5916\u9876\u5c16\u7684\u591a\u6a21\u6001\u6280\u672f\u56e2\u961f\u6df1\u5ea6\u5408\u4f5c\uff0c\u5c06\u5148\u8fdb\u7684\u89c6\u89c9\u4e0e\u8bed\u8a00\u5904\u7406\u80fd\u529b\u65e0\u7f1d\u96c6\u6210\u5230 SGLang \u00a0\u4e2d\u3002\u73b0\u6709\u65b9\u6848\u4f7f\u5f97\u7cfb\u7edf\u80fd\u591f\u540c\u65f6\u5e94\u5bf9\u5355\u56fe\u50cf\u3001\u591a\u56fe\u50cf\u4ee5\u53ca\u89c6\u9891\u4efb\u52a1\uff0c\u5b9e\u73b0\u4e86\u5728\u4e09\u5927\u8ba1\u7b97\u673a\u89c6\u89c9\u573a\u666f\u4e2d\u7684\u5148\u8fdb\u6027\u80fd\uff0c\u4e3a\u540e\u7eed\u591a\u6a21\u6001\u5e94\u7528\u5960\u5b9a\u4e86\u575a\u5b9e\u57fa\u7840\u3002<\/p>\n<p>\u5728\u5b9e\u73b0\u4e0a\uff0cSGLang \u652f\u6301\u901a\u8fc7 OpenAI \u517c\u5bb9\u7684\u89c6\u89c9 API \u63d0\u4f9b\u670d\u52a1\u3002\u8be5\u63a5\u53e3\u80fd\u591f\u5904\u7406\u7eaf\u6587\u672c\u8f93\u5165\uff0c\u8fd8\u53ef\u4ee5\u63a5\u53d7\u4ea4\u9519\u6587\u672c\u3001\u56fe\u50cf\u548c\u89c6\u9891\u7684\u6df7\u5408\u8f93\u5165\uff0c\u6ee1\u8db3\u590d\u6742\u5e94\u7528\u573a\u666f\u4e0b\u591a\u6a21\u6001\u6570\u636e\u7684\u534f\u540c\u5904\u7406\u9700\u6c42\u3002\u7528\u6237\u65e0\u9700\u989d\u5916\u5f00\u53d1\uff0c\u5373\u53ef\u901a\u8fc7\u7edf\u4e00\u7684 API \u8c03\u7528\u4f53\u9a8c\u591a\u6a21\u6001\u63a8\u7406\u7684\u4fbf\u6377\u4e0e\u9ad8\u6548\u3002<\/p>\n<p>\u5b98\u65b9\u63d0\u4f9b\u7684 benchmark \u7ed3\u679c\u663e\u793a\uff0c\u5728 VideoDetailDescriptions \u548c LLaVA-in-the-wild \u6570\u636e\u96c6\u4e0a\uff0c\u96c6\u6210\u540e\u7684\u591a\u6a21\u6001\u6a21\u578b\u5728\u4fdd\u8bc1\u63a8\u7406\u51c6\u786e\u6027\u7684\u540c\u65f6\uff0c\u76f8\u8f83\u4e8e HuggingFace\/transformers \u7684\u539f\u59cb\u5b9e\u73b0\uff0c\u6027\u80fd\u6700\u9ad8\u53ef\u63d0\u5347 4.5 \u500d\u3002\u8fd9\u4e00\u52a0\u901f\u6548\u679c\u5f97\u76ca\u4e8e SGLang Runtime \u7684\u9ad8\u6548\u8c03\u5ea6\u548c\u8f7b\u91cf\u5316\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u7cfb\u7edf\u5728\u5904\u7406\u591a\u7c7b\u578b\u6570\u636e\u65f6\u59cb\u7ec8\u80fd\u591f\u4fdd\u6301\u8f83\u9ad8\u7684\u541e\u5410\u7387\u3002<\/p>\n<p>\u76ee\u524d\u4e3a\u6b62\uff0cSGLang \u5df2\u7ecf\u5728\u591a\u6a21\u6001\u652f\u6301\u65b9\u9762\u5c55\u793a\u4e86\u5353\u8d8a\u7684\u517c\u5bb9\u6027\u548c\u6269\u5c55\u80fd\u529b\uff0c\u540e\u7eed\u8fd8\u5c06\u9080\u8bf7\u66f4\u591a\u5f00\u53d1\u8005\u91cd\u6784\u76f8\u5173\u4ee3\u7801\u5e76\u4e14\u8fdb\u884c\u66f4\u591a\u6a21\u578b\u4e43\u81f3\u6700\u65b0\u7684 cosmos \u4e16\u754c\u6a21\u578b\u548c -o \u6d41\u5f0f\u6a21\u578b\u7684\u652f\u6301\u3002\u901a\u8fc7\u4ea4\u4e92\u5f0f\u7684\u6587\u672c\u3001\u56fe\u50cf\u548c\u89c6\u9891\u8f93\u5165\uff0cSGLang \u4e0d\u4ec5\u5927\u5e45\u63d0\u5347\u4e86\u591a\u6a21\u6001\u4efb\u52a1\u7684\u5904\u7406\u6548\u7387\uff0c\u540c\u65f6\u4e5f\u4e3a\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e0b\u7684\u590d\u6742\u6570\u636e\u534f\u540c\u8ba1\u7b97\u63d0\u4f9b\u4e86\u6709\u529b\u7684\u6280\u672f\u4fdd\u969c\u3002\u66f4\u591a\u8be6\u7ec6\u7684\u4f7f\u7528\u65b9\u6cd5\u548c\u6027\u80fd\u6570\u636e\uff0c\u8bf7\u53c2\u8003\u5b98\u65b9\u6280\u672f\u6587\u6863\u53ca benchmark \u62a5\u544a\u3002<\/p>\n<h3>\u56db\u3001X-Grammar\uff1a\u7ed3\u6784\u5316\u751f\u6210\u7684\u8303\u5f0f\u91cd\u6784<\/h3>\n<p>\u5728\u7ea6\u675f\u89e3\u7801\u9886\u57df\uff0cSGLang \u5229\u7528\u4e86 XGrammar \u7cfb\u7edf\u5728\u7ed3\u6784\u5316\u751f\u6210\u65b9\u9762\u66f4\u662f\u5b9e\u73b0\u4e86\u5168\u65b0\u7684\u8303\u5f0f\u91cd\u6784\uff0c\u663e\u8457\u7a81\u7834\u4e86\u4f20\u7edf\u7ea6\u675f\u89e3\u7801\u7684\u6027\u80fd\u74f6\u9888\u3002<\/p>\n<p>\u5728\u4e0a\u4e0b\u6587\u6269\u5c55\u65b9\u9762\uff0cXGrammar \u9488\u5bf9\u6bcf\u6761\u8bed\u6cd5\u89c4\u5219\u589e\u52a0\u4e86\u989d\u5916\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\u68c0\u6d4b\uff0c\u4ece\u800c\u6709\u6548\u51cf\u5c11\u4e86\u4e0e\u4e0a\u4e0b\u6587\u4f9d\u8d56\u76f8\u5173\u7684 token \u6570\u91cf\u3002\u8fd9\u4e00\u6539\u8fdb\u4f7f\u7cfb\u7edf\u5728\u5904\u7406\u590d\u6742\u8bed\u6cd5\u65f6\u80fd\u591f\u66f4\u65e9\u8bc6\u522b\u5e76\u5229\u7528\u89c4\u5219\u9690\u542b\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u89e3\u7801\u8fc7\u7a0b\u4e2d\u4e0d\u5fc5\u8981\u7684\u72b6\u6001\u5207\u6362\u5f00\u9500\u3002<\/p>\n<p>\u4e3a\u4e86\u9ad8\u6548\u7ba1\u7406\u591a\u6761\u6269\u5c55\u8def\u5f84\u4ea7\u751f\u7684\u6267\u884c\u72b6\u6001\uff0cXGrammar \u91c7\u7528\u4e86\u57fa\u4e8e\u6811\u7ed3\u6784\u7684\u6570\u636e\u7ec4\u7ec7\u65b9\u5f0f\uff0c\u6784\u5efa\u4e86\u6301\u4e45\u5316\u6267\u884c\u6808\u3002\u8be5\u8bbe\u8ba1\u4e0d\u4ec5\u80fd\u591f\u9ad8\u6548\u5730\u7ba1\u7406\u591a\u4e2a\u6267\u884c\u6808\uff0c\u8fd8\u53ef\u4ee5\u5728\u9762\u5bf9\u62c6\u5206\u4e0e\u5408\u5e76\u64cd\u4f5c\u65f6\u4fdd\u6301\u6570\u636e\u7ed3\u6784\u7684\u7a33\u5b9a\u6027\u548c\u9ad8\u6548\u6027\uff0c\u786e\u4fdd\u6574\u4e2a\u89e3\u7801\u6d41\u7a0b\u59cb\u7ec8\u6d41\u7545\u8fd0\u884c\u3002<\/p>\n<p>\u5728\u4e0b\u63a8\u81ea\u52a8\u673a\u7ed3\u6784\u4f18\u5316\u65b9\u9762\uff0cXGrammar \u501f\u9274\u4e86\u7f16\u8bd1\u5668\u8bbe\u8ba1\u4e2d\u7684\u5185\u8054\u4f18\u5316\u548c\u7b49\u4ef7\u72b6\u6001\u5408\u5e76\u6280\u672f\uff0c\u5bf9\u81ea\u52a8\u673a\u4e2d\u7684\u8282\u70b9\u8fdb\u884c\u7cbe\u7b80\u3002\u901a\u8fc7\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u72b6\u6001\u8282\u70b9\uff0c\u7cfb\u7edf\u80fd\u591f\u66f4\u8fc5\u901f\u5730\u5b8c\u6210\u8bed\u6cd5\u89c4\u5219\u7684\u5339\u914d\u4e0e\u8f6c\u6362\uff0c\u4ece\u800c\u663e\u8457\u63d0\u5347\u4e86\u89e3\u7801\u6548\u7387\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4e3a\u5145\u5206\u53d1\u6325\u591a\u6838 CPU \u7684\u8ba1\u7b97\u80fd\u529b\uff0cXGrammar \u5bf9\u8bed\u6cd5\u7f16\u8bd1\u8fc7\u7a0b\u8fdb\u884c\u4e86\u5e76\u884c\u5316\u5904\u7406\u3002\u8bed\u6cd5\u89c4\u5219\u7684\u7f16\u8bd1\u4efb\u52a1\u88ab\u5206\u914d\u5230\u591a\u4e2a CPU \u6838\u5fc3\u4e0a\u540c\u65f6\u6267\u884c\uff0c\u4e0d\u4ec5\u5927\u5e45\u7f29\u77ed\u4e86\u7f16\u8bd1\u65f6\u95f4\uff0c\u4e5f\u4e3a\u540e\u7eed\u591a\u4efb\u52a1\u89e3\u6790\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u57fa\u7840\u3002<\/p>\n<p>\u7efc\u5408\u4e0a\u8ff0\u5404\u9879\u4f18\u5316\u63aa\u65bd\uff0cXGrammar \u6280\u672f\u7684\u96c6\u6210\uff0c\u4f7f SGLang \u5728 JSON \u89e3\u7801\u7b49\u7ea6\u675f\u89e3\u7801\u4efb\u52a1\u4e0a\u5b9e\u73b0\u4e86 10 \u500d\u7684\u52a0\u901f\u6548\u679c\u3002\u5728\u5904\u7406\u590d\u6742\u7ed3\u6784\u5316\u6570\u636e\u548c\u5de5\u5177\u8c03\u7528\u573a\u666f\u65f6\uff0cXGrammar \u4e0d\u4ec5\u5927\u5e45\u964d\u4f4e\u4e86\u89e3\u7801\u5ef6\u8fdf\uff0c\u8fd8\u4e3a\u5927\u89c4\u6a21\u5728\u7ebf\u670d\u52a1\u63d0\u4f9b\u4e86\u53ef\u9760\u7684\u6027\u80fd\u4fdd\u969c\u3002<\/p>\n<p>\u6709\u5173 XGrammar \u7684\u8fdb\u4e00\u6b65\u4ecb\u7ecd\uff0cSGLang \u56e2\u961f\u5df2\u5728\u5b98\u65b9\u535a\u5ba2\u4e2d\u8fdb\u884c\u4e86\u6df1\u5165\u63a2\u8ba8\uff0c\u76f8\u5173\u6280\u672f\u6587\u6863\u53ef\u4f9b\u53c2\u8003\u3002<\/p>\n<h3>\u4e94\u3001Cache-Aware Load Balancer\uff1a\u667a\u80fd\u8def\u7531\u7684\u67b6\u6784\u7a81\u7834<\/h3>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s6.51cto.com\/oss\/202503\/07\/798b4ce248e659128dd835fa0d19a573659b2a.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u5728 SGLang v0.4 \u4e2d\uff0c\u5f15\u5165\u4e86\u72ec\u51fa\u5fc3\u88c1\u7684\u5168\u65b0 Cache-Aware Load Balancer\uff0c\u4e3a\u5927\u6a21\u578b\u63a8\u7406\u7cfb\u7edf\u63d0\u4f9b\u4e86\u667a\u80fd\u8def\u7531\u7684\u67b6\u6784\u7a81\u7834\uff0c\u5168\u90e8\u4ee5 Rust \u7f16\u5199\uff0c\u76f8\u6bd4\u4e8e Python \u5927\u5e45\u51cf\u5c11 Service Overhead\u3002\u8be5\u8d1f\u8f7d\u5747\u8861\u5668\u91c7\u7528\u57fa\u4e8e\u5b57\u7b26\u7ea7\u524d\u7f00\u5339\u914d\u7684\u8def\u7531\u7b97\u6cd5\uff0c\u901a\u8fc7\u5408\u5e76\u540e\u7684 Radix Tree \u5b9e\u73b0\u65e0\u9700 Tokenization \u7684\u5339\u914d\u3002\u7cfb\u7edf\u80fd\u591f\u6839\u636e\u5404\u5de5\u4f5c\u8282\u70b9\u7684\u524d\u7f00 KV \u7f13\u5b58\u547d\u4e2d\u7387\u8fdb\u884c\u52a8\u6001\u8bc4\u4f30\uff0c\u5e76\u81ea\u52a8\u9009\u62e9\u7f13\u5b58\u547d\u4e2d\u7387\u8f83\u9ad8\u7684\u8282\u70b9\u6765\u5904\u7406\u8bf7\u6c42\u3002\u4e0e\u4f20\u7edf\u7684\u8f6e\u8be2\u8c03\u5ea6\u65b9\u5f0f\u76f8\u6bd4\uff0c\u6b64\u65b9\u6848\u5728\u5b9e\u9645\u6d4b\u8bd5\u4e2d\u5c55\u793a\u4e86\u6700\u9ad8\u53ef\u8fbe\u5c06\u8fd1\u4e24\u500d\u7684\u541e\u5410\u91cf\u63d0\u5347\uff0c\u4ee5\u53ca\u5c06\u8fd1\u56db\u500d\u7684\u7f13\u5b58\u547d\u4e2d\u7387\u6539\u8fdb\u3002\u968f\u7740\u5de5\u4f5c\u8282\u70b9\u6570\u91cf\u7684\u589e\u52a0\uff0c\u8fd9\u79cd\u4f18\u52bf\u66f4\u4e3a\u660e\u663e\uff0c\u5145\u5206\u4f53\u73b0\u4e86\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\u5728\u591a\u8282\u70b9\u5206\u5e03\u5f0f\u90e8\u7f72\u4e2d\u7684\u6269\u5c55\u6027\u3002<\/p>\n<p>\u4e3a\u4e86\u6709\u6548\u7ba1\u7406\u7f13\u5b58\u8d44\u6e90\uff0cSGLang \u7684\u8d1f\u8f7d\u5747\u8861\u5668\u5185\u90e8\u5f15\u5165\u4e86\u61d2\u66f4\u65b0\u7684 LRU \u6dd8\u6c70\u7b56\u7565\uff0c\u5bf9\u8fd1\u4f3c Radix Tree \u4e2d\u8bbf\u95ee\u9891\u7387\u8f83\u4f4e\u7684\u53f6\u5b50\u8282\u70b9\u8fdb\u884c\u5b9a\u671f\u6e05\u7406\uff0c\u4ece\u800c\u9632\u6b62\u5185\u5b58\u8fc7\u5ea6\u81a8\u80c0\u5e76\u4fdd\u6301\u6811\u7ed3\u6784\u7684\u9ad8\u6548\u6027\u3002\u6b64\u4e3e\u4e0d\u4ec5\u4f18\u5316\u4e86\u5185\u5b58\u4f7f\u7528\uff0c\u8fd8\u4e3a\u6574\u4e2a\u63a8\u7406\u7cfb\u7edf\u5e26\u6765\u4e86\u66f4\u7a33\u5b9a\u7684\u7f13\u5b58\u6027\u80fd\u3002\u5728\u5206\u5e03\u5f0f\u90e8\u7f72\u573a\u666f\u4e0b\uff0c\u7cfb\u7edf\u901a\u8fc7 HTTP \u63a5\u53e3\u5b9e\u73b0\u4e86\u79d2\u7ea7\u52a8\u6001\u6269\u7f29\u5bb9\uff0c\u5141\u8bb8\u5728\u96c6\u7fa4\u4e2d\u5feb\u901f\u589e\u51cf\u5de5\u4f5c\u8282\u70b9\u3002\u5f97\u76ca\u4e8e\u8fd9\u4e00\u667a\u80fd\u8def\u7531\u8bbe\u8ba1\uff0cSGLang \u5728\u591a\u8282\u70b9\u96c6\u7fa4\u4e2d\u7684\u541e\u5410\u6027\u80fd\u5448\u73b0\u51fa\u8fd1\u7ebf\u6027\u7684\u6269\u5c55\u8d8b\u52bf\uff0c\u4e3a\u5927\u89c4\u6a21\u5728\u7ebf\u670d\u52a1\u63d0\u4f9b\u4e86\u575a\u5b9e\u7684\u6027\u80fd\u548c\u53ef\u9760\u6027\u4fdd\u969c\u3002<\/p>\n<h3>\u516d\u3001\u5f00\u53d1\u8005\u5de5\u5177\u94fe<\/h3>\n<p>\u5728\u53ef\u7528\u6027\u548c\u6613\u7528\u6027\u65b9\u9762\uff0cSGLang \u63d0\u4f9b\u4e86\u4e0e OpenAI API \u517c\u5bb9\u7684\u63a5\u53e3\u5c42\uff0c\u652f\u6301 Chat\u3001Completions\u3001Embeddings \u7b49\u5e38\u89c1\u529f\u80fd\uff0c\u5f00\u53d1\u8005\u4ec5\u9700\u66ff\u6362\u7aef\u70b9\u5373\u53ef\u5feb\u901f\u65e0\u7f1d\u8fc1\u79fb\u3002\u5bf9\u4e8e\u66f4\u7075\u6d3b\u7684\u90e8\u7f72\u65b9\u5f0f\uff0c\u79bb\u7ebf\u5f15\u64ce\u6a21\u5f0f\uff08Offline Engine\uff09\u5141\u8bb8\u5355\u811a\u672c\u540c\u65f6\u9a71\u52a8\u591a\u8282\u70b9\u63a8\u7406\uff0c\u65e0\u9700\u72ec\u7acb\u670d\u52a1\u5316\uff0c\u4ece\u800c\u5927\u5e45\u7b80\u5316\u4e86\u8fd0\u7ef4\u6210\u672c\u3002<\/p>\n<p>\u4e3a\u4e86\u8ba9\u5f00\u53d1\u8005\u80fd\u591f\u6df1\u5165\u4e86\u89e3\u6a21\u578b\u72b6\u6001\u5e76\u8fdb\u884c\u7cbe\u7ec6\u8c03\u4f18\uff0cSGLang \u5185\u7f6e\u4e86 Prometheus \u76d1\u63a7\u96c6\u6210\uff0c\u5b9e\u65f6\u8ffd\u8e2a\u541e\u5410\u91cf\uff08Throughput\uff09\u3001\u5ef6\u8fdf\uff08Latency\uff09\u548c\u663e\u5b58\u4f7f\u7528\uff08GPU Memory Pressure\uff09\u7b49\u6838\u5fc3\u6307\u6807\uff1b\u591a LoRA \u52a8\u6001\u52a0\u8f7d\uff08Dynamic LoRA Switching\uff09\u5219\u8ba9\u540c\u4e00\u670d\u52a1\u53ef\u5728\u663e\u5b58\u590d\u7528\u7387\u9ad8\u8fbe 90% \u7684\u60c5\u51b5\u4e0b\uff0c\u70ed\u5207\u6362\u591a\u4e2a\u4e0d\u540c\u7684 LoRA \u9002\u914d\u5668\uff08Low-Rank Adaptation\uff09\uff1b\u800c\u7ea6\u675f\u89e3\u7801\uff08Constrained Decoding\uff09\u63d0\u4f9b\u4e86 JSON\u3001GBNF \u7b49\u683c\u5f0f\u7684\u5f3a\u5236\u6821\u9a8c\u80fd\u529b\uff0c\u5c06\u751f\u6210\u9519\u8bef\u7387\u964d\u81f3\u6781\u4f4e\u6c34\u5e73\uff0c\u6ee1\u8db3\u751f\u4ea7\u573a\u666f\u5bf9\u8f93\u51fa\u683c\u5f0f\u7684\u4e00\u81f4\u6027\u8981\u6c42\u3002<\/p>\n<h3>\u4e03\u3001\u793e\u533a\u4e0e\u672a\u6765\u89c4\u5212<\/h3>\n<p>\u76ee\u524d\uff0cSGLang \u5728\u5168\u7403\u8303\u56f4\u5185\u5df2\u7ecf\u6c47\u805a\u4e86 30 \u4f59\u4f4d\u6838\u5fc3\u8d21\u732e\u8005\u3002\u5728\u63a5\u4e0b\u6765\u7684 2025 H1 \u9636\u6bb5\u4e2d\uff0c\u56e2\u961f\u5c06\u81f4\u529b\u4e8e\u5b8c\u5584\u5b9e\u6218\u573a\u666f\u4e0b\u7684 PD \u5206\u79bb\u3001Speculative Decoding \u7684\u957f\u6587\u672c\u4f18\u5316\u3001\u63a8\u52a8\u591a\u7ea7\u7f13\u5b58\uff08GPU\/CPU\/Disk\uff09\u7b56\u7565\u843d\u5730\uff0c\u5e76\u7ee7\u7eed\u5f3a\u5316\u5e76\u884c\u7b56\u7565\u4ee5\u9002\u914d\u5343\u4ebf\u7ea7 MoE \u6a21\u578b\u3002\u9664\u5f00\u672c\u8eab\u63a8\u7406\u6548\u679c\u7684\u4f18\u5316\uff0cSGLang \u56e2\u961f\u4e5f\u5c06\u81f4\u529b\u63a8\u7406\u5f15\u64ce\u7684\u5e7f\u6cdb\u843d\u5730\uff0c\u7ee7\u7eed\u652f\u6301 RAG\u3001multi-Agent\u3001Reasoning\u3001RLHF \u7b49\u7b49\u9886\u57df\u7684 AI \u843d\u5730\u3002\u6700\u540e\uff0cSGLang \u4e5f\u5c06\u5728\u7b97\u5b50\u8986\u76d6\u7387\u4e0e\u6027\u80fd\u4e0a\u6301\u7eed\u4f18\u5316\uff0c\u652f\u6301\u66f4\u591a\u7684\u66f4\u5e7f\u6cdb\u7684\u786c\u4ef6\uff0c\u529b\u4e89\u4e3a\u5f00\u6e90\u793e\u533a\u63d0\u4f9b\u66f4\u52a0\u5148\u8fdb\u7684\u4e00\u7ad9\u5f0f\u5927\u6a21\u578b\u63a8\u7406\u65b9\u6848\u3002<\/p>\n<h3>\u516b\u3001\u65b0\u7684\u4e00\u5e74\uff0c\u4e0e\u793e\u533a\u5171\u8d74\u661f\u8fb0\u5927\u6d77<\/h3>\n<p><img decoding=\"async\" title=\"\u56fe\u7247\" src=\"https:\/\/s3.51cto.com\/oss\/202503\/07\/636a6cf78c1f40d18102160759d9adeb830175.webp\" alt=\"\u56fe\u7247\" data-type=\"inline\" \/><\/p>\n<p>\u5f00\u6e90\u4e00\u5468\u5e74\uff0cSGLang \u7684\u6210\u957f\u8f68\u8ff9\u5370\u8bc1\u4e86\u4e00\u4e2a\u6280\u672f\u771f\u7406\uff1a<strong>\u9876\u5c16\u7684\u5de5\u7a0b\u5b9e\u8df5\uff0c\u6c38\u8fdc\u8bde\u751f\u4e8e\u5f00\u53d1\u8005\u793e\u533a\u7684\u534f\u4f5c\u5171\u632f\u3002<\/strong>\u4ece\u9996\u4e2a\u652f\u6301 Prefix Cache \u7684\u63a8\u7406\u6846\u67b6\uff0c\u5230\u65a9\u83b7 11K Star\u3001\u6708\u5747 10 \u4e07\u4e0b\u8f7d\u91cf\u7684\u5f00\u6e90\u660e\u661f\uff1b\u4ece xAI\u3001NVIDIA\u3001AMD \u7b49\u5de8\u5934\u7684\u6df1\u5ea6\u96c6\u6210\uff0c\u5230\u4e3a DeepSeek \u6a21\u578b\u7684\u6700\u4f73\u5f00\u6e90\u63a8\u7406\u5f15\u64ce \u2014\u2014<strong>SGLang \u7684\u6bcf\u4e00\u6b21\u6280\u672f\u7a81\u7834\uff0c\u90fd\u6e90\u4e8e\u793e\u533a\u5f00\u53d1\u8005\u7684\u771f\u5b9e\u9700\u6c42\u4e0e\u5171\u521b\u667a\u6167\u3002<\/strong><\/p>\n<p>\u5728 SGLang \u7684\u4ee3\u7801\u4ed3\u5e93\u4e2d\uff0c200+ \u793e\u533a\u8d21\u732e\u4e0d\u4ec5\u5e26\u6765\u4e86 Multi-head Latent Attention\u3001Block-wise FP8 \u7b49\u6838\u5fc3\u521b\u65b0\uff0c\u66f4\u50ac\u751f\u4e86\u5f00\u53d1\u8005\u5de5\u5177\u94fe\u7684\u5168\u9762\u8fdb\u5316\uff1a\u652f\u6301\u591a\u6a21\u6001\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u3001\u70ed\u5207\u6362 LoRA \u7684\u663e\u5b58\u590d\u7528\u3001JSON \u7ed3\u6784\u5316\u751f\u6210\u7684\u6781\u901f\u6821\u9a8c\u2026\u2026 \u8fd9\u4e9b\u80fd\u529b\u80cc\u540e\uff0c\u662f\u4e09\u5341\u4f59\u4f4d\u6838\u5fc3\u8d21\u732e\u8005\u4e0e\u6570\u767e\u5f00\u53d1\u8005\u7684\u6280\u672f\u63a5\u529b\u3002\u6b63\u5982 LMSYS Org \u7ec4\u7ec7\u79c9\u6301\u7684\u6838\u5fc3\u7406\u5ff5\uff0c\u6280\u672f\u751f\u6001\u7684\u7e41\u8363\uff0c\u4ece\u4e0d\u662f\u5355\u6253\u72ec\u6597\u7684\u5947\u8ff9\u3002\u5f53\u6211\u4eec\u770b\u5230 SGLang \u5728 DeepSeek-R1 \u4e0a\u5b9e\u73b0\u72c2\u98d9\u5f0f\u541e\u5410\uff0c\u5728 128k \u957f\u6587\u672c\u4e2d\u8fbe\u5230\u8d85\u4f4e\u5ef6\u8fdf\uff0c\u8fd9\u4e0d\u4ec5\u662f\u6846\u67b6\u7684\u80dc\u5229\uff0c\u66f4\u662f\u5f00\u6e90\u793e\u533a \u201c\u767e\u82b1\u9f50\u653e\u201d \u7684\u660e\u8bc1 \u2014\u2014 \u4ece LinkedIn \u7684\u5206\u5e03\u5f0f\u90e8\u7f72\u5b9e\u8df5\uff0c\u5230 NVIDIA \u7684 Triton \u5185\u6838\u4f18\u5316\uff0c\u518d\u5230\u666e\u901a\u5f00\u53d1\u8005\u63d0\u4ea4\u7684\u6570\u767e\u4e2a\u6027\u80fd\u8c03\u4f18 PR\uff0c\u6bcf\u4e2a\u53c2\u4e0e\u8005\u90fd\u5728\u4e3a\u5927\u6a21\u578b\u843d\u5730\u8fd9\u4e00\u7ec8\u6781\u76ee\u6807\u6dfb\u7816\u52a0\u74e6\u3002<\/p>\n<p>\u4e00\u679d\u72ec\u79c0\u4e0d\u662f\u6625\uff0c\u767e\u82b1\u9f50\u653e\u6625\u6ee1\u56ed\u3002SGLang \u7684\u5f00\u6e90\u6545\u4e8b\uff0c\u6b63\u5728\u8be0\u91ca\u4e00\u4e2a\u5c5e\u4e8e\u5f00\u53d1\u8005\u7684\u9ec4\u91d1\u65f6\u4ee3\uff1a\u5728\u8fd9\u91cc\uff0c\u5de5\u7a0b\u5e08\u7684\u6bcf\u4e00\u884c\u4ee3\u7801\u90fd\u53ef\u80fd\u5f71\u54cd\u6570\u767e\u4e07\u7528\u6237\u7684\u4f53\u9a8c\uff0c\u7814\u7a76\u8005\u7684\u6bcf\u4e2a\u521b\u610f\u90fd\u6709\u673a\u4f1a\u91cd\u5851\u6280\u672f\u8fb9\u754c\u3002SGLang \u56e2\u961f\u4e5f\u8bda\u9080\u6bcf\u4e00\u4f4d\u611f\u5174\u8da3\u7684\u670b\u53cb\u52a0\u5165\u793e\u533a\uff0c\u53c2\u4e0e Slack \u4e0e Github \u4e0a\u7684\u6280\u672f\u8ba8\u8bba\uff0c\u5728\u5168\u7403 Meetup \u4e2d\u5206\u4eab\u5b9e\u6218\u6d1e\u5bdf \u2014\u2014 \u56e0\u4e3a\u4e0b\u4e00\u4e2a\u6539\u53d8\u884c\u4e1a\u7684\u4f18\u5316\u65b9\u6848\uff0c\u6216\u8bb8\u5c31\u8bde\u751f\u4e8e\u4f60\u7684\u4e00\u6b21 git commit\u3002<\/p>\n<ul data-id=\"ua73dd4b-nMd6f9OG\">\n<li data-id=\"ld70c578-plZ0ZftQ\">GitHub \u4ed3\u5e93:\u00a0https:\/\/github.com\/sgl-project\/sglang<\/li>\n<li data-id=\"ld70c578-KOBstaFe\">Slack \u793e\u533a\uff1aslack.sglang.ai<\/li>\n<li data-id=\"ld70c578-Fd5Q9S2y\">DeepSeek \u4f18\u5316\u6307\u5357: https:\/\/docs.sglang.ai\/references\/deepseek.html<\/li>\n<\/ul>\n<p>\u7acb\u523b\u4f53\u9a8c\u6700\u65b0\u7248\u672c\u7684 SGLang\uff0c\u8ba9\u5927\u6a21\u578b\u63a8\u7406\u518d\u4e0a\u65b0\u53f0\u9636\uff01<\/p>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_24943\" class=\"pvc_stats total_only  \" data-element-id=\"24943\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u4ece DeepSeek \u6a21\u578b\u53d1\u5e03\u5f53\u5929\u4fbf\u5b9e\u73b0\u6700\u4f73\u9002\u914d\uff0c\u5230\u957f\u671f\u7a33\u5c45 SOTA \u6027\u80fd\u699c\u9996\uff0cSGLang \u7684\u8fdb\u5316\u8f68\u8ff9\u63ed\u793a [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_24943\" class=\"pvc_stats total_only  \" data-element-id=\"24943\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":24945,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3,23,20,80],"tags":[124,880,945],"class_list":["post-24943","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-23","category-20","category-80","tag-ai","tag-880","tag-945"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24943","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=24943"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24943\/revisions"}],"predecessor-version":[{"id":24946,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/24943\/revisions\/24946"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/24945"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=24943"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=24943"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=24943"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}