{"id":15563,"date":"2026-10-30T15:05:10","date_gmt":"2026-10-30T07:05:10","guid":{"rendered":"https:\/\/ai-stack.ai\/?p=15563"},"modified":"2026-10-30T15:13:05","modified_gmt":"2026-10-30T07:13:05","slug":"what-is-llm-benchmark","status":"publish","type":"post","link":"https:\/\/ai-stack.ai\/zh-hant\/what-is-llm-benchmark","title":{"rendered":"\u5341\u5206\u9418\u770b\u61c2 LLM Benchmark\uff1a\u5ef6\u9072\u3001\u541e\u5410\u91cf\u8207\u8ca0\u8f09\u6a21\u578b"},"content":{"rendered":"\n

2026\/10\/22   INFINITIX Product Dep.<\/p>\n\n\n\n

\u524d\u8a00\uff1a\u5982\u4f55\u56de\u7b54\u300c\u9019\u53f0\u4f3a\u670d\u5668\u80fd\u540c\u6642\u670d\u52d9\u591a\u5c11\u4eba\uff1f\u300d<\/strong><\/h1>\n\n\n\n

\u5718\u968a\u5728\u4e00\u53f0 GPU \u4f3a\u670d\u5668\u4e0a\u90e8\u7f72\u4e86\u958b\u6e90\u5927\u578b\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u3002\u4f60\u7528 curl \u9001\u51fa\u4e00\u6bb5 prompt\uff08\u8f38\u5165\u63d0\u793a\uff09\uff0c\u4e00\u79d2\u591a\u5c31\u958b\u59cb\u56de\u50b3\u6587\u5b57\uff0c\u770b\u8d77\u4f86\u5f88\u5feb\u3002\u4f46\u300c\u770b\u8d77\u4f86\u5f88\u5feb\u300d\u4e26\u4e0d\u80fd\u56de\u7b54\u4ee5\u4e0b\u9019\u4e9b\u554f\u984c\uff1a\u7cfb\u7d71\u80fd\u540c\u6642\u627f\u53d7 10 \u4eba\u7684\u8acb\u6c42\uff0c\u9084\u662f 200 \u4eba\u7684\u8acb\u6c42\uff1f\u4f7f\u7528\u8005\u5728\u5e7e\u79d2\u5167\u770b\u5230\u7b2c\u4e00\u500b\u56de\u8986\u6587\u5b57\u51fa\u73fe\u624d\u7b97\u5408\u683c\uff1f\u4ee5\u4e32\u6d41\u65b9\u5f0f\u986f\u793a\u56de\u8986\u6587\u5b57\u7684\u901f\u5ea6\u8981\u591a\u5feb\u624d\u7b97\u6d41\u66a2\uff1f<\/p>\n\n\n\n

\u8981\u56de\u7b54\u4ee5\u4e0a\u9019\u985e\u554f\u984c\uff0c\u5c31\u9700\u8981\u9032\u884c LLM benchmark\uff08\u6548\u80fd\u57fa\u6e96\u6e2c\u8a66\uff09\u3002\u672c\u6587\u4e0d\u4ecb\u7d39\u5de5\u5177\u5b89\u88dd\uff0c\u800c\u662f\u805a\u7126\u6e2c\u8a66\u80cc\u5f8c\u7684\u908f\u8f2f\uff1a\u8a72\u91cf\u6e2c\u54ea\u4e9b\u6307\u6a19\u3001\u70ba\u4ec0\u9ebc\u4e0d\u80fd\u53ea\u770b\u55ae\u4e00\u6578\u5b57\u3001\u8a72\u5982\u4f55\u65bd\u52a0\u8ca0\u8f09\uff0c\u4ee5\u53ca\u54ea\u4e9b\u7d30\u7bc0\u53ef\u80fd\u8b93\u6e2c\u8a66\u7d50\u679c\u5931\u771f\u3002\u6587\u4e2d\u4e26\u6703\u4ee5 NVIDIA \u958b\u6e90\u7684 AIPerf \u5de5\u5177\u70ba\u4f8b\uff0c\u85c9\u6b64\u8aaa\u660e\u9019\u4e9b\u6982\u5ff5\u5728\u5be6\u969b\u64cd\u4f5c\u4e2d\u8a72\u5982\u4f55\u5448\u73fe\u3002<\/p>\n\n\n\n

\u4e00\u3001\u4ec0\u9ebc\u662f LLM Benchmark\uff1f<\/strong><\/h1>\n\n\n\n

\u300cLLM benchmark\u300d\u5e38\u898b\u6709\u5169\u7a2e\u610f\u601d\u3002\u80fd\u529b\u8a55\u6e2c benchmark<\/strong>\uff08\u4f8b\u5982 MMLU\u3001HumanEval \u7b49\u984c\u5eab\uff09\u95dc\u5fc3\u7684\u662f\u6a21\u578b\u7b54\u5f97\u5c0d\u4e0d\u5c0d\uff0c\u800c\u4e0d\u662f\u5728\u8a55\u4f30\u786c\u9ad4\u6548\u80fd\uff1b\u63a8\u8ad6\u6548\u80fd benchmark<\/strong>\u5247\u95dc\u5fc3\u540c\u4e00\u500b\u6a21\u578b\u90e8\u7f72\u5728\u7279\u5b9a\u786c\u9ad4\u8207\u63a8\u8ad6\u5f15\u64ce\u5f8c\uff0c\u56de\u61c9\u901f\u5ea6\u6709\u591a\u5feb\uff0c\u4ee5\u53ca\u53ef\u4ee5\u540c\u6642\u670d\u52d9\u591a\u5c11\u4f7f\u7528\u8005\u3002\u672c\u6587\u8a0e\u8ad6\u7684\u662f\u5f8c\u8005\uff0c\u4e5f\u5c31\u662f\u5927\u5bb6\u771f\u6b63\u60f3\u77e5\u9053\u7684\u4e8b\u3002<\/p>\n\n\n\n

\u4e0d\u904e\uff0c\u9019\u985e\u6e2c\u8a66\u4e0d\u80fd\u76f4\u63a5\u7167\u642c\u4e00\u822c HTTP \u58d3\u529b\u6e2c\u8a66\u7684\u505a\u6cd5\uff0c\u56e0\u70ba LLM \u63a8\u8ad6\u548c\u50b3\u7d71 API \u6709\u5e7e\u500b\u91cd\u8981\u5dee\u7570\uff1a<\/p>\n\n\n\n

    \n
  • \u56de\u61c9\u6703\u4ee5\u4e32\u6d41\u65b9\u5f0f\u9010\u500b token \u50b3\u56de\u3002<\/strong>\u7b2c\u4e00\u500b token\uff08Time to First Token\uff0cTTFT\uff09\u4f55\u6642\u51fa\u73fe\u3001\u5f8c\u7e8c token \u9593\u9694\uff08Inter-Token Latency\uff0cITL\uff09\u591a\u4e45\uff0c\u4ee5\u53ca\u6574\u6bb5\u56de\u7b54\uff08End-to-End Request Latency\uff0cE2E Latency\uff09\u4f55\u6642\u5b8c\u6210\uff0c\u662f\u4e09\u7a2e\u4e0d\u540c\u7684\u6642\u9593\u3002<\/li>\n\n\n\n
  • \u6bcf\u500b\u8acb\u6c42\u7684\u5de5\u4f5c\u91cf\u4e0d\u56fa\u5b9a\u3002<\/strong>\u6a21\u578b\u53ef\u80fd\u5728\u4e0d\u540c\u6642\u9593\u9ede\u505c\u6b62\u751f\u6210\uff1b\u8f38\u51fa\u8d8a\u9577\uff0c\u901a\u5e38\u5360\u7528 GPU \u7684\u6642\u9593\u4e5f\u8d8a\u4e45\u3002\u5982\u679c\u5169\u6b21\u6e2c\u8a66\u7684\u8f38\u51fa\u9577\u5ea6\u5206\u5e03\u4e0d\u540c\uff0c\u541e\u5410\u91cf\u5c31\u96e3\u4ee5\u76f4\u63a5\u6bd4\u8f03\u3002<\/li>\n\n\n\n
  • \u904b\u7b97\u5206\u6210 prefill \u8207 decode \u5169\u500b\u968e\u6bb5\u3002<\/strong>Prefill \u6703\u8655\u7406\u6574\u6bb5\u8f38\u5165 prompt\uff1bdecode \u5247\u9010\u6b65\u7522\u751f\u8f38\u51fa token\u3002\u5169\u500b\u968e\u6bb5\u7684\u8cc7\u6e90\u9700\u6c42\u548c\u6548\u80fd\u74f6\u9838\u4e0d\u540c\u3002<\/li>\n\n\n\n
  • \u4e0d\u540c\u8acb\u6c42\u6703\u4e92\u76f8\u5f71\u97ff\u3002<\/strong>\u63a8\u8ad6\u5f15\u64ce\u53ef\u80fd\u900f\u904e continuous batching\uff08\u9023\u7e8c\u6279\u6b21\u8655\u7406\uff09\uff0c\u628a\u591a\u500b\u8acb\u6c42\u7684 decode \u5de5\u4f5c\u5408\u4f75\u6392\u7a0b\u3002\u56e0\u6b64\uff0c\u55ae\u4e00\u8acb\u6c42\u7522\u751f token \u7684\u901f\u5ea6\uff0c\u4e5f\u6703\u53d7\u5230\u540c\u4e00\u6642\u9593\u5176\u4ed6\u8acb\u6c42\u7684\u5f71\u97ff\u3002<\/li>\n\n\n\n
  • GPU \u8a18\u61b6\u9ad4\u53ef\u80fd\u5148\u6210\u70ba\u74f6\u9838\u3002<\/strong>\u6bcf\u500b\u5c1a\u672a\u5b8c\u6210\u7684\u8acb\u6c42\u90fd\u6703\u5360\u7528 KV cache\uff0c\u9700\u6c42\u4e5f\u6703\u96a8\u5e8f\u5217\u9577\u5ea6\u589e\u52a0\u3002\u7576\u8a18\u61b6\u9ad4\u63a5\u8fd1\u6216\u9054\u5230\u4e0a\u9650\u6642\uff0c\u5ef6\u9072\u53ef\u80fd\u7a81\u7136\u60e1\u5316\uff0c\u800c\u4e0d\u662f\u9010\u6b65\u8b8a\u6162\u3002<\/li>\n<\/ul>\n\n\n\n

    \u4e00\u6b21 benchmark \u81f3\u5c11\u8981\u5148\u56fa\u5b9a\u4e94\u4ef6\u4e8b\uff1a\u6e2c\u8a66\u76ee\u6a19\uff08API \u670d\u52d9\u7aef\u9ede endpoint\u3001\u6a21\u578b\u3001\u662f\u5426\u555f\u7528\u4e32\u6d41\uff09\u3001\u5de5\u4f5c\u8ca0\u8f09\uff08prompt \u4f86\u6e90\u3001\u8f38\u5165\u5e8f\u5217\u9577\u5ea6 (Input Sequence Length\uff0cISL ) \u8207\u8f38\u51fa\u5e8f\u5217\u9577\u5ea6 (Output Sequence Length\uff0cOSL)\uff09\u3001\u8ca0\u8f09\u6a21\u578b\uff08\u540c\u6642\u6709\u591a\u5c11\u8acb\u6c42\u3001\u4ee5\u4ec0\u9ebc\u7bc0\u594f\u9001\u51fa\uff09\u3001\u91cf\u6e2c\u6307\u6a19\uff0c\u4ee5\u53ca\u5224\u5b9a\u6a19\u6e96 SLO\uff08Service Level Objective\uff0c\u670d\u52d9\u7b49\u7d1a\u76ee\u6a19\uff09\u3002\u628a\u5b83\u7576\u6210\u53d7\u63a7\u5be6\u9a57\uff0c\u4e00\u6b21\u53ea\u8abf\u6574\u4e00\u500b\u8b8a\u56e0\uff0c\u624d\u5bb9\u6613\u5224\u65b7\u7d50\u679c\u6539\u8b8a\u7684\u539f\u56e0\u3002<\/p>\n\n\n\n

    \u4e8c\u3001\u70ba\u4ec0\u9ebc\u4e0d\u80fd\u53ea\u770b\u55ae\u4e00\u6578\u5b57<\/strong><\/h1>\n\n\n\n

    2.1 \u540c\u4e00\u689d\u6642\u9593\u8ef8\u4e0a\u7684\u56db\u500b\u6838\u5fc3\u6307\u6a19<\/strong><\/h2>\n\n\n\n

    \u6240\u6709\u6307\u6a19\u90fd\u4f86\u81ea\u540c\u4e00\u689d\u6642\u9593\u8ef8\uff1at\u2080 \u662f\u9001\u51fa\u8acb\u6c42\u7684\u6642\u9593\uff0ct\u2081 \u662f\u6536\u5230\u7b2c\u4e00\u500b token \u7684\u6642\u9593\uff1b\u4e4b\u5f8c\u6703\u9678\u7e8c\u6536\u5230\u66f4\u591a token\uff0c\u76f4\u5230 t\u2099 \u6536\u5230\u6700\u5f8c\u4e00\u500b token\u3002<\/p>\n\n\n\n

    \"AI\u8acb\u6c42\u6d41\u7a0b\u6642\u9593\u8ef8\"<\/figure>\n\n\n\n

    \u5716 1\u3000\u5f9e\u7528\u6236\u7aef\uff08client\uff09\u89c0\u5bdf\u4e32\u6d41\u8acb\u6c42\u7684\u6642\u9593\u8ef8\uff0c\u4ee5\u53ca\u56db\u9805\u6838\u5fc3\u6307\u6a19\u5404\u81ea\u5c0d\u61c9\u7684\u6642\u9593\u5340\u6bb5<\/p>\n\n\n\n

      \n
    • TTFT\uff08time to first token\uff09\uff1d t\u2081 \u2212 t\u2080\u3002<\/strong>\u9019\u6bb5\u6642\u9593\u5305\u542b\u7db2\u8def\u5f80\u8fd4\u3001\u4f3a\u670d\u5668\u6392\u968a\uff0c\u4ee5\u53ca prefill \u8655\u7406\u5b8c\u6574\u6bb5 prompt \u7684\u6642\u9593\u3002\u5b83\u53cd\u6620\u4f7f\u7528\u8005\u662f\u5426\u611f\u89ba\u5230\u300c\u7cfb\u7d71\u6709\u56de\u61c9\u300d\uff0c\u5c0d prompt \u9577\u5ea6\u76f8\u7576\u654f\u611f\uff0c\u4e5f\u5e38\u5728\u4e26\u767c\u91cf\u5347\u9ad8\u6642\u7387\u5148\u60e1\u5316\u3002<\/li>\n\n\n\n
    • ITL\uff08inter-token latency\uff09\uff1d (request latency \u2212 TTFT) \u00f7 (OSL \u2212 1)\u3002<\/strong>\u9019\u662f\u6536\u5230\u7b2c\u4e00\u500b token \u4e4b\u5f8c\uff0c\u5e73\u5747\u6bcf\u9694\u591a\u4e45\u6536\u5230\u4e0b\u4e00\u500b token\uff0c\u80fd\u53cd\u6620 decode \u8f38\u51fa\u7684\u6d41\u66a2\u5ea6\u3002TPOT \u662f\u696d\u754c\u5e38\u898b\u7684\u76f8\u95dc\u540d\u7a31\u3002\u5206\u6bcd\u4f7f\u7528 OSL \u2212 1\uff0c\u662f\u56e0\u70ba\u7b2c\u4e00\u500b token \u7684\u7b49\u5f85\u6642\u9593\u5df2\u8a08\u5165 TTFT\u3002<\/li>\n\n\n\n
    • Request latency \uff1d t\u2099 \u2212 t\u2080\u3002<\/strong>\u9019\u662f\u6574\u6bb5\u56de\u7b54\u7684\u7e3d\u8017\u6642\u3002\u5373\u4f7f\u5169\u500b\u8acb\u6c42\u7684\u7e3d\u8017\u6642\u76f8\u540c\uff0c\u4f7f\u7528\u8005\u9ad4\u9a57\u4e5f\u53ef\u80fd\u5dee\u5f88\u591a\uff1a\u4e00\u500b\u53ef\u80fd\u300c\u7b49\u5f88\u4e45\u3001\u4f46\u5f8c\u9762\u8f38\u51fa\u5f88\u5feb\u300d\uff0c\u53e6\u4e00\u500b\u5247\u662f\u300c\u5f88\u5feb\u958b\u59cb\u3001\u4f46\u6162\u6162\u8f38\u51fa\u300d\u3002\u56e0\u6b64\uff0c\u7e3d\u8017\u6642\u5fc5\u9808\u642d\u914d TTFT \u548c ITL \u4e00\u8d77\u770b\u3002<\/li>\n\n\n\n
    • \u6bcf\u4f4d\u4f7f\u7528\u8005\u7684\u8f38\u51fa\u901f\u5ea6 \uff1d 1 \u00f7 ITL\u3002<\/strong>\u9019\u8868\u793a\u55ae\u4e00\u4f7f\u7528\u8005\u6bcf\u79d2\u5927\u7d04\u80fd\u6536\u5230\u591a\u5c11 token\u3002\u82e5\u901f\u5ea6\u964d\u5230\u6bcf\u79d2\u500b\u4f4d\u6578 token\uff0c\u4f7f\u7528\u8005\u901a\u5e38\u6703\u660e\u986f\u611f\u89ba\u8f38\u51fa\u504f\u6162\u3002<\/li>\n<\/ul>\n\n\n\n

      2.2 \u541e\u5410\u91cf\uff1a\u7cfb\u7d71\u7684\u6574\u9ad4\u7522\u80fd<\/strong><\/h2>\n\n\n\n
        \n
      • Benchmark duration\uff1a<\/strong>\u5f9e\u7b2c\u4e00\u500b\u8acb\u6c42\u9001\u51fa\uff0c\u5230\u6700\u5f8c\u4e00\u500b\u56de\u61c9\u6536\u5230\u70ba\u6b62\u7684\u6642\u9593\u3002\u8a08\u7b97\u541e\u5410\u91cf\u6642\uff0c\u6703\u4ee5\u9019\u6bb5\u6642\u9593\u4f5c\u70ba\u5206\u6bcd\u3002<\/li>\n\n\n\n
      • Output token throughput \uff1d \u6240\u6709\u8f38\u51fa token \u7684\u7e3d\u6578 \u00f7 duration\u3002<\/strong>\u4ee3\u8868\u6574\u500b\u7cfb\u7d71\u6bcf\u79d2\u7e3d\u5171\u7522\u751f\u591a\u5c11\u8f38\u51fa token\uff08Tokens Per Second\uff0cTPS\uff09\u3002<\/li>\n\n\n\n
      • Request throughput \uff1d \u5b8c\u6210\u7684\u8acb\u6c42\u6578 \u00f7 duration\u3002<\/strong>\u4ee3\u8868\u7cfb\u7d71\u6bcf\u79d2\u5b8c\u6210\u591a\u5c11\u500b\u8acb\u6c42\uff08Requests Per Second\uff0cRPS\uff09\u3002\u5b83\u8207 token \u541e\u5410\u91cf\u4e4b\u9593\u7684\u63db\u7b97\uff0c\u53d6\u6c7a\u65bc\u5e73\u5747\u6bcf\u500b\u8acb\u6c42\u7684\u8f38\u51fa\u9577\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n

        \u6700\u5bb9\u6613\u6df7\u6dc6\u7684\u4f8b\u5b50\uff0c\u662f\u300c\u6bcf\u4f4d\u4f7f\u7528\u8005\u6bcf\u79d2\u6536\u5230 20 \u500b token\u300d\u548c\u300c\u6574\u500b\u7cfb\u7d71\u6bcf\u79d2\u7522\u751f 2,000 \u500b token\u300d\u3002\u9019\u5169\u500b\u6578\u5b57\u4ee3\u8868\u4e0d\u540c\u7684\u4e8b\u60c5\uff0c\u8b8a\u5316\u8da8\u52e2\u4e5f\u53ef\u80fd\u76f8\u53cd\uff1a\u4e26\u767c\u91cf\u63d0\u9ad8\u6642\uff0cbatching \u53ef\u80fd\u8b93\u7cfb\u7d71\u7684\u7e3d\u7522\u51fa\u589e\u52a0\uff0c\u4f46\u6bcf\u500b decode \u6b65\u9a5f\u4e5f\u53ef\u80fd\u8b8a\u6162\uff0c\u5c0e\u81f4\u55ae\u4e00\u4f7f\u7528\u8005\u7684\u8f38\u51fa\u901f\u5ea6\u4e0b\u964d\u3002\u770b\u5230 tokens\/s \u9019\u500b\u55ae\u4f4d\u6642\uff0c\u8a18\u5f97\u5148\u78ba\u8a8d\u5b83\u6307\u7684\u662f\u300c\u6bcf\u4f4d\u4f7f\u7528\u8005\u300d\u9084\u662f\u300c\u6574\u500b\u7cfb\u7d71\u300d\u3002<\/p>\n\n\n\n

        2.3 \u70ba\u4ec0\u9ebc\u4e0d\u80fd\u53ea\u770b\u5e73\u5747\u503c<\/strong><\/h2>\n\n\n\n

        \u5ef6\u9072\u5206\u5e03\u901a\u5e38\u5448\u73fe\u53f3\u504f\u3001\u5e36\u6709\u9577\u5c3e\u7684\u6a23\u5b50\uff1a\u5927\u591a\u6578\u8acb\u6c42\u5f88\u5feb\u5b8c\u6210\uff0c\u4f46\u5c11\u6578\u8acb\u6c42\u53ef\u80fd\u56e0\u6392\u968a\u6216\u8a18\u61b6\u9ad4\u58d3\u529b\u800c\u6162\u4e0a\u597d\u5e7e\u500d\u3002\u5e73\u5747\u503c\u5bb9\u6613\u88ab\u9019\u4e9b\u6162\u8acb\u6c42\u62c9\u9ad8\uff0c\u537b\u7121\u6cd5\u6e05\u695a\u5448\u73fe\u5c3e\u7aef\u5ef6\u9072\u7684\u7a0b\u5ea6\u3002AIPerf \u65b9\u6cd5\u8ad6\u6587\u4ef6\u4e2d\u7684\u4f8b\u5b50\u986f\u793a\uff0cp99 \u6bd4\u5e73\u5747\u503c\u9ad8\u51fa\u4e03\u6210\u4ee5\u4e0a\u3002\u4e00\u822c\u4f86\u8aaa\uff0cp50 \u53ef\u7528\u4f86\u89c0\u5bdf\u5178\u578b\u8acb\u6c42\u7684\u5ef6\u9072\uff1bp95 \u4ee3\u8868 95% \u7684\u8acb\u6c42\u5ef6\u9072\u4e0d\u8d85\u904e\u9019\u500b\u503c\uff1bp99 \u5247\u4ee3\u8868 99% \u7684\u8acb\u6c42\u5ef6\u9072\u4e0d\u8d85\u904e\u9019\u500b\u503c\uff0c\u53ea\u6709\u6700\u6162\u7684 1% \u843d\u5728\u5176\u5f8c\u3002SLO \u901a\u5e38\u6703\u4ee5 p95 \u6216 p99 \u4f5c\u70ba\u9580\u6abb\u3002\u56e0\u6b64\uff0c\u597d\u7684\u6e2c\u8a66\u5de5\u5177\u6703\u4fdd\u7559\u6bcf\u500b\u8acb\u6c42\u7684\u539f\u59cb\u7d00\u9304\uff0c\u65b9\u4fbf\u4e8b\u5f8c\u91cd\u65b0\u8a08\u7b97\u4e0d\u540c\u7684\u767e\u5206\u4f4d\u6578\uff08percentile\uff09\u3002<\/p>\n\n\n\n

        2.4 Goodput\uff1a\u540c\u6642\u8861\u91cf\u901f\u5ea6\u8207\u670d\u52d9\u54c1\u8cea<\/strong><\/h2>\n\n\n\n

        \u5ef6\u9072\u8207\u541e\u5410\u91cf\u5f80\u5f80\u4e92\u76f8\u727d\u5236\u3002\u8981\u56de\u7b54\u300c\u7cfb\u7d71\u80fd\u627f\u53d7\u591a\u5c11\u4f7f\u7528\u8005\u300d\uff0c\u5c31\u5f97\u628a\u5169\u8005\u653e\u5728\u540c\u4e00\u500b\u5224\u6e96\u4e0b\u770b\uff1agoodput \uff1d \u540c\u6642\u7b26\u5408\u6240\u6709 SLO \u7684\u8acb\u6c42\u6578 \u00f7 duration<\/strong>\u3002\u4f8b\u5982\uff0c\u53ef\u4ee5\u8a2d\u5b9a TTFT < 500 ms \u4e14 ITL < 50 ms\uff1b\u53ea\u6709\u6240\u6709\u689d\u4ef6\u90fd\u9054\u6a19\u7684\u8acb\u6c42\uff0c\u624d\u6703\u8a08\u5165 goodput \u7684\u5206\u5b50\u3002\u56e0\u6b64\uff0c goodput \u4e0d\u6703\u9ad8\u65bc request throughput\u3002\u5bb9\u91cf\u898f\u5283\u82e5\u9078\u932f\u6307\u6a19\uff0c\u7d50\u679c\u53ef\u80fd\u5dee\u5f88\u591a\uff1aAIPerf \u6587\u4ef6\u4e2d\u7684\u4f8b\u5b50\u986f\u793a\uff0c\u4ee5 throughput \u4f30\u7b97\u9700\u8981 38 \u53f0\u4f3a\u670d\u5668\uff0c\u4ee5 goodput \u4f30\u7b97\u5247\u9700\u8981 135 \u53f0\u3002\u5f8c\u8005\u624d\u771f\u6b63\u4ee3\u8868\u300c\u5de5\u4f5c\u505a\u5f97\u5b8c\uff0c\u800c\u4e14\u4f7f\u7528\u8005\u9ad4\u9a57\u4e5f\u7b26\u5408\u8981\u6c42\u300d\u3002<\/p>\n\n\n\n

        \u6307\u6a19<\/strong><\/th>\u516c\u5f0f<\/strong><\/th>\u4ee3\u8868\u7684\u610f\u7fa9\uff0f\u95dc\u6ce8\u9762\u5411<\/strong><\/th><\/tr><\/thead>
        TTFT<\/td>t\u2081 \u2212 t\u2080<\/td>\u4f7f\u7528\u8005\u662f\u5426\u611f\u89ba\u5230\u7cfb\u7d71\u6709\u56de\u61c9<\/td><\/tr>
        ITL\uff08TPOT\uff09<\/td>(request latency \u2212 TTFT) \u00f7 (OSL \u2212 1)<\/td>\u8f38\u51fa\u6587\u5b57\u662f\u5426\u6d41\u66a2<\/td><\/tr>
        Request latency<\/td>t\u2099 \u2212 t\u2080<\/td>\u6279\u6b21\u5de5\u4f5c\u3001Agent \u6d41\u7a0b<\/td><\/tr>
        Output token throughput<\/td>\u03a3 OSL \u00f7 duration<\/td>\u71df\u904b\u6210\u672c\u3001GPU \u5229\u7528\u7387<\/td><\/tr>
        Goodput<\/td>\u7b26\u5408 SLO \u7684\u8acb\u6c42\u6578 \u00f7 duration<\/td>\u5bb9\u91cf\u898f\u5283\u8207\u670d\u52d9\u54c1\u8cea<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

        2.5 \u70ba\u4ec0\u9ebc\u95dc\u9589\u4e32\u6d41\u5f8c\uff0c\u5c31\u91cf\u4e0d\u5230 TTFT \u8207 ITL<\/strong><\/h2>\n\n\n\n

        \u5982\u679c\u8acb\u6c42\u6c92\u6709\u555f\u7528\u4e32\u6d41\uff0c\u4f3a\u670d\u5668\u6703\u7b49\u6574\u6bb5\u56de\u7b54\u5b8c\u6210\u5f8c\u624d\u4e00\u6b21\u56de\u50b3\u3002\u6b64\u6642\uff0cclient \u7aef\u7684\u6642\u9593\u8ef8\u53ea\u770b\u5f97\u5230 t\u2080 \u8207 t\u2099\uff1a\u7b2c\u4e00\u500b token \u548c\u6700\u5f8c\u4e00\u500b token \u6703\u96a8\u6574\u6bb5\u56de\u7b54\u4e00\u8d77\u5230\u9054\u3002\u5982\u6b64\u4e00\u4f86\uff0cTTFT \u5c31\u6703\u7b49\u540c\u65bc request latency\uff0cITL \u7684\u8a08\u7b97\u4e5f\u5931\u53bb\u610f\u7fa9\u3002\u56e0\u6b64\uff0cTTFT\u3001ITL \u8207\u6bcf\u4f4d\u4f7f\u7528\u8005\u7684\u8f38\u51fa\u901f\u5ea6\u90fd\u662f\u53ea\u9069\u7528\u65bc\u4e32\u6d41\u8f38\u51fa\u7684\u6307\u6a19<\/strong>\uff08AIPerf \u5c07\u5b83\u5011\u6a19\u793a\u70ba streaming-only\uff09\u3002\u53cd\u904e\u4f86\u8aaa\uff0c\u5982\u679c\u7522\u54c1\u672c\u4f86\u5c31\u4e0d\u4f7f\u7528\u4e32\u6d41\uff0c\u4f8b\u5982\u5f8c\u7aef\u6279\u6b21\u6458\u8981\uff0c\u5c31\u4e0d\u4e00\u5b9a\u9700\u8981\u95dc\u6ce8\u9019\u4e9b\u6307\u6a19\u3002\u9078\u64c7\u6307\u6a19\u6642\uff0c\u61c9\u4ee5\u7522\u54c1\u5be6\u969b\u7684\u4f7f\u7528\u65b9\u5f0f\u70ba\u6e96\u3002<\/p>\n\n\n\n

        \u6709\u5169\u500b\u8a2d\u5b9a\u7279\u5225\u5bb9\u6613\u8b93\u91cf\u6e2c\u7d50\u679c\u5931\u771f\u3002<\/strong>\u7b2c\u4e00\uff0c\u8f38\u51fa\u9577\u5ea6\u6c92\u6709\u63a7\u5236\u597d\uff1a\u6a21\u578b\u53ef\u80fd\u63d0\u65e9\u9047\u5230\u7d50\u675f\u7b26\u865f\u800c\u505c\u6b62\u751f\u6210\uff0c\u5c0e\u81f4\u5be6\u969b OSL \u77ed\u65bc\u9810\u671f\uff0c\u9032\u800c\u5f71\u97ff\u541e\u5410\u91cf\u8207 ITL\u3002\u8f03\u56b4\u8b39\u7684\u6e2c\u8a66\u65b9\u5f0f\uff0c\u662f\u900f\u904e ignore_eos \u6216 min_tokens \u7b49\u53c3\u6578\uff0c\u8b93\u6a21\u578b\u4f9d\u8a2d\u5b9a\u7522\u751f\u8db3\u5920\u7684 token\u3002\u53e6\u9808\u7559\u610f\uff0c\u6709\u4e9b\u670d\u52d9\u4ecd\u4f7f\u7528 max_tokens \u9019\u500b\u6b04\u4f4d\u540d\u7a31\uff0c\u4e5f\u53ef\u80fd\u975c\u9ed8\u5ffd\u7565\u4e0d\u652f\u63f4\u7684\u53c3\u6578\u3002\u7b2c\u4e8c\uff0ctokenizer \u8a2d\u5b9a\u932f\u8aa4\uff1aclient \u7aef\u6703\u900f\u904e tokenizer \u8a08\u7b97 ISL \u8207 OSL\uff1b\u82e5\u4f7f\u7528\u7684 tokenizer \u8207\u6a21\u578b\u4e0d\u4e00\u81f4\uff0c\u6240\u6709\u4ee5 token \u70ba\u55ae\u4f4d\u7684\u6307\u6a19\u90fd\u53ef\u80fd\u7522\u751f\u504f\u5dee\u3002<\/p>\n\n\n\n

        \u4e09\u3001\u8ca0\u8f09\u6a21\u578b\uff1a\u65bd\u52a0\u8ca0\u8f09\u7684\u65b9\u5f0f\uff0c\u6c7a\u5b9a\u4f60\u91cf\u5230\u4ec0\u9ebc<\/strong><\/h1>\n\n\n\n

        \u8ca0\u8f09\u6a21\u578b\u4e0d\u53ea\u662f\u8a2d\u5b9a\u4e00\u500b\u300c\u5927\u5c0f\u300d\uff0c\u66f4\u91cd\u8981\u7684\u662f\u6c7a\u5b9a\u8acb\u6c42\u4f55\u6642\u9001\u51fa\u3001\u4ee5\u53ca\u4f9d\u7167\u4ec0\u9ebc\u898f\u5247\u9001\u51fa\u7684\u6a5f\u5236\u3002<\/p>\n\n\n\n

        \"\u9589\u74b0\u3001\u958b\u74b0\u8ca0\u8f09\u6a21\u578b\u7684\u5dee\u7570\"<\/figure>\n\n\n\n

        \u5716 2\u3000\u9589\u74b0\uff08\u56fa\u5b9a\u4e26\u767c\u6578\uff09\u8207\u958b\u74b0\uff08\u56fa\u5b9a\u5230\u9054\u7387\uff09\u5169\u7a2e\u8ca0\u8f09\u6a21\u578b\u7684\u5dee\u7570<\/p>\n\n\n\n

        Concurrency\uff08\u56fa\u5b9a\u4e26\u767c\u6578\uff0c\u9589\u74b0\uff09\uff1a<\/strong>\u6a21\u64ec N \u4f4d\u865b\u64ec\u4f7f\u7528\u8005\uff0c\u6bcf\u4f4d\u4f7f\u7528\u8005\u9001\u51fa\u4e00\u500b\u8acb\u6c42\uff0c\u7b49\u6536\u5230\u5b8c\u6574\u56de\u61c9\u5f8c\uff0c\u518d\u9001\u51fa\u4e0b\u4e00\u500b\u3002\u4efb\u4f55\u6642\u5019\uff0c\u5c1a\u672a\u5b8c\u6210\u7684\u8acb\u6c42\u6578\u90fd\u4e0d\u6703\u8d85\u904e N\u3002\u9019\u7a2e\u65b9\u5f0f\u6709\u56de\u994b\u6548\u679c\uff1a\u7cfb\u7d71\u6108\u6162\uff0c\u4e0b\u4e00\u500b\u8acb\u6c42\u5c31\u6108\u665a\u9001\u51fa\uff0c\u56e0\u6b64\u6574\u9ad4\u9001\u51fa\u901f\u7387\u6703\u81ea\u52d5\u4e0b\u964d\u3002\u5b83\u9069\u5408\u91cf\u6e2c\u300cN \u4f4d\u4f7f\u7528\u8005\u6301\u7e8c\u4f7f\u7528\u6642\u300d\u7cfb\u7d71\u80fd\u7dad\u6301\u7684\u7a69\u614b\u5bb9\u91cf\uff0c\u7d50\u679c\u8f03\u7a69\u5b9a\u3001\u5bb9\u6613\u91cd\u73fe\uff0c\u4f46\u5c0d\u7a81\u767c\u6d41\u91cf\u4f86\u8aaa\uff0c\u8ca0\u8f09\u76f8\u5c0d\u6eab\u548c\u3002<\/p>\n\n\n\n

        Request rate\uff08\u56fa\u5b9a\u5230\u9054\u7387\uff0c\u958b\u74b0\uff09\uff1a<\/strong>\u6bcf\u79d2\u4f9d\u7167\u6392\u7a0b\u9001\u51fa \u03bb \u500b\u8acb\u6c42\uff0c\u4e0d\u7ba1\u524d\u4e00\u500b\u8acb\u6c42\u662f\u5426\u5df2\u5b8c\u6210\u3002\u82e5\u7cfb\u7d71\u8655\u7406\u4e0d\u53ca\uff0c\u5c1a\u672a\u5b8c\u6210\u7684\u8acb\u6c42\u5c31\u6703\u4e0d\u65b7\u7d2f\u7a4d\uff0c\u5ef6\u9072\u4e5f\u53ef\u80fd\u5feb\u901f\u4e0a\u5347\u3002\u9019\u7a2e\u65b9\u5f0f\u66f4\u63a5\u8fd1\u56fa\u5b9a\u901f\u7387\u5230\u9054\u7684\u5be6\u969b\u6d41\u91cf\uff0c\u9069\u5408\u9a57\u8b49\u300c\u9810\u671f\u5c16\u5cf0\u4e0b\uff0c\u5ef6\u9072\u662f\u5426\u4ecd\u7b26\u5408 SLO\u300d\u3002\u4e0d\u904e\uff0c\u5b83\u5c0d \u03bb \u7684\u8a2d\u5b9a\u975e\u5e38\u654f\u611f\uff1a\u7565\u4f4e\u65bc\u7cfb\u7d71\u53ef\u627f\u53d7\u7684\u5bb9\u91cf\u6642\u53ef\u80fd\u4e00\u5207\u6b63\u5e38\uff0c\u7565\u9ad8\u6642\u6392\u968a\u5c31\u53ef\u80fd\u6301\u7e8c\u7d2f\u7a4d\u3002<\/p>\n\n\n\n

        \u9019\u5169\u7a2e\u65b9\u6cd5\u56de\u7b54\u7684\u662f\u4e0d\u540c\u554f\u984c\uff1a\u300c\u6700\u591a\u80fd\u540c\u6642\u670d\u52d9\u591a\u5c11\u4eba\uff1f\u300d\u9069\u5408\u7528\u9589\u74b0\u6e2c\u8a66\uff1b\u300c\u6bcf\u79d2\u6536\u5230 50 \u500b\u8acb\u6c42\u6642\uff0cp99 \u5ef6\u9072\u662f\u591a\u5c11\uff1f\u300d\u5247\u9069\u5408\u7528\u958b\u74b0\u6e2c\u8a66\u3002\u5e38\u898b\u7684\u8aa4\u7528\u662f\uff1a\u9589\u74b0\u6e2c\u8a66\u986f\u793a\u300c100 \u4e26\u767c\u6642\uff0cTTFT \u7684 p99 \u53ea\u6709 800 ms\u300d\uff0c\u4e0a\u7dda\u5f8c\u6bcf\u79d2\u6e67\u5165 100 \u500b\u8acb\u6c42\uff0c\u7cfb\u7d71\u537b\u53ef\u80fd\u6490\u4e0d\u4f4f\u3002\u539f\u56e0\u662f\u7576\u7cfb\u7d71\u8b8a\u6162\u6642\uff0c\u9589\u74b0\u6e2c\u8a66\u4e2d\u7684 100 \u4e26\u767c\uff0c\u6bcf\u79d2\u5be6\u969b\u9001\u51fa\u7684\u8acb\u6c42\u53ef\u80fd\u53ea\u6709 30 \u500b\u3002<\/p>\n\n\n\n

        \u8acb\u6c42\u5230\u9054\u6a21\u5f0f\uff1a<\/strong>\u958b\u74b0\u6e2c\u8a66\u9084\u8981\u6c7a\u5b9a\u8acb\u6c42\u9593\u9694\u591a\u4e45\u3002Constant \u6a21\u5f0f\u6703\u6bcf\u9694 1\/\u03bb \u79d2\u9001\u51fa\u4e00\u500b\u8acb\u6c42\uff0c\u9593\u9694\u5b8c\u5168\u56fa\u5b9a\uff1bPoisson \u6a21\u5f0f\u7684\u8acb\u6c42\u9593\u9694\u5247\u670d\u5f9e\u6307\u6578\u5206\u5e03\uff0c\u5e73\u5747\u9593\u9694\u540c\u6a23\u662f 1\/\u03bb\uff0c\u4f46\u6703\u81ea\u7136\u51fa\u73fe\u8acb\u6c42\u96c6\u4e2d\u62b5\u9054\u8207\u66ab\u6642\u6c92\u6709\u8acb\u6c42\u7684\u7a7a\u6a94\u3002\u9019\u662f\u6a21\u64ec\u7368\u7acb\u4f7f\u7528\u8005\u96a8\u6a5f\u9001\u51fa\u8acb\u6c42\u7684\u6578\u5b78\u6a21\u578b\uff0c\u4e5f\u662f AIPerf \u7684\u9810\u8a2d\u6a21\u5f0f\u3002Constant \u6a21\u5f0f\u4e0b\uff0c\u8acb\u6c42\u9593\u9694\u56fa\u5b9a\uff0c\u4f47\u5217\u8f03\u4e0d\u5bb9\u6613\u7a81\u7136\u5806\u7a4d\uff1bPoisson \u6a21\u5f0f\u5247\u53ef\u80fd\u5728\u540c\u4e00\u79d2\u6e67\u5165\u4e94\u500b\u8acb\u6c42\uff0c\u9032\u800c\u5f62\u6210\u5ef6\u9072\u9577\u5c3e\u3002\u56e0\u6b64\uff0c\u53ea\u7528 constant \u6e2c\u5f97\u7684 p99\uff0c\u901a\u5e38\u6703\u6bd4\u8cbc\u8fd1\u771f\u5be6\u6d41\u91cf\u7684\u6e2c\u8a66\u7d50\u679c\u6a02\u89c0\u3002<\/p>\n\n\n\n

        3.1 \u6383\u63cf\u8ca0\u8f09\u8207 knee \u8f49\u6298\u9ede<\/strong><\/h2>\n\n\n\n

        \u55ae\u4e00\u8ca0\u8f09\u9ede\u53ea\u80fd\u544a\u8a34\u4f60\u300c\u7cfb\u7d71\u5728\u9019\u500b\u8ca0\u8f09\u4e0b\u7684\u8868\u73fe\u300d\u3002\u82e5\u8981\u56de\u7b54\u300c\u7cfb\u7d71\u6700\u591a\u80fd\u627f\u53d7\u591a\u5c11\u58d3\u529b\u300d\uff0c\u5c31\u8981\u9010\u6b65\u63d0\u9ad8\u4e26\u767c\u6578\uff0c\u4f8b\u5982\u4f9d\u5c0d\u6578\u5c3a\u5ea6\u6e2c\u8a66 1\u30012\u30014\u30018\u2026\u2026\uff0c\u4e26\u5728\u6bcf\u500b\u8ca0\u8f09\u9ede\u57f7\u884c\u5b8c\u6574\u7684 benchmark\uff0c\u518d\u7e6a\u88fd\u541e\u5410\u91cf\u8207\u5ef6\u9072\u66f2\u7dda\u3002<\/p>\n\n\n\n

        \"Token\u541e\u5410\u91cf\u8207TTFT\u5ef6\u9072\u66f2\u7dda\"<\/figure>\n\n\n\n

        \u5716 3\u3000\u4e26\u767c\u6383\u63cf\u7684\u541e\u5410\u91cf\u8207\u5ef6\u9072\u66f2\u7dda\uff1a\u541e\u5410\u91cf\u5148\u8da8\u65bc\u98fd\u548c\uff0c\u5ef6\u9072\u5247\u53ef\u80fd\u6025\u907d\u4e0a\u5347(\u5716\u5de6)\uff1b\u8207 SLO \u9580\u6abb\u7684\u4ea4\u9ede\u53ef\u7528\u4f86\u5224\u5b9a\u5bb9\u91cf(\u5716\u53f3)<\/p>\n\n\n\n

        \u96a8\u8457\u4e26\u767c\u6578\u589e\u52a0\uff0c\u541e\u5410\u91cf\u901a\u5e38\u6703\u6301\u7e8c\u4e0a\u5347\uff0c\u4f46\u589e\u5e45\u9010\u6f38\u7e2e\u5c0f\uff0c\u5230\u67d0\u500b\u9ede\u5f8c\u53ef\u80fd\u6301\u5e73\uff0c\u751a\u81f3\u7565\u5fae\u4e0b\u964d\u3002\u5ef6\u9072\u5247\u53ef\u80fd\u5728\u4e00\u958b\u59cb\u8b8a\u5316\u4e0d\u5927\uff0c\u8d85\u904e\u67d0\u500b\u9ede\u5f8c\u958b\u59cb\u6025\u907d\u4e0a\u5347\u3002\u9019\u6bb5\u8f49\u6298\u5340\u57df\u7a31\u70ba knee\uff08\u8f49\u6298\u9ede\uff09\u3002\u8f03\u7406\u60f3\u7684\u5de5\u4f5c\u9ede\u901a\u5e38\u843d\u5728 knee \u4e4b\u524d\u4e0d\u9060\u8655\u3002\u82e5\u5728\u5716\u4e0a\u52a0\u4e0a SLO \u7684\u5ef6\u9072\u9580\u6abb\u7dda\uff0c\u5ef6\u9072\u66f2\u7dda\u8207\u9580\u6abb\u7dda\u7684\u4ea4\u9ede\uff0c\u5c31\u80fd\u7528\u4f86\u5224\u5b9a\u300c\u7b26\u5408 SLO \u7684\u6700\u9ad8\u4e26\u767c\u6578\u300d\uff1b\u518d\u642d\u914d\u8a72\u8ca0\u8f09\u4e0b\u7684 goodput\uff0c\u5c31\u80fd\u66f4\u5b8c\u6574\u5730\u56de\u7b54\u6709\u95dc\u904b\u884c\u6548\u80fd\u7684\u554f\u984c\u3002<\/p>\n\n\n\n

        3.2 \u6696\u6a5f\u8207\u907f\u514d\u6e2c\u8a66\u4e92\u76f8\u5e72\u64fe<\/strong><\/h2>\n\n\n\n

        \u6696\u6a5f\uff1a<\/strong>\u525b\u555f\u52d5\u7684\u63a8\u8ad6\u670d\u52d9\uff0c\u524d\u5e7e\u5341\u79d2\u7684\u8868\u73fe\u53ef\u80fd\u8207\u7a69\u5b9a\u904b\u4f5c\u6642\u5dee\u5f88\u591a\uff0c\u539f\u56e0\u5305\u62ec\u8f09\u5165\u6a21\u578b\u6b0a\u91cd\u3001\u9996\u6b21\u7de8\u8b6f CUDA kernel\u3001\u914d\u7f6e KV cache \u8a18\u61b6\u9ad4\u6c60\uff0c\u4ee5\u53ca\u5efa\u7acb\u9023\u7dda\u3002AIPerf \u65b9\u6cd5\u8ad6\u6587\u4ef6\u4e2d\u7684\u4f8b\u5b50\u986f\u793a\uff0c\u524d 10 \u79d2\u7684 TTFT \u6bd4\u7a69\u614b\u9ad8\u7d04 41%\u3002\u82e5\u628a\u9019\u6bb5\u6642\u9593\u4e5f\u7b97\u9032\u7d71\u8a08\uff0cp99 \u5c31\u53ef\u80fd\u6bd4\u7a69\u614b\u5be6\u969b\u8868\u73fe\u66f4\u5dee\uff0c\u800c\u4e14\u5f71\u97ff\u7a0b\u5ea6\u9084\u6703\u96a8 benchmark \u57f7\u884c\u6642\u9593\u800c\u6539\u8b8a\u3002\u8f03\u597d\u7684\u505a\u6cd5\u662f\u5148\u57f7\u884c\u4e00\u6bb5 warmup\uff08\u6696\u6a5f\u6e2c\u8a66\uff09\uff0c\u5c07\u9019\u6bb5\u8acb\u6c42\u6a19\u8a18\u70ba\u6696\u6a5f\u8cc7\u6599\u4e26\u6392\u9664\uff0c\u518d\u958b\u59cb\u6b63\u5f0f\u91cf\u6e2c\u3002<\/p>\n\n\n\n

        \u907f\u514d\u540c\u6642\u57f7\u884c\u591a\u7d44\u6e2c\u8a66\uff1a<\/strong>\u5047\u8a2d\u5169\u4f4d\u5de5\u7a0b\u5e2b\u5404\u81ea\u5c0d\u540c\u4e00\u500b endpoint \u57f7\u884c benchmark\uff0c\u4e26\u5404\u81ea\u8a2d\u5b9a 32 \u4e26\u767c\uff0c\u4f3a\u670d\u5668\u5be6\u969b\u627f\u53d7\u7684\u53ef\u80fd\u662f 64 \u500b\u5c1a\u672a\u5b8c\u6210\u7684\u8acb\u6c42\u3002\u9019\u4e9b\u8acb\u6c42\u5171\u7528\u540c\u4e00\u5f35 GPU\u3001\u540c\u4e00\u5957 batch \u6392\u7a0b\u8207\u540c\u4e00\u6c60 KV cache\uff0c\u56e0\u6b64\u5169\u4efd\u5831\u544a\u4e2d\u7684\u5ef6\u9072\u90fd\u53ef\u80fd\u6bd4\u55ae\u7368\u6e2c\u8a66 32 \u4e26\u767c\u6642\u66f4\u5dee\uff0c\u537b\u4e0d\u5bb9\u6613\u5f9e\u500b\u5225\u5831\u544a\u770b\u51fa\u539f\u56e0\u3002\u56e0\u6b64\u540c\u4e00\u500b endpoint \u7684 benchmark \u61c9\u4e00\u6b21\u53ea\u57f7\u884c\u4e00\u7d44\uff1b\u958b\u59cb\u524d\u5c31\u8981\u78ba\u8a8d\u6c92\u6709\u5176\u4ed6\u6e2c\u8a66\u540c\u6642\u57f7\u884c\uff0c\u800c\u4e0d\u662f\u53ea\u9760\u53e3\u982d\u63d0\u9192\u3002\u9019\u500b\u539f\u5247\u4e5f\u9069\u7528\u65bc\u7522\u751f\u8ca0\u8f09\u7684 client \u6a5f\u5668\uff1a\u5982\u679c CPU \u8cc7\u6e90\u4e0d\u8db3\u3001\u53ef\u7528\u9023\u63a5\u57e0\u8017\u76e1\uff0c\u6216\u4e8b\u4ef6\u8ff4\u5708\u88ab JSON \u89e3\u6790\u5de5\u4f5c\u5361\u4f4f\uff0c\u91cf\u5230\u7684\u5ef6\u9072\u4e5f\u6703\u6df7\u5165 client \u7aef\u7684\u74f6\u9838\uff0c\u7121\u6cd5\u55ae\u7d14\u4ee3\u8868\u4f3a\u670d\u5668\u6548\u80fd\u3002<\/p>\n\n\n\n

        \u53e6\u5916\u9084\u6709\u4e09\u500b\u56e0\u7d20\u6703\u5f71\u97ff\u7d50\u679c\u7684\u53ef\u6bd4\u8f03\u6027\u3002<\/strong>\u7b2c\u4e00\u662f\u6a23\u672c\u6578\uff1a\u6bcf\u500b\u8ca0\u8f09\u9ede\u6700\u597d\u81f3\u5c11\u6536\u96c6\u6578\u767e\u500b\u8acb\u6c42\uff0cp99 \u624d\u8f03\u6709\u53c3\u8003\u50f9\u503c\uff1b\u53ea\u6709\u4e00\u767e\u500b\u8acb\u6c42\u6642\uff0cp99 \u5e7e\u4e4e\u5c31\u662f\u6700\u6162\u7684\u90a3\u4e00\u7b46\u3002\u7b2c\u4e8c\u662f\u5de5\u4f5c\u8ca0\u8f09\u7279\u6027\uff1a\u5408\u6210\u8cc7\u6599\uff08\u56fa\u5b9a ISL\/OSL\uff09\u80fd\u63d0\u4f9b\u53ef\u63a7\u5236\u3001\u53ef\u91cd\u73fe\u7684\u5bb9\u91cf\u6578\u5b57\uff1b\u771f\u5be6\u5c0d\u8a71\u7d00\u9304\u7684\u9577\u5ea6\u5206\u5e03\u5247\u5e38\u5e36\u6709\u9577\u5c3e\uff0c\u53ef\u80fd\u66b4\u9732\u5408\u6210\u8cc7\u6599\u770b\u4e0d\u5230\u7684\u554f\u984c\u3002\u5169\u7a2e\u90fd\u503c\u5f97\u6e2c\u8a66\uff0c\u4f46\u4e0d\u9069\u5408\u76f4\u63a5\u6bd4\u8f03\u3002\u7b2c\u4e09\u662f\u53ef\u91cd\u73fe\u6027\uff1a\u56fa\u5b9a\u96a8\u6a5f\u7a2e\u5b50\u53ef\u4ee5\u8b93 prompt \u5167\u5bb9\u8207 Poisson \u5230\u9054\u9593\u9694\u91cd\u73fe\uff0c\u4f46\u5ef6\u9072\u672c\u8eab\u4ecd\u6709\u81ea\u7136\u6ce2\u52d5\u3002\u6bd4\u8f03\u5169\u7a2e\u914d\u7f6e\u6642\uff0c\u5efa\u8b70\u5404\u8dd1\u6578\u6b21\u4e26\u6bd4\u8f03\u4e2d\u4f4d\u6578\uff0c\u540c\u6642\u8a18\u9304\u6a21\u578b\u3001\u63a8\u8ad6\u5f15\u64ce\u7248\u672c\u3001\u786c\u9ad4\u3001tokenizer\u3001\u8ca0\u8f09\u6a21\u578b\u3001\u6696\u6a5f\u8a2d\u5b9a\u8207 SLO \u9580\u6abb\uff0c\u5426\u5247\u65e5\u5f8c\u5f88\u96e3\u91cd\u73fe\u540c\u4e00\u4efd\u5831\u544a\u3002<\/p>\n\n\n\n

        \u56db\u3001AIPerf\uff1a\u9019\u4e9b\u6982\u5ff5\u5982\u4f55\u51fa\u73fe\u5728\u5be6\u969b\u5de5\u5177\u4e2d<\/strong><\/h1>\n\n\n\n

        AIPerf \u662f NVIDIA Dynamo \u5718\u968a\u958b\u6e90\u7684 LLM \u63a8\u8ad6\u6548\u80fd benchmark \u5de5\u5177\uff0c\u63a1\u7528 Apache 2.0 \u6388\u6b0a\uff0c\u4e5f\u662f GenAI-Perf \u7684\u5f8c\u7e7c\u5de5\u5177\u3002\u672c\u6587\u4e0d\u4ecb\u7d39\u64cd\u4f5c\u6d41\u7a0b\uff0c\u800c\u662f\u900f\u904e\u4e09\u9805\u8a2d\u8a08\uff0c\u5c0d\u7167\u524d\u9762\u63d0\u5230\u7684\u6982\u5ff5\u3002<\/p>\n\n\n\n

        \"AIPerf\u67b6\u69cb\"<\/figure>\n\n\n\n

        \u5716 4\u3000AIPerf \u7684\u4e09\u500b\u5e73\u9762\uff1a\u63a7\u5236\u5e73\u9762\u6c7a\u5b9a\u8ca0\u8f09\u6a21\u5f0f\uff0c\u8cc7\u6599\u5e73\u9762\u8ca0\u8cac\u91cf\u6e2c\uff0c\u5206\u6790\u5e73\u9762\u8ca0\u8cac\u7d71\u8a08<\/p>\n\n\n\n

        \u591a\u7a0b\u5e8f\u8207 credit\uff08\u8acb\u6c42\u8a31\u53ef\uff09\uff1a<\/strong>AIPerf \u6703\u8b93\u7d04\u5341\u500b\u670d\u52d9\u5206\u5225\u57f7\u884c\u5728\u7368\u7acb\u7a0b\u5e8f\u4e2d\uff0c\u4e26\u900f\u904e ZeroMQ \u6e9d\u901a\u3002\u6574\u9ad4\u5206\u70ba\u63a7\u5236\u5e73\u9762\uff08\u6c7a\u5b9a\u4f55\u6642\u9001\u51fa\u591a\u5c11\u8acb\u6c42\uff09\u3001\u8cc7\u6599\u5e73\u9762\uff08\u7531 worker \u5be6\u969b\u767c\u9001 HTTP \u8acb\u6c42\u4e26\u8a18\u9304\u6642\u9593\u6233\uff09\uff0c\u4ee5\u53ca\u5206\u6790\u5e73\u9762\uff08\u8a08\u7b97\u6307\u6a19\u3001\u532f\u51fa\u7d50\u679c\uff09\u3002AIPerf \u7684\u8ca0\u8f09\u63a7\u5236\u6838\u5fc3\u4e4b\u4e00\u662f credit\uff0c\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\u4e00\u5f35\u300c\u8acb\u6c42\u901a\u884c\u8b49\u300d\uff1aworker \u53d6\u5f97 credit \u5f8c\u624d\u80fd\u9001\u51fa\u4e00\u500b\u8acb\u6c42\uff0c\u6c92\u6709 credit \u5c31\u4e0d\u80fd\u9001\u3002\u56fa\u5b9a\u4e26\u767c\u6a21\u5f0f\u4e0b\uff0cTiming Manager \u4e00\u958b\u59cb\u767c\u51fa N \u5f35 credit\uff1b\u6bcf\u6536\u56de\u4e00\u5f35\uff0c\u624d\u88dc\u767c\u4e00\u5f35\uff0c\u9019\u5c31\u662f\u9589\u74b0\u3002Request rate \u6a21\u5f0f\u5247\u4f9d Poisson \u5230\u9054\u6392\u7a0b\u767c\u653e credit\uff0c\u4e0d\u5fc5\u7b49\u524d\u4e00\u5f35\u6b78\u9084\uff0c\u9019\u5c31\u662f\u958b\u74b0\u3002\u63db\u53e5\u8a71\u8aaa\uff0c\u8ca0\u8f09\u6a21\u578b\u7684\u6838\u5fc3\uff0c\u5c31\u662f\u6c7a\u5b9a\u300c\u5728\u4ec0\u9ebc\u689d\u4ef6\u4e0b\u767c\u51fa\u4e0b\u4e00\u5f35\u901a\u884c\u8b49\u300d\u3002<\/p>\n\n\n\n

        \u6642\u9593\u6233\u8207 tokenizer\uff1a<\/strong>Worker \u6703\u8a18\u9304\u8acb\u6c42\u9001\u51fa\u7684\u6642\u9593\uff0c\u4ee5\u53ca\u6bcf\u500b\u4e32\u6d41 chunk \u62b5\u9054\u7528\u6236\u7aef\u7684\u6642\u9593\u3002TTFT\u3001ITL \u8207 request latency \u90fd\u662f\u900f\u904e\u6642\u9593\u6233\u76f8\u6e1b\u8a08\u7b97\uff1bbenchmark duration \u5247\u4f9d\u6574\u6bb5\u6e2c\u8a66\u5be6\u969b\u7d93\u904e\u7684\u6642\u9593\u8a08\u7b97\u3002ISL \u8207 OSL \u7531 tokenizer \u8a08\u7b97\uff0c\u56e0\u6b64\u57f7\u884c benchmark \u6642\u5fc5\u9808\u6307\u5b9a tokenizer\uff0c\u800c\u4e14\u8981\u8207\u4f3a\u670d\u5668\u5be6\u969b\u8f09\u5165\u7684\u6a21\u578b\u4e00\u81f4\u3002<\/p>\n\n\n\n

        \u6383\u63cf\u8207\u81ea\u9069\u61c9\u641c\u5c0b\uff1a<\/strong>\u683c\u9ede\u6383\u63cf\u9810\u8a2d\u6703\u6e2c\u8a66\u516b\u500b\u4ee5\u5c0d\u6578\u5c3a\u5ea6\u5206\u5e03\u7684\u4e26\u767c\u9ede\uff1a1\u30013\u30017\u300119\u300152\u3001139\u3001373\u30011000\u3002\u6bcf\u500b\u9ede\u4e5f\u53ef\u4ee5\u91cd\u8907\u6e2c\u8a66\u591a\u6b21\uff0c\u4f30\u8a08\u7d50\u679c\u7684\u4fe1\u8cf4\u5340\u9593\u3002\u5982\u679c\u60f3\u76f4\u63a5\u627e\u51fa\u300c\u7b26\u5408 SLO \u7684\u6700\u9ad8\u4e26\u767c\u6578\u300d\uff0c\u53ef\u4ee5\u4f7f\u7528 max-concurrency-under-sla \u9019\u985e search recipe\u3002\u9810\u8a2d\u6f14\u7b97\u6cd5\u6703\u5148\u9010\u6b65\u63d0\u9ad8\u8ca0\u8f09\uff0c\u76f4\u5230\u627e\u5230\u7b2c\u4e00\u500b\u8d85\u51fa\u9580\u6abb\u7684\u9ede\uff0c\u518d\u628a\u53ef\u80fd\u7684\u908a\u754c\u7e2e\u5c0f\u5230\u4e00\u500b\u5340\u9593\uff1b\u63a5\u8457\u900f\u904e\u7b49\u5f35\u56de\u6b78\uff08isotonic regression\uff09\u964d\u4f4e\u91cf\u6e2c\u96dc\u8a0a\u3001\u63d2\u503c\u4f30\u7b97\u4ea4\u9ede\uff0c\u4e26\u5728\u908a\u754c\u9644\u8fd1\u88dc\u505a\u6e2c\u8a66\uff0c\u76f4\u5230\u5340\u9593\u7a84\u65bc 5%\u3002\u6574\u500b\u6d41\u7a0b\u5927\u7d04\u9700\u8981 13 \u5230 25 \u6b21\u8fed\u4ee3\u3002\u76f8\u8f03\u4e4b\u4e0b\uff0c\u82e5\u53ea\u6e2c\u56fa\u5b9a\u7684\u516b\u500b\u9ede\uff0c\u53ef\u80fd\u53ea\u80fd\u5224\u65b7\u5bb9\u91cf\u908a\u754c\u843d\u5728 139 \u5230 373 \u4e4b\u9593\u3002\u5982\u679c\u76ee\u6a19\u6539\u6210\u300c\u5728\u7b26\u5408\u9580\u6abb\u7684\u689d\u4ef6\u4e0b\uff0c\u627e\u51fa\u6700\u5927\u541e\u5410\u91cf\u300d\uff0cAIPerf \u5247\u53ef\u63a1\u7528\u8c9d\u8449\u65af\u6700\u4f73\u5316\uff08Bayesian optimization\uff09\uff1a\u4ee5\u9ad8\u65af\u904e\u7a0b\uff08Gaussian process\uff09\u5efa\u7acb\u300c\u4e26\u767c\u6578\u5c0d\u61c9\u541e\u5410\u91cf\u300d\u7684\u9810\u6e2c\u6a21\u578b\uff0c\u6bcf\u6b21\u6311\u9078\u9810\u6e2c\u8868\u73fe\u8f03\u597d\u6216\u4e0d\u78ba\u5b9a\u6027\u8f03\u9ad8\u7684\u9ede\u9032\u884c\u6e2c\u8a66\uff1b\u82e5\u9023\u7e8c\u591a\u6b21\u6c92\u6709\u6539\u5584\uff0c\u5c31\u505c\u6b62\u641c\u5c0b\u3002\u7576\u9700\u8981\u63a2\u7d22\u7684\u53c3\u6578\u7dad\u5ea6\u589e\u52a0\u6642\uff0c\u9019\u7a2e\u65b9\u6cd5\u76f8\u8f03\u65bc\u56fa\u5b9a\u683c\u9ede\u6383\u63cf\u6703\u66f4\u6709\u512a\u52e2\u3002<\/p>\n\n\n\n

        \u7d50\u8a9e<\/strong><\/h1>\n\n\n\n

        \u904b\u884c\u6548\u80fd\u7684\u554f\u984c\u8981\u5148\u6539\u5beb\u6210\u53ef\u4ee5\u91cf\u6e2c\u3001\u9a57\u8b49\u7684\u5f62\u5f0f\uff0c\u624d\u6709\u8fa6\u6cd5\u56de\u7b54\uff1a\u300c\u5728\u6307\u5b9a\u7684\u6a21\u578b\u3001\u63a8\u8ad6\u5f15\u64ce\u8207 ISL\/OSL \u5de5\u4f5c\u8ca0\u8f09\u4e0b\uff0c\u63a1\u7528\u9589\u74b0\u4e26\u767c\u6383\u63cf\uff0c\u4e26\u6392\u9664\u6696\u6a5f\u8cc7\u6599\u5f8c\uff0c\u80fd\u8b93 TTFT \u7684 p95 \u4f4e\u65bc X ms\u3001\u4e14 ITL \u7684 p95 \u4f4e\u65bc Y ms \u7684\u6700\u9ad8\u4e26\u767c\u6578\u662f\u591a\u5c11\uff1f\u5c0d\u61c9\u7684 goodput \u53c8\u662f\u591a\u5c11\uff1f\u300d\u6700\u5f8c\u53ef\u4ee5\u7528\u4e00\u5f35\u541e\u5410\u91cf\u8207\u5ef6\u9072\u66f2\u7dda\u3001\u4e00\u689d SLO \u9580\u6abb\u7dda\uff0c\u4ee5\u53ca\u5169\u8005\u7684\u4ea4\u9ede\u4f86\u5448\u73fe\u7b54\u6848\u3002<\/p>\n\n\n\n

        \u958b\u59cb\u6e2c\u8a66\u524d\uff0c\u5148\u78ba\u8a8d\u4ee5\u4e0b\u5e7e\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n

          \n
        1. \u662f\u5426\u555f\u7528\u4e32\u6d41\uff1f\u9019\u6703\u6c7a\u5b9a\u80fd\u5426\u91cf\u6e2c TTFT \u8207 ITL\u3002<\/li>\n\n\n\n
        2. \u662f\u5426\u63a7\u5236\u8f38\u51fa\u9577\u5ea6\uff1ftokenizer \u662f\u5426\u8207\u6a21\u578b\u4e00\u81f4\uff1f<\/li>\n\n\n\n
        3. \u8981\u56de\u7b54\u7cfb\u7d71\u5bb9\u91cf\uff0c\u4f7f\u7528\u9589\u74b0\u4e26\u767c\u6383\u63cf\uff1b\u8981\u9a57\u8b49\u5c16\u5cf0\u6d41\u91cf\u4e0b\u7684\u5ef6\u9072\uff0c\u4f7f\u7528\u958b\u74b0\u56fa\u5b9a\u5230\u9054\u7387\u3002\u958b\u74b0\u6e2c\u8a66\u53ef\u63a1\u7528 Poisson \u5230\u9054\u6a21\u5f0f\u3002<\/li>\n\n\n\n
        4. \u662f\u5426\u6392\u9664\u6696\u6a5f\u8acb\u6c42\uff1f\u540c\u4e00\u500b endpoint \u662f\u5426\u6709\u5176\u4ed6 benchmark \u6b63\u5728\u57f7\u884c\uff1f<\/li>\n\n\n\n
        5. \u9664\u4e86\u5e73\u5747\u503c\uff0c\u4e5f\u8981\u770b p95\u3001p99 \u7b49\u767e\u5206\u4f4d\u6578\uff1b\u9664\u4e86 throughput\uff0c\u4e5f\u8981\u770b goodput\u3002\u5b8c\u6574\u6e2c\u8a66\u8a2d\u5b9a\u8207\u6bcf\u7b46\u8acb\u6c42\u7d00\u9304\u90fd\u61c9\u4fdd\u5b58\u3002<\/li>\n<\/ol>\n\n\n\n

          \u5de5\u5177\u6703\u66f4\u65b0\u3001GPU \u6703\u66f4\u63db\uff0c\u4f46\u9019\u4efd\u6aa2\u67e5\u6e05\u55ae\u4ecd\u7136\u9069\u7528\u3002<\/p>\n\n\n\n

          \u53c3\u8003\u8cc7\u6599<\/strong><\/h1>\n\n\n\n
            \n
          • AIPerf \u539f\u59cb\u78bc\u8207\u6587\u4ef6\uff1ahttps:\/\/github.com\/ai-dynamo\/aiperf\uff08\u5305\u542b architecture\u3001metrics-reference\u3001comprehensive-llm-benchmarking\u3001timing-modes-reference\u3001search-recipes \u7b49\u6587\u4ef6\uff09<\/li>\n\n\n\n
          • Yu et al., “Orca: A Distributed Serving System for Transformer-Based Generative Models”, OSDI 2022\uff08continuous batching\uff09<\/li>\n\n\n\n
          • Zhong et al., “DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving”, OSDI 2024 (goodput\uff09<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"

            \u4ec0\u9ebc\u662f LLM Benchmark\uff1f
            \n\u300cLLM benchmark\u300d\u5e38\u898b\u6709\u5169\u7a2e\u610f\u601d\u3002\u80fd\u529b\u8a55\u6e2c benchmark\uff08\u4f8b\u5982 MMLU\u3001HumanEval \u7b49\u984c\u5eab\uff09\u95dc\u5fc3\u7684\u662f\u6a21\u578b\u7b54\u5f97\u5c0d\u4e0d\u5c0d\uff0c\u800c\u4e0d\u662f\u5728\u8a55\u4f30\u786c\u9ad4\u6548\u80fd\uff1b\u63a8\u8ad6\u6548\u80fd benchmark\u5247\u95dc\u5fc3\u540c\u4e00\u500b\u6a21\u578b\u90e8\u7f72\u5728\u7279\u5b9a\u786c\u9ad4\u8207\u63a8\u8ad6\u5f15\u64ce\u5f8c\uff0c\u56de\u61c9\u901f\u5ea6\u6709\u591a\u5feb\uff0c\u4ee5\u53ca\u53ef\u4ee5\u540c\u6642\u670d\u52d9\u591a\u5c11\u4f7f\u7528\u8005\u3002\u672c\u6587\u8a0e\u8ad6\u7684\u662f\u5f8c\u8005\uff0c\u4e5f\u5c31\u662f\u5927\u5bb6\u771f\u6b63\u60f3\u77e5\u9053\u7684\u4e8b\u3002<\/p>\n","protected":false},"author":253372376,"featured_media":15592,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_post_was_ever_published":false},"categories":[96987561,96987582],"tags":[96987624,96987632],"class_list":["post-15563","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technical-support","category-product-features","tag-ai-stack","tag-llm"],"blocksy_meta":[],"acf":[],"jetpack_shortlink":"https:\/\/wp.me\/ph344V-431","jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/i0.wp.com\/ai-stack.ai\/wp-content\/uploads\/2026\/10\/b67b9458.png?fit=1920%2C1080&quality=100&ct=202603031250&ssl=1","_links":{"self":[{"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/posts\/15563","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/users\/253372376"}],"replies":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/comments?post=15563"}],"version-history":[{"count":2,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/posts\/15563\/revisions"}],"predecessor-version":[{"id":15600,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/posts\/15563\/revisions\/15600"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/media\/15592"}],"wp:attachment":[{"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/media?parent=15563"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/categories?post=15563"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ai-stack.ai\/zh-hant\/wp-json\/wp\/v2\/tags?post=15563"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}