{"id":32471,"date":"2026-05-16T08:00:34","date_gmt":"2026-05-16T00:00:34","guid":{"rendered":"https:\/\/www.varidata.com\/uncategorized-zh-tw\/how-to-measure-real-llm-server-throughput\/"},"modified":"2026-05-15T15:25:40","modified_gmt":"2026-05-15T07:25:40","slug":"how-to-measure-real-llm-server-throughput","status":"publish","type":"post","link":"https:\/\/www.varidata.com\/zh-tw\/blog-tc\/how-to-measure-real-llm-server-throughput\/","title":{"rendered":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf"},"content":{"rendered":"\n<p>\u5982\u679c\u4f60\u5728\u73fe\u4ee3\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u4e0a\u57f7\u884c\u63a8\u8ad6\u5de5\u4f5c\u8ca0\u8f09\uff0c\u90a3\u9ebc<a href=\"https:\/\/www.varidata.com\/zh-tw\/blog-en\/llm-fine-tuning-gpu-hours-explained\/\" target=\"_blank\">LLM\u4f3a\u670d\u5668<\/a>\u541e\u5410\u91cf\u9019\u500b\u8a5e\u5c31\u4e0d\u8a72\u53ea\u662f\u4f9b\u61c9\u5546\u6295\u5f71\u7247\u4e0a\u7684\u8aaa\u6cd5\uff0c\u6216\u5be6\u9a57\u5ba4\u5c55\u793a\u4e2d\u7684\u6578\u5b57\u3002\u771f\u5be6\u5bb9\u91cf\u53d6\u6c7a\u65bc\u6392\u968a\u884c\u70ba\u3001token \u4e32\u6d41\u3001\u8acb\u6c42\u985e\u578b\u7d44\u5408\uff0c\u4ee5\u53ca\u7cfb\u7d71\u5728\u8ca0\u8f09\u4e0b\u7684\u5be6\u969b\u8868\u73fe\u3002\u5c0d\u65bc\u5728<a href=\"https:\/\/www.varidata.com\/zh-tw\/server\/los-angeles\/\" target=\"_blank\">\u7f8e\u570b\u4f3a\u670d\u5668\u74b0\u5883<\/a>\u4e2d\u5efa\u69cb\u7cfb\u7d71\u7684\u6280\u8853\u5718\u968a\u4f86\u8aaa\uff0c\u76ee\u6a19\u5176\u5be6\u5f88\u76f4\u63a5\uff1a\u8861\u91cf\u9019\u5957\u6280\u8853\u5806\u758a\u5728\u4e26\u767c\u4f7f\u7528\u8005\u3001\u9577\u63d0\u793a\u8a5e\u548c\u6301\u7e8c\u6d41\u91cf\u540c\u6642\u51fa\u73fe\u6642\uff0c\u7a76\u7adf\u80fd\u771f\u6b63\u4ea4\u4ed8\u4ec0\u9ebc\u3002<\/p>\n<p>\u5f88\u591a\u95dc\u65bc\u541e\u5410\u91cf\u7684\u8a0e\u8ad6\u4e4b\u6240\u4ee5\u5931\u771f\uff0c\u662f\u56e0\u70ba\u5b83\u5011\u628a\u8a9e\u8a00\u6a21\u578b\u670d\u52d9\u7576\u6210\u4e00\u822c Web \u6d41\u91cf\u4f86\u770b\u5f85\u3002\u9019\u7a2e\u7c21\u5316\u5f88\u5feb\u5c31\u6703\u5931\u6548\u3002\u751f\u6210\u5f0f\u5de5\u4f5c\u8ca0\u8f09\u81f3\u5c11\u5305\u542b\u5169\u500b\u4e0d\u540c\u968e\u6bb5\uff1a\u63d0\u793a\u8a5e\u8655\u7406\u548c token \u751f\u6210\u3002\u5b98\u65b9\u57fa\u6e96\u6e2c\u8a66\u8cc7\u6599\u901a\u5e38\u5c07\u5176\u7a31\u70ba prefill \u548c decode\uff0c\u800c\u5b83\u5011\u5c0d\u5ef6\u9072\u7684\u5f71\u97ff\u65b9\u5f0f\u5b8c\u5168\u4e0d\u540c\u3002\u9996 token \u6642\u9593\u4e0d\u50c5\u5305\u542b\u6392\u968a\u3001\u63d0\u793a\u8a5e\u8655\u7406\uff0c\u4e5f\u5305\u62ec\u7db2\u8def\u958b\u92b7\uff1b\u800c\u8f38\u51fa token \u541e\u5410\u91cf\u5247\u53cd\u6620\u4e32\u6d41\u8f38\u51fa\u958b\u59cb\u5f8c\u7cfb\u7d71\u6301\u7e8c\u751f\u6210\u7684\u6548\u7387\u3002\u8207\u55ae\u7d14\u7684\u8acb\u6c42\u6578\u76f8\u6bd4\uff0c\u9019\u4e9b\u6307\u6a19\u91cd\u8981\u5f97\u591a\u3002<\/p>\n<h2><strong>\u70ba\u4ec0\u9ebc\u300c\u771f\u5be6\u541e\u5410\u91cf\u300d\u4e0d\u540c\u65bc\u7406\u8ad6\u541e\u5410\u91cf<\/strong><\/h2>\n<p>\u7406\u8ad6\u6548\u80fd\u7e3d\u662f\u5f88\u6574\u9f4a\uff0c\u771f\u5be6\u6548\u80fd\u537b\u5f80\u5f80\u5145\u6eff\u96dc\u8a0a\u3002\u67d0\u53f0\u4f3a\u670d\u5668\u5728\u7d19\u9762\u898f\u683c\u4e0a\u53ef\u80fd\u76f8\u7576\u6f02\u4eae\uff0c\u4f46\u7576\u63d0\u793a\u8a5e\u9577\u5ea6\u589e\u52a0\u3001\u8a18\u61b6\u9ad4\u58d3\u529b\u5347\u9ad8\uff0c\u6216\u591a\u500b\u5de5\u4f5c\u968e\u6bb5\u540c\u6642\u9032\u5165 prefill \u968e\u6bb5\u6642\uff0c\u5b83\u7684\u8868\u73fe\u53ef\u80fd\u6703\u5b8c\u5168\u4e0d\u540c\u3002\u5b98\u65b9\u57fa\u6e96\u6e2c\u8a66\u6307\u5357\u6307\u51fa\uff0c\u66f4\u9577\u7684\u8f38\u5165\u5e8f\u5217\u6703\u63d0\u9ad8 prefill \u6210\u672c\uff0c\u4e26\u53ef\u80fd\u62c9\u9577\u9996 token \u6642\u9593\uff1b\u800c\u66f4\u9577\u7684\u8f38\u51fa\u5247\u6703\u5c0d decode \u548c token \u9593\u5ef6\u9072\u5e36\u4f86\u66f4\u5927\u58d3\u529b\u3002\u63db\u53e5\u8a71\u8aaa\uff0c\u4e00\u500b\u57fa\u6e96\u6578\u5b57\u6839\u672c\u7121\u6cd5\u6982\u62ec\u6574\u500b LLM \u670d\u52d9\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u5de5\u7a0b\u5e2b\u4e0d\u8a72\u518d\u554f\u300c\u541e\u5410\u91cf\u662f\u591a\u5c11\u300d\uff0c\u800c\u61c9\u8a72\u63d0\u51fa\u66f4\u6709\u50f9\u503c\u7684\u554f\u984c\uff1a<\/p>\n<ul>\n<li>\u5728\u4f4e\u3001\u4e2d\u3001\u9ad8\u4e26\u767c\u4e0b\uff0ctoken \u901f\u7387\u5206\u5225\u662f\u591a\u5c11\uff1f<\/li>\n<li>\u5728\u6392\u968a\u58d3\u529b\u4e0b\uff0c\u9996 token \u6642\u9593\u7684\u7a69\u5b9a\u6027\u5982\u4f55\uff1f<\/li>\n<li>\u5ef6\u9072\u5f9e\u54ea\u500b\u9ede\u958b\u59cb\u4e0a\u5347\u5f97\u6bd4\u541e\u5410\u91cf\u66f4\u5feb\uff1f<\/li>\n<li>\u63d0\u793a\u8a5e\u9577\u5ea6\u6703\u5982\u4f55\u6539\u8b8a\u6574\u9ad4\u66f2\u7dda\u5f62\u614b\uff1f<\/li>\n<li>\u7cfb\u7d71\u5728\u6301\u7e8c\u8ca0\u8f09\u4e0b\u662f\u5426\u80fd\u5920\u5e73\u9806\u9000\u5316\uff0c\u800c\u4e0d\u662f\u7a81\u7136\u5931\u7a69\uff1f<\/li>\n<\/ul>\n<p>\u9019\u4e9b\u554f\u984c\u63ed\u793a\u7684\u662f\u5be6\u969b\u904b\u4f5c\u72c0\u614b\uff0c\u800c\u4e0d\u662f\u5ba3\u50b3\u8cc7\u6599\u4e2d\u7684\u7406\u60f3\u5316\u7b97\u8853\u3002<\/p>\n<h2><strong>\u4f60\u771f\u6b63\u61c9\u8a72\u8ffd\u8e64\u7684\u6838\u5fc3\u6307\u6a19<\/strong><\/h2>\n<p>\u5728\u64b0\u5beb\u4efb\u4f55\u57fa\u6e96\u6e2c\u8a66\u8173\u672c\u4e4b\u524d\uff0c\u5148\u628a\u6307\u6a19\u5b9a\u7fa9\u6e05\u695a\u3002\u5982\u679c\u8861\u91cf\u541e\u5410\u91cf\u7684\u8a9e\u8a00\u672c\u8eab\u662f\u6a21\u7cca\u7684\uff0c\u90a3\u9ebc\u6700\u7d42\u7d50\u679c\u4e5f\u4e00\u5b9a\u662f\u6a21\u7cca\u7684\u3002<\/p>\n<ol>\n<li><strong>\u9996 token \u6642\u9593\uff08TTFT, Time to First Token\uff09\uff1a<\/strong>\u5f9e\u9001\u51fa\u8acb\u6c42\u5230\u6536\u5230\u7b2c\u4e00\u500b\u751f\u6210 token \u4e4b\u9593\u7684\u5ef6\u9072\u3002\u9019\u662f\u4e92\u52d5\u5f0f\u5834\u666f\u4e2d\u6700\u76f4\u89c0\u7684\u56de\u61c9\u6307\u6a19\u3002\u5b98\u65b9\u6587\u4ef6\u5c07\u5176\u8996\u70ba\u9762\u5411\u4f7f\u7528\u8005\u7684\u56de\u61c9\u6027\u6307\u6a19\uff0c\u4e26\u6307\u51fa\u5176\u4e2d\u5305\u542b\u7db2\u8def\u5ef6\u9072\u3001\u6392\u968a\u548c\u63d0\u793a\u8a5e\u8655\u7406\u3002<\/li>\n<li><strong>token \u9593\u5ef6\u9072\uff08Inter-Token Latency\uff09\uff1a<\/strong>\u4e32\u6d41\u8f38\u51fa\u958b\u59cb\u5f8c\uff0c\u76f8\u9130\u751f\u6210 token \u4e4b\u9593\u7684\u6642\u9593\u9593\u9694\u3002\u9019\u500b\u6307\u6a19\u80fd\u5920\u63ed\u793a decode \u968e\u6bb5\u7684\u8868\u73fe\uff0c\u4e26\u5e6b\u52a9\u5224\u65b7\u8f38\u51fa\u662f\u9806\u66a2\u9084\u662f\u5361\u9813\u3002<\/li>\n<li><strong>\u8f38\u51fa token \u541e\u5410\u91cf\uff08Output Token Throughput\uff09\uff1a<\/strong>\u6574\u500b\u7cfb\u7d71\u6bcf\u79d2\u751f\u6210\u7684 token \u6578\u91cf\u3002\u9019\u662f\u8861\u91cf\u751f\u6210\u5f0f\u670d\u52d9\u541e\u5410\u80fd\u529b\u6700\u76f4\u63a5\u7684\u6307\u6a19\u3002<\/li>\n<li><strong>\u8acb\u6c42\u541e\u5410\u91cf\uff08Request Throughput\uff09\uff1a<\/strong>\u6bcf\u79d2\u5b8c\u6210\u7684\u8acb\u6c42\u6578\u3002\u5b83\u7576\u7136\u6709\u7528\uff0c\u4f46\u53ea\u6709\u5728\u8f38\u5165\u548c\u8f38\u51fa\u9577\u5ea6\u88ab\u56b4\u683c\u63a7\u5236\u6642\u624d\u771f\u6b63\u6709\u53c3\u8003\u50f9\u503c\u3002<\/li>\n<li><strong>\u5c3e\u7aef\u5ef6\u9072\uff08Tail Latency\uff09\uff1a<\/strong>p95 \u6216 p99 \u5f80\u5f80\u6bd4\u4e00\u500b\u597d\u770b\u7684\u5e73\u5747\u503c\u66f4\u91cd\u8981\uff0c\u56e0\u70ba\u751f\u7522\u74b0\u5883\u4e2d\u7684\u771f\u5be6\u554f\u984c\u901a\u5e38\u90fd\u51fa\u73fe\u5728\u5c3e\u7aef\u3002<\/li>\n<li><strong>\u5931\u6557\u7387\uff08Failure Rate\uff09\uff1a<\/strong>\u903e\u6642\u3001\u7a7a\u56de\u61c9\u3001\u91cd\u8a66\u4ee5\u53ca\u8207\u8a18\u61b6\u9ad4\u76f8\u95dc\u7684\u5d29\u6f70\uff0c\u90fd\u61c9\u8a72\u7d0d\u5165\u6700\u7d42\u5831\u544a\u3002<\/li>\n<\/ol>\n<p>\u5c0d\u6280\u8853\u8b80\u8005\u4f86\u8aaa\uff0c\u6700\u95dc\u9375\u7684\u539f\u5247\u662f\uff1a\u4e0d\u8981\u5728\u6c92\u6709\u5ef6\u9072\u8cc7\u6599\u7684\u60c5\u6cc1\u4e0b\u8ac7\u541e\u5410\u91cf\uff0c\u4e5f\u4e0d\u8981\u5728\u4e0d\u8aaa\u660e\u8ca0\u8f09\u578b\u614b\u7684\u524d\u63d0\u4e0b\u8ac7\u5ef6\u9072\u3002<\/p>\n<h2><strong>\u5728\u958b\u59cb\u6e2c\u8a66\u4e4b\u524d\uff0c\u5148\u7406\u89e3 prefill \u548c decode<\/strong><\/h2>\n<p>\u5927\u591a\u6578\u57fa\u6e96\u6e2c\u8a66\u932f\u8aa4\uff0c\u6839\u6e90\u90fd\u5728\u65bc\u5ffd\u8996\u4e86\u968e\u6bb5\u884c\u70ba\u3002Prefill \u662f\u6a21\u578b\u8655\u7406\u8f38\u5165\u4e0a\u4e0b\u6587\u7684\u968e\u6bb5\uff0cdecode \u5247\u662f\u6a21\u578b\u9010\u6b65\u8f38\u51fa token \u7684\u968e\u6bb5\u3002\u591a\u4efd\u5b98\u65b9\u8cc7\u6599\u90fd\u5f37\u8abf\uff0c\u9019\u5169\u500b\u968e\u6bb5\u5c0d\u7cfb\u7d71\u9020\u6210\u7684\u58d3\u529b\u4f86\u6e90\u4e26\u4e0d\u76f8\u540c\uff0c\u800c\u4e14 prefill \u968e\u6bb5\u7684\u4e26\u767c\u60c5\u6cc1\u6703\u986f\u8457\u5f71\u97ff\u6392\u968a\u6df1\u5ea6\u8207\u9996 token \u5ef6\u9072\u3002<\/p>\n<p>\u9019\u4e4b\u6240\u4ee5\u91cd\u8981\uff0c\u662f\u56e0\u70ba\u300c\u77ed\u63d0\u793a\u8a5e + \u9577\u8f38\u51fa\u300d\u548c\u300c\u9577\u63d0\u793a\u8a5e + \u77ed\u8f38\u51fa\u300d\u53ef\u80fd\u5448\u73fe\u5b8c\u5168\u4e0d\u540c\u7684\u7cfb\u7d71\u74f6\u9838\u3002\u5982\u679c\u4f60\u7684\u5de5\u4f5c\u8ca0\u8f09\u6d89\u53ca\u6aa2\u7d22\u3001\u7a0b\u5f0f\u78bc\u4e0a\u4e0b\u6587\u6216\u591a\u8f2a\u8a18\u61b6\uff0c\u90a3\u9ebc prefill \u5f88\u53ef\u80fd\u4e3b\u5c0e\u4f7f\u7528\u8005\u9ad4\u9a57\uff1b\u5982\u679c\u4f60\u7684\u5de5\u4f5c\u8ca0\u8f09\u9700\u8981\u6301\u7e8c\u751f\u6210\u9577\u6587\u672c\uff0c\u90a3\u9ebc decode \u6548\u7387\u5c31\u53ef\u80fd\u6210\u70ba\u771f\u6b63\u7684\u4e0a\u9650\u3002<\/p>\n<p>\u4e00\u500b\u5be6\u7528\u7684\u601d\u8003\u6a21\u578b\u53ef\u4ee5\u9019\u6a23\u7406\u89e3\uff1a<\/p>\n<ul>\n<li><strong>Prefill \u4e3b\u5c0e\u578b\u8ca0\u8f09\uff1a<\/strong>\u5c0d\u63d0\u793a\u8a5e\u9577\u5ea6\u3001\u6392\u968a\u8207\u6e96\u5165\u63a7\u5236\u66f4\u654f\u611f<\/li>\n<li><strong>Decode \u4e3b\u5c0e\u578b\u8ca0\u8f09\uff1a<\/strong>\u5c0d token \u6392\u7a0b\u3001\u8a18\u61b6\u9ad4\u983b\u5bec\u548c\u6279\u6b21\u6548\u7387\u66f4\u654f\u611f<\/li>\n<li><strong>\u6df7\u5408\u578b\u8ca0\u8f09\uff1a<\/strong>\u5169\u8005\u90fd\u654f\u611f\uff0c\u800c\u9019\u5176\u5be6\u624d\u662f\u5f88\u591a\u751f\u7522\u7cfb\u7d71\u7684\u771f\u5be6\u72c0\u614b<\/li>\n<\/ul>\n<h2><strong>\u5982\u4f55\u8a2d\u8a08\u4e00\u500b\u8cbc\u8fd1\u751f\u7522\u74b0\u5883\u7684\u57fa\u6e96\u6e2c\u8a66<\/strong><\/h2>\n<p>\u53ea\u6709\u7576\u57fa\u6e96\u6e2c\u8a66\u63a5\u8fd1\u4f60\u8a08\u756b\u57f7\u884c\u7684\u771f\u5be6\u670d\u52d9\u6642\uff0c\u5b83\u624d\u6709\u50f9\u503c\u3002\u9019\u4e26\u4e0d\u610f\u5473\u8457\u8981\u8907\u88fd\u6bcf\u4e00\u500b\u908a\u7de3\u60c5\u5883\uff0c\u800c\u662f\u8981\u63a7\u5236\u90a3\u4e9b\u771f\u6b63\u5f71\u97ff\u7cfb\u7d71\u884c\u70ba\u7684\u95dc\u9375\u8b8a\u6578\u3002<\/p>\n<ol>\n<li><strong>\u56fa\u5b9a\u6a21\u578b\u8207\u57f7\u884c\u8def\u5f91\u3002<\/strong>\u9664\u975e\u67d0\u9805\u8b8a\u66f4\u672c\u8eab\u5c31\u662f\u5be6\u9a57\u76ee\u6a19\uff0c\u5426\u5247\u4e0d\u8981\u5728\u91cf\u5316\u65b9\u5f0f\u3001\u6392\u7a0b\u53c3\u6578\u6216\u5e73\u884c\u914d\u7f6e\u8b8a\u5316\u5f8c\u6a6b\u5411\u6bd4\u8f03\u7d50\u679c\u3002<\/li>\n<li><strong>\u56fa\u5b9a\u63d0\u793a\u8a5e\u8207\u8f38\u51fa\u7bc4\u570d\u3002<\/strong>\u6c92\u6709 token \u6578\u91cf\u9650\u5236\u7684\u541e\u5410\u91cf\u8cc7\u6599\uff0c\u5e7e\u4e4e\u6c92\u6709\u89e3\u91cb\u50f9\u503c\u3002<\/li>\n<li><strong>\u5148\u5b8c\u6210\u9810\u71b1\u3002<\/strong>\u9996\u6b21\u57f7\u884c\u7684\u984d\u5916\u958b\u92b7\u53ef\u80fd\u4f86\u81ea\u60f0\u6027\u521d\u59cb\u5316\u3001\u5716\u5f62\u64f7\u53d6\u6216\u5feb\u53d6\u586b\u5145\uff0c\u9019\u6703\u6c61\u67d3\u7d50\u679c\u3002<\/li>\n<li><strong>\u5c07\u55ae\u4f7f\u7528\u8005\u6e2c\u8a66\u548c\u4e26\u767c\u6e2c\u8a66\u5206\u958b\u3002<\/strong>\u524d\u8005\u7528\u65bc\u89c0\u5bdf\u57fa\u790e\u56de\u61c9\uff0c\u5f8c\u8005\u7528\u65bc\u8a55\u4f30\u670d\u52d9\u5bb9\u91cf\u3002<\/li>\n<li><strong>\u6e2c\u8a66\u6642\u9593\u8981\u8db3\u5920\u9577\u3002<\/strong>\u904e\u77ed\u7684\u7a81\u767c\u6e2c\u8a66\u6703\u63a9\u84cb\u6eab\u5ea6\u3001\u6392\u968a\u6216\u8a18\u61b6\u9ad4\u5c64\u9762\u7684\u554f\u984c\u3002<\/li>\n<li><strong>\u8490\u96c6\u7cfb\u7d71\u9059\u6e2c\u8cc7\u6599\u3002<\/strong>GPU \u4f7f\u7528\u7387\u3001\u8a18\u61b6\u9ad4\u5360\u7528\u3001CPU \u58d3\u529b\u548c\u7db2\u8def\u8868\u73fe\u4e0d\u662f\u53ef\u6709\u53ef\u7121\u7684\u9644\u8a3b\uff0c\u5b83\u5011\u662f\u89e3\u91cb\u7d50\u679c\u7684\u95dc\u9375\u5c64\u3002<\/li>\n<\/ol>\n<p>\u5b98\u65b9\u7684\u751f\u6210\u5f0f\u7cfb\u7d71\u57fa\u6e96\u6e2c\u8a66\u5de5\u5177\u901a\u5e38\u652f\u63f4\u5408\u6210\u8f38\u5165\u6216\u8cc7\u6599\u96c6\u9a45\u52d5\u8f38\u5165\u3001\u57fa\u65bc\u4e26\u767c\u7684\u58d3\u529b\u6e2c\u8a66\u3001\u57fa\u65bc\u8acb\u6c42\u901f\u7387\u7684\u58d3\u529b\u6e2c\u8a66\uff0c\u4ee5\u53ca\u532f\u51fa\u65e5\u8a8c\u4f9b\u5f8c\u7e8c\u5206\u6790\u3002\u9019\u4f7f\u5de5\u7a0b\u5718\u968a\u53ef\u4ee5\u5efa\u7acb\u53ef\u91cd\u8907\u7684\u6e2c\u8a66\u6d41\u7a0b\uff0c\u800c\u4e0d\u662f\u53ea\u622a\u5e7e\u5f35\u7d42\u7aef\u6a5f\u756b\u9762\u4e86\u4e8b\u3002<\/p>\n<h2><strong>\u4e00\u5957\u9069\u5408\u6975\u5ba2\u7dad\u904b\u8996\u89d2\u7684\u5be6\u7528\u6e2c\u8a66\u77e9\u9663<\/strong><\/h2>\n<p>\u5982\u679c\u4f60\u7684\u7db2\u7ad9\u9762\u5411\u7f8e\u570b\u57fa\u790e\u8a2d\u65bd\u63a1\u8cfc\u8005\uff0c\u90a3\u9ebc\u8b80\u8005\u591a\u534a\u662f\u5728\u505a\u67b6\u69cb\u9078\u578b\uff0c\u800c\u4e0d\u662f\u95b1\u8b80\u7d14\u7406\u8ad6\u3002\u56e0\u6b64\uff0c\u63d0\u4f9b\u4e00\u5957\u80fd\u76f4\u63a5\u8dd1\u8d77\u4f86\u7684\u6e2c\u8a66\u77e9\u9663\uff0c\u9060\u6bd4\u62bd\u8c61\u89e3\u91cb\u66f4\u6709\u50f9\u503c\u3002<\/p>\n<ul>\n<li><strong>\u60c5\u5883 A\uff1a<\/strong>\u77ed\u63d0\u793a\u8a5e\u3001\u77ed\u8f38\u51fa\u3001\u4f4e\u4e26\u767c<\/li>\n<li><strong>\u60c5\u5883 B\uff1a<\/strong>\u77ed\u63d0\u793a\u8a5e\u3001\u9577\u8f38\u51fa\u3001\u4e2d\u7b49\u4e26\u767c<\/li>\n<li><strong>\u60c5\u5883 C\uff1a<\/strong>\u9577\u63d0\u793a\u8a5e\u3001\u77ed\u8f38\u51fa\u3001\u4e2d\u7b49\u4e26\u767c<\/li>\n<li><strong>\u60c5\u5883 D\uff1a<\/strong>\u9577\u63d0\u793a\u8a5e\u3001\u9577\u8f38\u51fa\u3001\u9010\u7d1a\u63d0\u5347\u4e26\u767c<\/li>\n<li><strong>\u60c5\u5883 E\uff1a<\/strong>\u6df7\u5408\u63d0\u793a\u8a5e\u9577\u5ea6\uff0c\u4e26\u758a\u52a0\u6301\u7e8c\u80cc\u666f\u6d41\u91cf<\/li>\n<\/ul>\n<p>\u7136\u5f8c\u4ee5\u968e\u68af\u65b9\u5f0f\u9010\u6b65\u63d0\u9ad8\u4e26\u767c\uff0c\u800c\u4e0d\u662f\u4e00\u958b\u59cb\u5c31\u628a\u4e26\u767c\u503c\u62c9\u5230\u6975\u9650\u3002\u9019\u6a23\u505a\u53ef\u4ee5\u627e\u51fa\u66f2\u7dda\u7684\u62d0\u9ede\uff0c\u4e5f\u5c31\u662f\u8f38\u51fa token \u541e\u5410\u91cf\u4e0d\u518d\u8db3\u4ee5\u62b5\u6d88\u5ef6\u9072\u4ee3\u50f9\u7684\u90a3\u500b\u4f4d\u7f6e\u3002\u5b98\u65b9\u7684\u751f\u6210\u5f0f\u670d\u52d9\u57fa\u6e96\u6e2c\u8a66\u5efa\u8b70\u4e5f\u5f37\u8abf\uff0c\u61c9\u4ed4\u7d30\u6383\u63cf\u4e26\u767c\u7a0b\u5ea6\u4ee5\u53ca\u8f38\u5165\u8f38\u51fa\u9577\u5ea6\uff0c\u56e0\u70ba\u9019\u4e9b\u53c3\u6578\u6703\u76f4\u63a5\u6c7a\u5b9a\u6e2c\u8a66\u7d50\u679c\u662f\u5426\u5177\u6709\u5be6\u969b\u610f\u7fa9\u3002<\/p>\n<p>\u5c0d\u5f88\u591a\u5718\u968a\u800c\u8a00\uff0c\u6700\u6709\u50f9\u503c\u7684\u5716\u8868\u4e26\u4e0d\u662f\u300c\u6700\u9ad8\u541e\u5410\u91cf\u300d\uff0c\u800c\u662f\u300c\u670d\u52d9\u5f9e\u56de\u61c9\u6d41\u66a2\u8b8a\u5f97\u9072\u6eef\u7684\u8f49\u6298\u9ede\u300d\u3002<\/p>\n<h2><strong>\u6e2c\u8a66\u904e\u7a0b\u4e2d\u61c9\u8a72\u8a18\u9304\u4ec0\u9ebc<\/strong><\/h2>\n<p>\u55ae\u7d14\u7684\u57fa\u6e96\u6e2c\u8a66\u8f38\u51fa\u6c38\u9060\u4e0d\u5920\u3002\u4f60\u9700\u8981\u628a\u8acb\u6c42\u7aef\u8996\u89d2\u548c\u7cfb\u7d71\u7aef\u8996\u89d2\u914d\u5c0d\u8d77\u4f86\u770b\u3002<\/p>\n<ul>\n<li>\u8acb\u6c42\u958b\u59cb\u6642\u9593\u548c\u5b8c\u6210\u6642\u9593<\/li>\n<li>\u9996 token \u6642\u9593\u6233\u8a18<\/li>\n<li>\u8f38\u5165 token \u6578\u548c\u8f38\u51fa token \u6578<\/li>\n<li>\u6bcf\u500b\u8acb\u6c42\u7684\u72c0\u614b\uff0c\u5305\u62ec\u5931\u6557\u689d\u4ef6<\/li>\n<li>\u6b63\u5728\u8655\u7406\u4e2d\u7684\u4e26\u767c\u8acb\u6c42\u6578<\/li>\n<li>GPU \u4f7f\u7528\u7387\u548c\u986f\u793a\u8a18\u61b6\u9ad4\u5360\u7528<\/li>\n<li>\u4e3b\u6a5f\u4e0a\u7684 CPU \u8207 RAM \u58d3\u529b<\/li>\n<li>\u5982\u679c\u7528\u6236\u7aef\u4f4d\u65bc\u9060\u7aef\uff0c\u9084\u8981\u8a18\u9304\u7db2\u8def\u5ef6\u9072<\/li>\n<\/ul>\n<p>\u7576\u6578\u5b57\u770b\u8d77\u4f86\u4e0d\u5c0d\u52c1\u6642\uff0c\u9019\u4e9b\u65e5\u8a8c\u80fd\u5e6b\u52a9\u4f60\u5224\u65b7\u554f\u984c\u7a76\u7adf\u4f86\u81ea\u6392\u968a\u5806\u7a4d\u3001\u63d0\u793a\u8a5e\u81a8\u8139\u3001decode \u8b8a\u6162\uff0c\u9084\u662f\u4e3b\u6a5f\u5c64\u9762\u7684\u74f6\u9838\u3002\u6c92\u6709\u9019\u4e9b\u8cc7\u8a0a\uff0c\u5de5\u7a0b\u5e2b\u5f80\u5f80\u6703\u628a\u554f\u984c\u6b78\u56e0\u5230\u932f\u8aa4\u7684\u5c64\u7d1a\u3002<\/p>\n<h2><strong>\u90a3\u4e9b\u6703\u8b93\u7d50\u679c\u5931\u53bb\u610f\u7fa9\u7684\u5e38\u898b\u6e2c\u8a66\u9677\u9631<\/strong><\/h2>\n<p>\u5728 LLM \u541e\u5410\u91cf\u6e2c\u8a66\u4e2d\uff0c\u6709\u4e00\u4e9b\u932f\u8aa4\u5e7e\u4e4e\u53cd\u8986\u51fa\u73fe\u3002<\/p>\n<ol>\n<li><strong>\u53ea\u6e2c\u4e00\u500b\u4e26\u767c\u503c\u3002<\/strong>\u9019\u6a23\u6839\u672c\u770b\u4e0d\u5230\u7cfb\u7d71\u66f2\u7dda\u7684\u5f62\u72c0\u3002<\/li>\n<li><strong>\u5ffd\u8996\u63d0\u793a\u8a5e\u9577\u5ea6\u3002<\/strong>\u6975\u5c0f\u7684\u5408\u6210\u63d0\u793a\u8a5e\u5e7e\u4e4e\u7121\u6cd5\u4ee3\u8868\u6aa2\u7d22\u578b\u6d41\u91cf\u3002<\/li>\n<li><strong>\u53ea\u770b\u5e73\u5747\u5ef6\u9072\u3002<\/strong>\u4f7f\u7528\u8005\u771f\u6b63\u6703\u62b1\u6028\u7684\u662f\u5c3e\u7aef\u5ef6\u9072\u3002<\/li>\n<li><strong>\u628a\u51b7\u555f\u52d5\u548c\u71b1\u555f\u52d5\u6df7\u5728\u4e00\u8d77\u3002<\/strong>\u521d\u59cb\u5316\u958b\u92b7\u6703\u6c61\u67d3\u6a23\u672c\u3002<\/li>\n<li><strong>\u5728\u4e0d\u540c token \u9810\u7b97\u4e0b\u6bd4\u8f03\u8acb\u6c42\u541e\u5410\u91cf\u3002<\/strong>\u9019\u6839\u672c\u4e0d\u662f\u540c\u4e00\u7dad\u5ea6\u7684\u6bd4\u8f03\u3002<\/li>\n<li><strong>\u5831\u544a tokens per second \u6642\uff0c\u4e0d\u8aaa\u660e\u662f\u7cfb\u7d71\u7e3d\u541e\u5410\u9084\u662f\u55ae\u4e00\u4f7f\u7528\u8005\u541e\u5410\u3002<\/strong>\u5b98\u65b9\u8cc7\u6599\u4e4b\u6240\u4ee5\u5340\u5206\u9019\u5169\u7a2e\u8996\u89d2\uff0c\u662f\u6709\u5145\u5206\u7406\u7531\u7684\u3002<\/li>\n<li><strong>\u8b93\u7528\u6236\u7aef\u5148\u6210\u70ba\u74f6\u9838\u3002<\/strong>\u6548\u80fd\u4e0d\u8db3\u7684\u58d3\u6e2c\u7aef\u6703\u507d\u9020\u4f3a\u670d\u5668\u4e0a\u9650\u3002<\/li>\n<\/ol>\n<p>\u9084\u6709\u4e00\u500b\u5bb9\u6613\u88ab\u5ffd\u7565\u7684\u9677\u9631\u5c24\u5176\u503c\u5f97\u55ae\u7368\u63d0\u51fa\uff1a\u5354\u8abf\u907a\u6f0f\u3002\u95dc\u65bc prefill \u4e26\u767c\u7684\u5b98\u65b9\u6587\u4ef6\u63d0\u5230\uff0c\u5982\u679c\u57fa\u6e96\u6e2c\u8a66\u5de5\u5177\u5728\u7b49\u5f85\u7a7a\u9592\u69fd\u4f4d\u6642\u81ea\u52d5\u653e\u6162\u7bc0\u594f\uff0c\u90a3\u9ebc\u6574\u500b\u6e2c\u8a66\u53ef\u80fd\u6703\u7121\u610f\u4e2d\u628a\u81ea\u8eab\u7bc0\u6d41\u5230\u8207\u4f3a\u670d\u5668\u5bb9\u91cf\u4e00\u81f4\u3002\u9019\u6703\u8b93\u5ef6\u9072\u8868\u73fe\u770b\u8d77\u4f86\u6bd4\u771f\u5be6\u958b\u653e\u6d41\u91cf\u60c5\u5883\u4e0b\u66f4\u597d\u3002<\/p>\n<h2><strong>\u5982\u4f55\u50cf\u7dad\u904b\u5de5\u7a0b\u5e2b\u4e00\u6a23\u89e3\u8b80\u7d50\u679c<\/strong><\/h2>\n<p>\u7576\u6e2c\u8a66\u7d50\u675f\u5f8c\uff0c\u4e0d\u8981\u6025\u8457\u628a\u6700\u5927\u7684\u6578\u5b57\u7576\u6210\u8d0f\u5bb6\u3002\u771f\u6b63\u6709\u50f9\u503c\u7684\u89e3\u8b80\uff0c\u6c38\u9060\u662f\u5728\u5206\u6790\u6b0a\u8861\u95dc\u4fc2\u3002<\/p>\n<ul>\n<li>\u5982\u679c TTFT \u4e00\u76f4\u5f88\u4f4e\uff0c\u4f46 token \u4e32\u6d41\u8b8a\u5f97\u65b7\u65b7\u7e8c\u7e8c\uff0c\u90a3\u9ebc\u9650\u5236\u56e0\u7d20\u5f88\u53ef\u80fd\u51fa\u5728 decode \u968e\u6bb5\u3002<\/li>\n<li>\u5982\u679c\u9577\u63d0\u793a\u8a5e\u8b93 TTFT \u6025\u5287\u4e0a\u5347\uff0c\u90a3\u9ebc\u554f\u984c\u53ef\u80fd\u4f86\u81ea prefill \u6216\u6392\u968a\u6e96\u5165\u3002<\/li>\n<li>\u5982\u679c\u541e\u5410\u91cf\u4e0a\u5347\u4e86\uff0c\u4f46 p99 \u5ef6\u9072\u4e5f\u66b4\u6f32\uff0c\u90a3\u9ebc\u4f60\u5f88\u53ef\u80fd\u5df2\u7d93\u8d85\u51fa\u53ef\u7528\u904b\u4f5c\u5340\u9593\u3002<\/li>\n<li>\u5982\u679c\u7d50\u679c\u5f88\u5dee\u6642 GPU \u4f7f\u7528\u7387\u537b\u4e0d\u9ad8\uff0c\u90a3\u9ebc\u74f6\u9838\u53ef\u80fd\u5728\u6392\u7a0b\u3001CPU \u8655\u7406\u6d41\u7a0b\u6216\u7528\u6236\u7aef\u8def\u5f91\u4e0a\u3002<\/li>\n<li>\u5982\u679c\u5931\u6557\u53ea\u5728\u6301\u7e8c\u8ca0\u8f09\u4e0b\u51fa\u73fe\uff0c\u90a3\u9ebc\u4f60\u7684\u6280\u8853\u5806\u758a\u66f4\u53ef\u80fd\u662f\u7a69\u5b9a\u6027\u51fa\u4e86\u554f\u984c\uff0c\u800c\u4e0d\u662f\u7b97\u529b\u672c\u8eab\u4e0d\u5920\u3002<\/li>\n<\/ul>\n<p>\u512a\u79c0\u7684\u7dad\u904b\u4eba\u54e1\u95dc\u6ce8\u7684\u4e0d\u662f\u300c\u6700\u5927\u6578\u5b57\u300d\uff0c\u800c\u662f\u300c\u9ad8\u6548\u5340\u9593\u300d\u3002\u6240\u8b02\u9ad8\u6548\u5340\u9593\uff0c\u5c31\u662f\u541e\u5410\u91cf\u5920\u5f37\u3001\u5ef6\u9072\u4ecd\u53ef\u9810\u6e2c\u3001\u5c3e\u7aef\u8868\u73fe\u4e5f\u4ecd\u7b26\u5408\u61c9\u7528\u8981\u6c42\u7684\u90a3\u4e00\u6bb5\u5de5\u4f5c\u5340\u9593\u3002<\/p>\n<h2><strong>\u70ba\u4ec0\u9ebc\u9019\u5c0d\u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528\u548c\u4f3a\u670d\u5668\u8a17\u7ba1\u898f\u5283\u5f88\u91cd\u8981<\/strong><\/h2>\n<p>\u5c0d\u6b63\u5728\u8a55\u4f30<strong>\u4f3a\u670d\u5668\u79df\u7528<\/strong>\u9084\u662f<strong>\u4f3a\u670d\u5668\u8a17\u7ba1<\/strong>\u7684\u5718\u968a\u4f86\u8aaa\uff0c\u541e\u5410\u91cf\u6e2c\u8a66\u7d55\u4e0d\u53ea\u662f\u5be6\u9a57\u5ba4\u88e1\u7684\u7df4\u7fd2\u3002\u5b83\u6703\u76f4\u63a5\u5f71\u97ff\u5bb9\u91cf\u898f\u5283\u3001\u8cc7\u6e90\u5bc6\u5ea6\u8a2d\u8a08\u548c\u670d\u52d9\u578b\u614b\u9078\u64c7\u3002\u4e00\u4efd\u7d2e\u5be6\u7684\u57fa\u6e96\u5831\u544a\u80fd\u544a\u8a34\u4f60\uff0c\u7576\u524d\u7cfb\u7d71\u66f4\u9069\u5408\u570d\u7e5e\u56de\u61c9\u901f\u5ea6\u6700\u4f73\u5316\u3001\u570d\u7e5e\u6279\u6b21\u6548\u7387\u6700\u4f73\u5316\uff0c\u9084\u662f\u570d\u7e5e\u8a18\u61b6\u9ad4\u9918\u88d5\u6700\u4f73\u5316\u3002\u5b83\u4e5f\u80fd\u5e6b\u52a9\u4f60\u4f30\u7b97\uff0c\u5728\u670d\u52d9\u54c1\u8cea\u958b\u59cb\u60e1\u5316\u4e4b\u524d\uff0c\u4e00\u5957\u90e8\u7f72\u7a76\u7adf\u80fd\u652f\u6490\u591a\u5c11\u5de5\u4f5c\u968e\u6bb5\u3002<\/p>\n<p>\u9019\u5728\u7f8e\u570b\u4f3a\u670d\u5668\u90e8\u7f72\u4e2d\u5c24\u5176\u91cd\u8981\uff0c\u56e0\u70ba\u9019\u88e1\u5c0d\u5ef6\u9072\u7684\u671f\u671b\u901a\u5e38\u6bd4\u8f03\u56b4\u683c\uff0c\u800c\u751f\u7522\u6d41\u91cf\u53c8\u53ef\u80fd\u96a8\u8457\u5730\u5340\u3001\u5de5\u4f5c\u8ca0\u8f09\u6a21\u5f0f\u548c\u696d\u52d9\u6642\u6bb5\u767c\u751f\u8b8a\u5316\u3002\u505a\u804a\u5929\u5f0f\u4e92\u52d5\u7684\u5718\u968a\uff0c\u4e5f\u8a31\u66f4\u770b\u91cd\u9996 token \u901f\u5ea6\uff1b\u57f7\u884c\u5f8c\u53f0\u6279\u91cf\u751f\u6210\u7684\u5718\u968a\uff0c\u5247\u53ef\u80fd\u66f4\u95dc\u6ce8\u7e3d\u9ad4\u8f38\u51fa token \u541e\u5410\u91cf\u3002\u57fa\u6e96\u6e2c\u8a66\u61c9\u8a72\u53cd\u6620\u9019\u7a2e\u71df\u904b\u73fe\u5be6\uff0c\u800c\u4e0d\u662f\u5f37\u884c\u8ffd\u6c42\u67d0\u500b\u653e\u8af8\u56db\u6d77\u7686\u6e96\u7684\u55ae\u4e00\u76ee\u6a19\u3002<\/p>\n<h2><strong>\u4e00\u5957\u53ef\u4ee5\u53cd\u8986\u91cd\u7528\u7684\u7c21\u55ae\u6e2c\u8a66\u6d41\u7a0b<\/strong><\/h2>\n<ol>\n<li>\u8a18\u9304\u786c\u9ad4\u548c\u8edf\u9ad4\u74b0\u5883\u3002<\/li>\n<li>\u7528\u4e00\u7d44\u56fa\u5b9a\u7684\u5c0f\u8acb\u6c42\u5c0d\u670d\u52d9\u9032\u884c\u9810\u71b1\u3002<\/li>\n<li>\u5148\u8dd1\u57fa\u7dda\u55ae\u8acb\u6c42\u6e2c\u8a66\u3002<\/li>\n<li>\u9010\u6b65\u6383\u63cf\u4e26\u767c\u7a0b\u5ea6\u3002<\/li>\n<li>\u7528\u66f4\u9577\u7684\u63d0\u793a\u8a5e\u91cd\u8907\u6383\u63cf\u3002<\/li>\n<li>\u518d\u7528\u66f4\u9577\u7684\u8f38\u51fa\u91cd\u8907\u4e00\u6b21\u3002<\/li>\n<li>\u57f7\u884c\u6301\u7e8c\u6027\u6e2c\u8a66\uff0c\u89c0\u5bdf\u6f02\u79fb\u548c\u5931\u6557\u60c5\u6cc1\u3002<\/li>\n<li>\u5c07 TTFT\u3001token \u9593\u5ef6\u9072\u3001\u5c3e\u7aef\u5ef6\u9072\u548c\u7e3d\u9ad4 token \u541e\u5410\u91cf\u7d50\u5408\u8d77\u4f86\u5206\u6790\u3002<\/li>\n<li>\u6a19\u8a18\u51fa\u6700\u7b26\u5408\u76ee\u6a19\u61c9\u7528\u7684\u904b\u4f5c\u5340\u9593\u3002<\/li>\n<\/ol>\n<p>\u9019\u5957\u6d41\u7a0b\u5920\u7c21\u55ae\uff0c\u4fbf\u65bc\u91cd\u8907\u57f7\u884c\uff1b\u540c\u6642\u4e5f\u5920\u56b4\u8b39\uff0c\u80fd\u5920\u7522\u51fa\u771f\u6b63\u6709\u5de5\u7a0b\u50f9\u503c\u7684\u8b49\u64da\u3002<\/p>\n<h2><strong>\u6700\u5f8c\u7684\u601d\u8003<\/strong><\/h2>\n<p><strong>LLM \u4f3a\u670d\u5668\u541e\u5410\u91cf<\/strong>\u4e26\u4e0d\u662f\u4e00\u500b\u884c\u92b7\u6578\u5b57\uff0c\u800c\u662f\u5728\u660e\u78ba\u5de5\u4f5c\u8ca0\u8f09\u4e0b\uff0cprefill\u3001decode\u3001\u6392\u968a\u3001\u4e26\u767c\u548c\u7cfb\u7d71\u908a\u754c\u5171\u540c\u4f5c\u7528\u5f8c\u7684\u7d50\u679c\u3002\u53ea\u8981\u4f60\u8a8d\u771f\u8861\u91cf\u9019\u4e9b\u52d5\u614b\u56e0\u7d20\uff0c\u5c31\u80fd\u5728\u67b6\u69cb\u3001\u5bb9\u91cf\u548c\u7f8e\u570b<strong>\u4f3a\u670d\u5668\u79df\u7528<\/strong>\u7b56\u7565\u4e0a\u505a\u51fa\u66f4\u660e\u667a\u7684\u6c7a\u7b56\u3002\u5982\u679c\u8df3\u904e\u9019\u4e00\u6b65\uff0c\u57fa\u6e96\u6e2c\u8a66\u5c31\u53ea\u5269\u4e0b\u8868\u6f14\u610f\u7fa9\u3002\u5c0d\u6280\u8853\u5718\u968a\u4f86\u8aaa\uff0c\u771f\u6b63\u7684\u76ee\u6a19\u4e0d\u662f\u8ffd\u9010\u6700\u97ff\u4eae\u7684\u6578\u5b57\uff0c\u800c\u662f\u627e\u5230\u4e00\u500b\u7a69\u5b9a\u904b\u4f5c\u9ede\uff1a\u5728\u90a3\u88e1\uff0c\u5ef6\u9072\u4ecd\u7136\u5408\u7406\uff0ctoken \u4e32\u6d41\u4fdd\u6301\u9806\u66a2\uff0c\u57fa\u790e\u8a2d\u65bd\u6548\u7387\u4e5f\u8db3\u5920\u9ad8\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5982\u679c\u4f60\u5728\u73fe\u4ee3\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u4e0a\u57f7\u884c\u63a8\u8ad6\u5de5\u4f5c\u8ca0\u8f09\uff0c\u90a3\u9ebcLLM\u4f3a\u670d\u5668\u541e\u5410\u91cf\u9019\u500b\u8a5e\u5c31\u4e0d\u8a72\u53ea\u662f\u4f9b\u61c9\u5546\u6295\u5f71\u7247\u4e0a\u7684\u8aaa\u6cd5\uff0c [&#8230;]<\/p>\n<p><a class=\"btn btn-secondary understrap-read-more-link\" href=\"https:\/\/www.varidata.com\/zh-tw\/blog-tc\/how-to-measure-real-llm-server-throughput\/\">Read More&#8230;<\/a><\/p>\n","protected":false},"author":12,"featured_media":32468,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[267],"tags":[],"class_list":["post-32471","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-blog-tc"],"acf":[],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf<\/title>\n<meta name=\"description\" content=\"\u4e86\u89e3\u5982\u4f55\u900f\u904e\u5be6\u7528\u65b9\u6cd5\u5c0d LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\u9032\u884c\u57fa\u6e96\u6e2c\u8a66\uff0c\u6db5\u84cb TTFT\u3001\u5ef6\u9072\u3001\u4e26\u767c\u8207 token \u4e32\u6d41\u8868\u73fe\uff0c\u9069\u7528\u65bc\u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u3002\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\" \/>\n<meta property=\"og:locale\" content=\"zh_TW\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\" \/>\n<meta property=\"og:site_name\" content=\"Varidata Limited\" \/>\n<meta property=\"article:published_time\" content=\"2026-05-16T00:00:34+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"615\" \/>\n\t<meta property=\"og:image:height\" content=\"346\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#article\",\"isPartOf\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\"},\"author\":\"Varidata\",\"headline\":\"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\",\"datePublished\":\"2026-05-16T00:00:34+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\"},\"wordCount\":93,\"publisher\":{\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#organization\"},\"image\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg\",\"articleSection\":[\"Varidata \u5b98\u65b9\u535a\u5ba2\"],\"inLanguage\":\"zh-TW\"},{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\",\"url\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\",\"name\":\"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\",\"isPartOf\":{\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage\"},\"image\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage\"},\"thumbnailUrl\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg\",\"datePublished\":\"2026-05-16T00:00:34+00:00\",\"description\":\"\u4e86\u89e3\u5982\u4f55\u900f\u904e\u5be6\u7528\u65b9\u6cd5\u5c0d LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\u9032\u884c\u57fa\u6e96\u6e2c\u8a66\uff0c\u6db5\u84cb TTFT\u3001\u5ef6\u9072\u3001\u4e26\u767c\u8207 token \u4e32\u6d41\u8868\u73fe\uff0c\u9069\u7528\u65bc\u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u3002\",\"breadcrumb\":{\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#breadcrumb\"},\"inLanguage\":\"zh-TW\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-TW\",\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage\",\"url\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg\",\"contentUrl\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg\",\"width\":615,\"height\":346,\"caption\":\"\u5c55\u793a LLM \u4f3a\u670d\u5668\u541e\u5410\u91cf\u6e2c\u8a66\u7684\u793a\u610f\u5716\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\/\/www.varidata.com\/zh-tw\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#website\",\"url\":\"https:\/\/www.varidata.com\/zh-tw\/\",\"name\":\"Varidata Limited\",\"description\":\"\",\"publisher\":{\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.varidata.com\/zh-tw\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-TW\"},{\"@type\":\"Organization\",\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#organization\",\"name\":\"Varidata\",\"url\":\"https:\/\/www.varidata.com\/zh-tw\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-TW\",\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/logo\/image\/\",\"url\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2021\/09\/varidata_logo_white_-748x480_hor_web-1.png\",\"contentUrl\":\"https:\/\/www.varidata.com\/wp-content\/uploads\/2021\/09\/varidata_logo_white_-748x480_hor_web-1.png\",\"width\":248,\"height\":94,\"caption\":\"Varidata\"},\"image\":{\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/logo\/image\/\"}},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/person\/8fe12b0dd65fd2ba7ccf0dea7d14de6c\",\"name\":\"Amy Yan\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-TW\",\"@id\":\"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/secure.gravatar.com\/avatar\/087e2e5704ab9b17ce8a5ebfda282b5f9d46ea62ff6c00f7b19ef4d6ce4e83aa?s=96&d=mm&r=g\",\"contentUrl\":\"https:\/\/secure.gravatar.com\/avatar\/087e2e5704ab9b17ce8a5ebfda282b5f9d46ea62ff6c00f7b19ef4d6ce4e83aa?s=96&d=mm&r=g\",\"caption\":\"Amy Yan\"}}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf","description":"\u4e86\u89e3\u5982\u4f55\u900f\u904e\u5be6\u7528\u65b9\u6cd5\u5c0d LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\u9032\u884c\u57fa\u6e96\u6e2c\u8a66\uff0c\u6db5\u84cb TTFT\u3001\u5ef6\u9072\u3001\u4e26\u767c\u8207 token \u4e32\u6d41\u8868\u73fe\uff0c\u9069\u7528\u65bc\u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u3002","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/","og_locale":"zh_TW","og_type":"article","og_title":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf","og_url":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/","og_site_name":"Varidata Limited","article_published_time":"2026-05-16T00:00:34+00:00","og_image":[{"width":615,"height":346,"url":"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg","type":"image\/jpeg"}],"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#article","isPartOf":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/"},"author":"Varidata","headline":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf","datePublished":"2026-05-16T00:00:34+00:00","mainEntityOfPage":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/"},"wordCount":93,"publisher":{"@id":"https:\/\/www.varidata.com\/zh-tw\/#organization"},"image":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage"},"thumbnailUrl":"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg","articleSection":["Varidata \u5b98\u65b9\u535a\u5ba2"],"inLanguage":"zh-TW"},{"@type":"WebPage","@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/","url":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/","name":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf","isPartOf":{"@id":"https:\/\/www.varidata.com\/zh-tw\/#website"},"primaryImageOfPage":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage"},"image":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage"},"thumbnailUrl":"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg","datePublished":"2026-05-16T00:00:34+00:00","description":"\u4e86\u89e3\u5982\u4f55\u900f\u904e\u5be6\u7528\u65b9\u6cd5\u5c0d LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf\u9032\u884c\u57fa\u6e96\u6e2c\u8a66\uff0c\u6db5\u84cb TTFT\u3001\u5ef6\u9072\u3001\u4e26\u767c\u8207 token \u4e32\u6d41\u8868\u73fe\uff0c\u9069\u7528\u65bc\u7f8e\u570b\u4f3a\u670d\u5668\u79df\u7528\u57fa\u790e\u8a2d\u65bd\u3002","breadcrumb":{"@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#breadcrumb"},"inLanguage":"zh-TW","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/"]}]},{"@type":"ImageObject","inLanguage":"zh-TW","@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#primaryimage","url":"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg","contentUrl":"https:\/\/www.varidata.com\/wp-content\/uploads\/2026\/05\/Screenshot-2026-05-15-150435.jpg","width":615,"height":346,"caption":"\u5c55\u793a LLM \u4f3a\u670d\u5668\u541e\u5410\u91cf\u6e2c\u8a66\u7684\u793a\u610f\u5716"},{"@type":"BreadcrumbList","@id":"https:\/\/www.varidata.com\/blog-en\/how-to-measure-real-llm-server-throughput\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/www.varidata.com\/zh-tw\/"},{"@type":"ListItem","position":2,"name":"\u5982\u4f55\u8861\u91cf LLM \u4f3a\u670d\u5668\u7684\u771f\u5be6\u541e\u5410\u91cf"}]},{"@type":"WebSite","@id":"https:\/\/www.varidata.com\/zh-tw\/#website","url":"https:\/\/www.varidata.com\/zh-tw\/","name":"Varidata Limited","description":"","publisher":{"@id":"https:\/\/www.varidata.com\/zh-tw\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.varidata.com\/zh-tw\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-TW"},{"@type":"Organization","@id":"https:\/\/www.varidata.com\/zh-tw\/#organization","name":"Varidata","url":"https:\/\/www.varidata.com\/zh-tw\/","logo":{"@type":"ImageObject","inLanguage":"zh-TW","@id":"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/logo\/image\/","url":"https:\/\/www.varidata.com\/wp-content\/uploads\/2021\/09\/varidata_logo_white_-748x480_hor_web-1.png","contentUrl":"https:\/\/www.varidata.com\/wp-content\/uploads\/2021\/09\/varidata_logo_white_-748x480_hor_web-1.png","width":248,"height":94,"caption":"Varidata"},"image":{"@id":"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/logo\/image\/"}},{"@type":"Person","@id":"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/person\/8fe12b0dd65fd2ba7ccf0dea7d14de6c","name":"Amy Yan","image":{"@type":"ImageObject","inLanguage":"zh-TW","@id":"https:\/\/www.varidata.com\/zh-tw\/#\/schema\/person\/image\/","url":"https:\/\/secure.gravatar.com\/avatar\/087e2e5704ab9b17ce8a5ebfda282b5f9d46ea62ff6c00f7b19ef4d6ce4e83aa?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/087e2e5704ab9b17ce8a5ebfda282b5f9d46ea62ff6c00f7b19ef4d6ce4e83aa?s=96&d=mm&r=g","caption":"Amy Yan"}}]}},"_links":{"self":[{"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/posts\/32471","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/users\/12"}],"replies":[{"embeddable":true,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/comments?post=32471"}],"version-history":[{"count":3,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/posts\/32471\/revisions"}],"predecessor-version":[{"id":32477,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/posts\/32471\/revisions\/32477"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/media\/32468"}],"wp:attachment":[{"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/media?parent=32471"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/categories?post=32471"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.varidata.com\/zh-tw\/wp-json\/wp\/v2\/tags?post=32471"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}