{"id":839,"date":"2026-07-12T01:15:20","date_gmt":"2026-07-11T17:15:20","guid":{"rendered":"https:\/\/blog.itsmygo.uk\/2026\/07\/12\/%e3%80%90vllm-%e5%b7%a5%e7%a8%8b%e5%b8%ab%e5%af%a6%e6%88%b0%e3%80%9103-%e6%a6%a8%e4%b9%be%e5%90%9e%e5%90%90%e9%87%8f%ef%bc%9a%e9%80%a3%e7%ba%8c%e6%89%b9%e6%ac%a1%e3%80%81%e4%b8%a6%e8%a1%8c%e8%88%87\/"},"modified":"2026-07-12T01:15:20","modified_gmt":"2026-07-11T17:15:20","slug":"%e3%80%90vllm-%e5%b7%a5%e7%a8%8b%e5%b8%ab%e5%af%a6%e6%88%b0%e3%80%9103-%e6%a6%a8%e4%b9%be%e5%90%9e%e5%90%90%e9%87%8f%ef%bc%9a%e9%80%a3%e7%ba%8c%e6%89%b9%e6%ac%a1%e3%80%81%e4%b8%a6%e8%a1%8c%e8%88%87","status":"publish","type":"post","link":"https:\/\/blog.itsmygo.uk\/wordpress\/2026\/07\/12\/%e3%80%90vllm-%e5%b7%a5%e7%a8%8b%e5%b8%ab%e5%af%a6%e6%88%b0%e3%80%9103-%e6%a6%a8%e4%b9%be%e5%90%9e%e5%90%90%e9%87%8f%ef%bc%9a%e9%80%a3%e7%ba%8c%e6%89%b9%e6%ac%a1%e3%80%81%e4%b8%a6%e8%a1%8c%e8%88%87\/","title":{"rendered":"\u3010vLLM \u5de5\u7a0b\u5e2b\u5be6\u6230\u3011#03 \u69a8\u4e7e\u541e\u5410\u91cf\uff1a\u9023\u7e8c\u6279\u6b21\u3001\u4e26\u884c\u8207\u91cf\u5316\u7684\u8abf\u6821\u624b\u518a"},"content":{"rendered":"\n<div id=\"quiz-55fa3b4f\">\n<style>\n#quiz-55fa3b4f .question { background: white; border-radius: 8px; padding: 20px; margin: 20px 0; box-shadow: 0 2px 4px rgba(0,0,0,0.1); }#quiz-55fa3b4f .question h3 { margin-top: 0; color: #333; }#quiz-55fa3b4f .options { list-style: none; padding: 0; }#quiz-55fa3b4f .options li { padding: 12px 15px; margin: 8px 0; border: 2px solid #ddd; border-radius: 6px; cursor: pointer; transition: all 0.2s; }#quiz-55fa3b4f .options li:hover:not(.disabled) { border-color: #4caf50; background: #f0fff0; }#quiz-55fa3b4f .options li.correct { border-color: #4caf50; background: #e8f5e9; }#quiz-55fa3b4f .options li.wrong { border-color: #f44336; background: #ffebee; }#quiz-55fa3b4f .options li.disabled { cursor: default; opacity: 0.7; }#quiz-55fa3b4f .options li.show-answer { border-color: #4caf50; background: #e8f5e9; }#quiz-55fa3b4f .feedback { margin-top: 15px; padding: 12px; border-radius: 6px; display: none; }#quiz-55fa3b4f .feedback.correct { background: #e8f5e9; color: #2e7d32; display: block; }#quiz-55fa3b4f .feedback.wrong { background: #ffebee; color: #c62828; display: block; }#quiz-55fa3b4f .score { text-align: center; font-size: 1.5em; padding: 20px; background: white; border-radius: 8px; margin-top: 20px; }#quiz-55fa3b4f .code-block { background: #1e1e1e; color: #d4d4d4; padding: 14px; border-radius: 6px; font-family: \"Monaco\",\"Consolas\",\"Courier New\",monospace; font-size: 0.9em; overflow-x: auto; margin: 12px 0; white-space: pre-wrap; }\n<\/style>\n<h1>\u6e2c\u9a57\uff1avLLM \u69a8\u4e7e\u541e\u5410\u91cf<\/h1>\n    <p>\u5171 5 \u984c\uff0c\u9ede\u9078\u7b54\u6848\u5f8c\u6703\u7acb\u5373\u986f\u793a\u7d50\u679c<\/p>\n\n    <div class=\"question\" id=\"q1\">\n      <h3>1. \u95dc\u65bc continuous batching\uff08\u9023\u7e8c\u6279\u6b21\uff09\uff0c\u6587\u7ae0\u7684\u767d\u8a71\u6bd4\u55bb\u662f\u4ec0\u9ebc\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(1, 'A', 'B')\">A. \u50cf\u516c\u8eca\uff0c\u8981\u7b49\u5ea7\u4f4d\u5750\u6eff\u624d\u767c\u8eca<\/li>\n        <li onclick=\"checkAnswer(1, 'B', 'B')\">B. \u50cf\u5171\u4e58\u8a08\u7a0b\u8eca\uff0c\u8eca\u5728\u8def\u4e0a\u8dd1\u3001\u6709\u4eba\u5230\u5c31\u4f75\u9032\u4f86\u3001\u6709\u4eba\u5230\u7ad9\u5c31\u653e\u4e0b<\/li>\n        <li onclick=\"checkAnswer(1, 'C', 'B')\">C. \u9700\u8981\u5de5\u7a0b\u5e2b\u624b\u52d5\u5beb code \u89f8\u767c\u5b83\u624d\u6703\u904b\u4f5c<\/li>\n        <li onclick=\"checkAnswer(1, 'D', 'B')\">D. \u53ea\u6709\u91cf\u5316\u6a21\u578b\u624d\u652f\u63f4\u7684\u7279\u6b8a\u529f\u80fd<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb1\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q2\">\n      <h3>2. <code>--tensor-parallel-size 2<\/code> \u9019\u500b\u53c3\u6578\u5728\u505a\u4ec0\u9ebc\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(2, 'A', 'C')\">A. \u4e00\u6b21\u6700\u591a\u8655\u7406 2 \u500b\u8acb\u6c42<\/li>\n        <li onclick=\"checkAnswer(2, 'B', 'C')\">B. \u628a\u6a21\u578b\u91cf\u5316\u6210 2-bit<\/li>\n        <li onclick=\"checkAnswer(2, 'C', 'C')\">C. \u628a\u4e00\u500b\u6a21\u578b\u5207\u958b\u6524\u5230 2 \u5f35 GPU \u4e0a\u4e00\u8d77\u7b97<\/li>\n        <li onclick=\"checkAnswer(2, 'D', 'C')\">D. \u9650\u5236\u4e00\u6279\u6700\u591a 2 \u500b token<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb2\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q3\">\n      <h3>3. \u4f60\u89c0\u5bdf\u5230\u300c\u541e\u5410\u91cf\u504f\u4f4e\uff0c\u4f46 GPU \u4f7f\u7528\u7387\u4e00\u76f4\u6c92\u6eff\u300d\uff0c\u6700\u53ef\u80fd\u8a72\u8abf\u54ea\u500b\u53c3\u6578\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(3, 'A', 'C')\">A. \u8abf\u4f4e <code>--tensor-parallel-size<\/code><\/li>\n        <li onclick=\"checkAnswer(3, 'B', 'C')\">B. \u63db\u4e00\u500b\u7cbe\u5ea6\u66f4\u9ad8\u7684 <code>--dtype<\/code><\/li>\n        <li onclick=\"checkAnswer(3, 'C', 'C')\">C. \u8abf\u9ad8 <code>--max-num-batched-tokens<\/code>\uff0c\u628a\u6279\u6b21\u5582\u98fd<\/li>\n        <li onclick=\"checkAnswer(3, 'D', 'C')\">D. \u62ff\u6389 <code>--quantization<\/code><\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb3\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q4\">\n      <h3>4. \u95dc\u65bc <code>--quantization awq<\/code>\uff0c\u6587\u7ae0\u5f37\u8abf\u7684\u95dc\u9375\u524d\u63d0\u662f\u4ec0\u9ebc\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(4, 'A', 'B')\">A. \u5b83\u6703\u5728\u8f09\u5165\u6642\u5373\u6642\u628a\u539f\u7248\u6a21\u578b\u58d3\u6210 4-bit<\/li>\n        <li onclick=\"checkAnswer(4, 'B', 'B')\">B. \u5b83\u662f\u7528 AWQ \u65b9\u5f0f\u8b80\u300c\u672c\u8eab\u5df2\u662f AWQ \u91cf\u5316\u7248\u300d\u7684\u6a21\u578b\uff0c\u8f09\u5165\u7684\u6a21\u578b\u540d\u7a31\u5fc5\u9808\u662f\u91cf\u5316\u7248<\/li>\n        <li onclick=\"checkAnswer(4, 'C', 'B')\">C. \u53ea\u8981\u986f\u5b58\u5920\uff0c\u4efb\u4f55\u6a21\u578b\u90fd\u8a72\u52a0\u4e0a\u5b83<\/li>\n        <li onclick=\"checkAnswer(4, 'D', 'B')\">D. \u5b83\u6703\u81ea\u52d5\u63d0\u9ad8\u6a21\u578b\u7cbe\u5ea6<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb4\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q5\">\n      <h3>5. \u58d3\u6e2c\u5831\u544a\u88e1\u7684 TTFT\uff08Time To First Token\uff09\u504f\u9ad8\uff0c\u4ee3\u8868\u4f7f\u7528\u8005\u611f\u53d7\u5230\u7684\u662f\u4ec0\u9ebc\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(5, 'A', 'A')\">A. \u76ef\u8457\u7a7a\u767d\u756b\u9762\u4e7e\u7b49\uff0c\u7b2c\u4e00\u500b\u5b57\u9072\u9072\u4e0d\u51fa\u4f86<\/li>\n        <li onclick=\"checkAnswer(5, 'B', 'A')\">B. \u5b57\u4e00\u500b\u4e00\u500b\u6162\u6162\u8e66\u3001\u6253\u5b57\u4e0d\u6d41\u66a2<\/li>\n        <li onclick=\"checkAnswer(5, 'C', 'A')\">C. \u6574\u53f0\u6a5f\u5668\u6bcf\u79d2\u7e3d\u7522\u51fa\u8b8a\u4f4e<\/li>\n        <li onclick=\"checkAnswer(5, 'D', 'A')\">D. \u6a21\u578b\u56de\u7b54\u7684\u7cbe\u5ea6\u4e0b\u964d<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb5\"><\/div>\n    <\/div>\n\n    <div class=\"score\" id=\"score\" style=\"display:none;\"><\/div>\n<script>\n\n      let answered = 0, correct = 0;\n      const total = 5;\n      const explanations = {\n        1: \"continuous batching \u50cf\u5171\u4e58\u8a08\u7a0b\u8eca\uff0c\u8acb\u6c42\u96a8\u5230\u96a8\u4f75\u5165\u6b63\u5728\u8dd1\u7684\u6279\u6b21\uff0c\u4e0d\u7528\u6e4a\u6eff\u4e00\u6279\u624d\u767c\u8eca\u3002\u5b83\u662f\u81ea\u52d5\u767c\u751f\u7684\uff0c\u4e0d\u9700\u5beb code \u89f8\u767c\uff0c\u4e5f\u4e0d\u9650\u91cf\u5316\u6a21\u578b\u3002\",\n        2: \"tensor-parallel-size\uff08TP\uff09\u662f\u628a\u300c\u4e00\u500b\u6a21\u578b\u300d\u62c6\u958b\u6524\u5230\u5e7e\u5f35 GPU \u4e0a\u4e00\u8d77\u7b97\uff0c\u7528\u4f86\u8dd1\u55ae\u5361\u585e\u4e0d\u4e0b\u7684\u5927\u6a21\u578b\u3002\u5b83\u8ddf\u8acb\u6c42\u6578\u3001token \u6578\u3001\u91cf\u5316\u90fd\u7121\u95dc\u3002\",\n        3: \"GPU \u4f7f\u7528\u7387\u6c92\u6eff\u4ee3\u8868\u6279\u6b21\u6c92\u5582\u98fd\u3002\u9577 prompt \u9032\u4f86\u6642\u5e38\u88ab max-num-batched-tokens \u7684\u4e0a\u9650\u5361\u4f4f\uff0c\u8abf\u5927\u5b83\u8b93\u6bcf\u500b step \u8655\u7406\u66f4\u591a token\uff0c\u541e\u5410\u624d\u62c9\u5f97\u4e0a\u53bb\u3002\",\n        4: \"--quantization awq \u4e0d\u662f\u5373\u6642\u58d3\u7e2e\uff0c\u800c\u662f\u544a\u8a34 vLLM\u300c\u9019\u500b\u6a21\u578b\u672c\u8eab\u5c31\u662f AWQ \u7248\u3001\u8acb\u7528 AWQ \u8b80\u300d\u3002\u6240\u4ee5\u6a21\u578b\u540d\u7a31\u5fc5\u9808\u662f\u91cf\u5316\u7248\uff08\u5982 -AWQ \u7d50\u5c3e\uff09\uff0c\u5426\u5247\u8f09\u5165\u5931\u6557\u3002\",\n        5: \"TTFT \u662f\u7b2c\u4e00\u500b token \u5410\u51fa\u4f86\u8981\u591a\u4e45\uff0c\u9ad8\u5c31\u4ee3\u8868\u4f7f\u7528\u8005\u76ef\u8457\u7a7a\u767d\u756b\u9762\u7b49\u5f85\u3002\u5b57\u6162\u6162\u8e66\u662f TPOT\uff0c\u6574\u9ad4\u7522\u51fa\u662f throughput\uff0c\u90fd\u4e0d\u662f TTFT\u3002\"\n      };\n      function checkAnswer(qNum, selected, correctAns) {\n        const question = document.getElementById(\"q\" + qNum);\n        if (question.classList.contains(\"answered\")) return;\n        question.classList.add(\"answered\");\n        answered++;\n        const options = question.querySelectorAll(\".options li\");\n        options.forEach((opt) => {\n          opt.classList.add(\"disabled\");\n          if (opt.textContent.startsWith(correctAns + \".\")) opt.classList.add(\"show-answer\");\n        });\n        const feedback = document.getElementById(\"fb\" + qNum);\n        if (selected === correctAns) {\n          correct++;\n          feedback.className = \"feedback correct\";\n          feedback.innerHTML = \"\u2713 \u6b63\u78ba\uff01\" + explanations[qNum];\n        } else {\n          feedback.className = \"feedback wrong\";\n          feedback.innerHTML = \"\u2717 \u932f\u8aa4\u3002\u6b63\u78ba\u7b54\u6848\u662f \" + correctAns + \"\u3002\" + explanations[qNum];\n          options.forEach((opt) => { if (opt.textContent.startsWith(selected + \".\")) opt.classList.add(\"wrong\"); });\n        }\n        if (answered === total) {\n          const scoreDiv = document.getElementById(\"score\");\n          scoreDiv.style.display = \"block\";\n          scoreDiv.innerHTML = \"\u6e2c\u9a57\u5b8c\u6210\uff01\u5f97\u5206\uff1a\" + correct + \" \/ \" + total +\n            (correct === total ? \" \ud83c\udf89 \u5168\u5c0d\uff01\" : correct >= Math.ceil(total * 0.6) ? \" \ud83d\udc4d \u4e0d\u932f\uff01\" : \" \ud83d\udcda \u518d\u8907\u7fd2\u4e00\u4e0b\u5427\");\n        }\n      }\n    \n<\/script>\n<\/div>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u7cfb\u5217\u7b2c 3 \u7bc7\uff0c\u5171 4 \u7bc7 \uff5c \u96e3\u5ea6\uff1aL2-\u9032\u968e<br>\u524d\u7f6e\u77e5\u8b58\uff1a\u5df2\u8b80\u7b2c 1\u30012 \u7bc7\uff0c\u7406\u89e3 <code>vllm serve<\/code> \u600e\u9ebc\u628a\u6a21\u578b\u8dd1\u8d77\u4f86\u3001KV Cache \u8207 <code>--gpu-memory-utilization<\/code> \u9019\u4e9b\u8a18\u61b6\u9ad4\u53c3\u6578\u3002<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u9019\u7bc7\u8981\u89e3\u6c7a\u7684\u554f\u984c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c 1\u30012 \u7bc7\u4f60\u5df2\u7d93\u80fd\u628a\u6a21\u578b serve \u8d77\u4f86\u3001\u4e5f\u77e5\u9053 KV Cache \u6703\u5403\u6389\u591a\u5c11\u986f\u5b58\u3002\u4f46\u771f\u6b63\u4e0a\u7dda\u5f8c\uff0c\u8001\u95c6\u554f\u7684\u662f\u53e6\u4e00\u4ef6\u4e8b\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u540c\u6642 50 \u500b\u4eba\u6253\u904e\u4f86\uff0c\u6703\u4e0d\u6703\u5361\uff1f\u4e00\u5f35\u5361\u5230\u5e95\u80fd\u625b\u591a\u5c11 QPS\uff1f\u8dd1\u4e0d\u52d5\u4e86\u8981\u52a0\u5361\u9084\u662f\u63db\u91cf\u5316\uff1f\u300d<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u7bc7\u5c31\u662f\u6559\u4f60\u600e\u9ebc\u300c\u770b\u61c2\u541e\u5410\u91cf\u7684\u65cb\u9215\u300d\u3002\u91cd\u9ede\u4e0d\u662f\u8981\u4f60\u624b\u523b\u63a8\u8ad6\u5f15\u64ce\u2014\u2014continuous batching \u9019\u4e9b\u6838\u5fc3\u6a5f\u5236 vLLM \u81ea\u52d5\u5e6b\u4f60\u505a\u597d\u4e86\u2014\u2014\u800c\u662f\u8b93\u4f60<strong>\u770b\u61c2 AI \u6216\u540c\u4e8b\u4e1f\u7d66\u4f60\u7684\u555f\u52d5\u6307\u4ee4\u6709\u6c92\u6709\u554f\u984c\u3001\u58d3\u6e2c\u6578\u5b57\u5f80\u54ea\u500b\u65b9\u5411\u8abf<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u662f\u9a57\u6536\u8005\uff1aAI \u5f88\u6703\u751f\u4e00\u9577\u4e32 <code>vllm serve<\/code> \u53c3\u6578\uff0c\u4f46\u5b83\u5e38\u5e38\u628a <code>--tensor-parallel-size<\/code> \u8a2d\u932f\u3001\u6216\u91cf\u5316\u6a21\u578b\u540d\u7a31\u5f35\u51a0\u674e\u6234\u3002\u4f60\u8981\u80fd\u4e00\u773c\u8a8d\u51fa\u7d05\u65d7\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e00\u53e5\u8a71\u8aaa\u660e\uff1acontinuous batching<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u767d\u8a71\uff1a<\/strong> \u8acb\u6c42\u96a8\u5230\u96a8\u4f75\u9032\u6b63\u5728\u8dd1\u7684\u90a3\u6279\uff0c\u4e0d\u7528\u7b49\u6e4a\u6eff\u4e00\u8eca\u624d\u767c\u8eca\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u50b3\u7d71 batching \u50cf\u516c\u8eca\u2014\u2014\u8981\u7b49\u5ea7\u4f4d\u5750\u6eff\uff08\u6216\u7b49\u6642\u9593\u5230\uff09\u624d\u958b\uff0c\u5148\u5230\u7684\u4eba\u5f97\u4e7e\u7b49\u3002continuous batching\uff08\u9023\u7e8c\u6279\u6b21\uff0cvLLM \u53c8\u53eb in-flight batching\uff09\u50cf\u5171\u4e58\u8a08\u7a0b\u8eca\u2014\u2014\u8eca\u5728\u8def\u4e0a\u8dd1\uff0c\u6709\u4eba\u8981\u4e0a\u8eca\u5c31\u5728\u4e0b\u4e00\u500b\u8def\u53e3\u4f75\u9032\u4f86\uff0c\u6709\u4eba\u5230\u7ad9\u5c31\u653e\u4e0b\uff0c\u8eca\u4e00\u76f4\u662f\u6eff\u8f09\u72c0\u614b\u3002<\/p>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-1\" data-shcb-language-name=\"CSS\" data-shcb-language-slug=\"css\"><span><code class=\"hljs language-css\">\u50b3\u7d71 <span class=\"hljs-selector-tag\">batch<\/span>\uff1a\n  <span class=\"hljs-selector-attr\">&#91;req1 req2]<\/span> \u4e00\u8d77\u8dd1\u5b8c \u2192 \u624d\u6536 <span class=\"hljs-selector-attr\">&#91;req3 req4]<\/span>\n  <span class=\"hljs-selector-tag\">req3<\/span> \u5230\u5f97\u65e9\u4e5f\u5f97\u7b49 <span class=\"hljs-selector-tag\">req1<\/span>\u3001<span class=\"hljs-selector-tag\">req2<\/span> \u5168\u90e8\u751f\u6210\u5b8c\n\n<span class=\"hljs-selector-tag\">continuous<\/span> <span class=\"hljs-selector-tag\">batch<\/span>\uff08<span class=\"hljs-selector-tag\">vLLM<\/span> \u9810\u8a2d\uff09\uff1a\n  <span class=\"hljs-selector-tag\">req1<\/span> \u751f\u6210\u4e2d ... <span class=\"hljs-selector-tag\">req2<\/span> \u4f75\u5165 ... <span class=\"hljs-selector-tag\">req1<\/span> \u7d50\u675f\u3001<span class=\"hljs-selector-tag\">req3<\/span> \u88dc\u4f4d ...\n  \u6bcf\u500b <span class=\"hljs-selector-tag\">decode<\/span> <span class=\"hljs-selector-tag\">step<\/span> \u91cd\u65b0\u7d44\u6279\uff0c<span class=\"hljs-selector-tag\">GPU<\/span> \u6c38\u9060\u4e0d\u9592\u7f6e\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-1\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">CSS<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">css<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4f60\u8981\u77e5\u9053\u7684\u53ea\u6709\u4e00\u4ef6\u4e8b\uff1a\u9019\u662f\u81ea\u52d5\u767c\u751f\u7684\uff0cvLLM \u9810\u8a2d\u5c31\u958b\u8457\u3002<\/strong> \u4f60\u4e0d\u7528\u5beb\u4efb\u4f55 code \u53bb\u89f8\u767c\u5b83\u3002\u4f60\u8981\u505a\u7684\u662f\u9935\u7d66\u5b83\u5c0d\u7684\u53c3\u6578\uff0c\u8b93\u300c\u9019\u53f0\u5171\u4e58\u8eca\u300d\u80fd\u8f09\u66f4\u591a\u4eba\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea6 \u70ba\u4ec0\u9ebc\u5de5\u7a0b\u5e2b\u8981\u5728\u4e4e\u5b83<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u70ba\u5b83\u6c7a\u5b9a\u4e86\u5169\u500b\u6b64\u6d88\u5f7c\u9577\u7684\u6307\u6a19\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>Throughput\uff08\u541e\u5410\u91cf\uff09<\/strong>\uff1a\u6574\u53f0\u4f3a\u670d\u5668\u6bcf\u79d2\u80fd\u5410\u591a\u5c11 token \/ \u670d\u52d9\u591a\u5c11\u8acb\u6c42\u3002\u6279\u6b21\u8d8a\u5927\u8d8a\u9ad8\u3002<\/li>\n\n\n<li><strong>Latency\uff08\u5ef6\u9072\uff09<\/strong>\uff1a\u55ae\u4e00\u4f7f\u7528\u8005\u611f\u53d7\u5230\u7684\u5feb\u6162\u3002\u6279\u6b21\u592a\u5927\u6642\uff0c\u4f60\u7684\u8acb\u6c42\u8981\u8ddf\u4e00\u5806\u4eba\u6436 GPU\uff0c\u53cd\u800c\u8b8a\u6162\u3002<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">continuous batching \u8b93\u4f60\u5728<strong>\u9ad8\u541e\u5410<\/strong>\u7684\u540c\u6642\u76e1\u91cf\u4e0d\u72a7\u7272\u5ef6\u9072\uff0c\u4f46\u5929\u4e0b\u6c92\u6709\u767d\u5403\u7684\u5348\u9910\u2014\u2014\u6279\u6b21\u585e\u8d8a\u6eff\uff0c\u500b\u5225\u8acb\u6c42\u7684\u5ef6\u9072\u5c31\u8d8a\u5bb9\u6613\u88ab\u62c9\u9ad8\u3002\u8abf\u6821\u5c31\u662f\u5728\u9019\u689d\u7dda\u4e0a\u627e\u4f60\u5834\u666f\u8981\u7684\u5e73\u8861\u9ede\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u95dc\u9375\u65cb\u9215\uff1a\u4e94\u500b\u4f60\u6703\u53cd\u8986\u770b\u5230\u7684\u53c3\u6578<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">AI \u751f\u6210\u7684\u555f\u52d5\u6307\u4ee4\u901a\u5e38\u9577\u9019\u6a23\uff08\u6700\u5c0f\u53ef\u8b80\u7248\u672c\uff09\uff1a<\/p>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\"><span><code class=\"hljs\">vllm serve Qwen\/Qwen2.5-7B-Instruct \\\n  --tensor-parallel-size 2 \\\n  --max-num-seqs 256 \\\n  --max-num-batched-tokens 8192 \\\n  --quantization awq \\\n  --dtype auto\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u884c\u4e00\u884c\u7ffb\u8b6f\u6210\u767d\u8a71\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead>\n<tr>\n<th>\u53c3\u6578<\/th>\n<th>\u9019\u5728\u5e79\u561b<\/th>\n<th>\u8abf\u5927\u6703\u600e\u6a23<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>--tensor-parallel-size<\/code>\uff08TP\uff09<\/td>\n<td>\u628a\u4e00\u500b\u6a21\u578b\u300c\u5207\u958b\u300d\u6524\u5230\u5e7e\u5f35 GPU \u4e0a<\/td>\n<td>\u80fd\u8dd1\u66f4\u5927\u7684\u6a21\u578b\uff0c\u4f46\u5361\u9593\u901a\u8a0a\u6709\u958b\u92b7<\/td>\n<\/tr>\n<tr>\n<td><code>--max-num-seqs<\/code><\/td>\n<td>\u4e00\u6279\u6700\u591a\u540c\u6642\u8655\u7406\u5e7e\u500b\u8acb\u6c42\uff08\u4f75\u767c\u6578\u4e0a\u9650\uff09<\/td>\n<td>\u541e\u5410\u2191\uff0c\u4f46\u5ef6\u9072\u53ef\u80fd\u2191\u3001\u986f\u5b58\u2191<\/td>\n<\/tr>\n<tr>\n<td><code>--max-num-batched-tokens<\/code><\/td>\n<td>\u4e00\u500b step \u6700\u591a\u8655\u7406\u5e7e\u500b token\uff08batch \u7684\u300c\u7e3d\u5bb9\u91cf\u300d\uff09<\/td>\n<td>\u541e\u5410\u2191\uff0c\u8a2d\u592a\u5c0f\u6703\u5361\u4f4f\u541e\u5410<\/td>\n<\/tr>\n<tr>\n<td><code>--quantization<\/code><\/td>\n<td>\u7528\u54ea\u7a2e\u91cf\u5316\u683c\u5f0f\u58d3\u7e2e\u6b0a\u91cd\uff08awq\/gptq\/fp8\u2026\uff09<\/td>\n<td>\u7701\u986f\u5b58\u3001\u53ef\u63d0\u4f75\u767c\uff0c\u4f46\u7cbe\u5ea6\u7565\u964d<\/td>\n<\/tr>\n<tr>\n<td><code>--dtype<\/code><\/td>\n<td>\u6b0a\u91cd\u7684\u904b\u7b97\u7cbe\u5ea6\uff08auto\/float16\/bfloat16\uff09<\/td>\n<td>\u901a\u5e38\u653e <code>auto<\/code> \u5c31\u597d<\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--tensor-parallel-size<\/code>\uff08TP\uff09\uff1a\u628a\u6a21\u578b\u6524\u5230\u591a\u5f35\u5361<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-2\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\">--tensor-parallel-size <span class=\"hljs-number\">2<\/span>   <span class=\"hljs-comment\"># \u9019\u9846\u6a21\u578b\u62c6\u6210 2 \u4efd\uff0c\u7528 2 \u5f35 GPU \u4e00\u8d77\u7b97<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-2\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u767d\u8a71\uff1a<\/strong> \u4e00\u9846 70B \u7684\u6a21\u578b\u55ae\u5f35\u5361\u585e\u4e0d\u4e0b\uff0c\u5c31\u628a\u5b83\u300c\u6a6b\u8457\u5207\u958b\u300d\u5206\u7d66 2 \u5f35\u30014 \u5f35\u5361\uff0c\u5927\u5bb6\u7b97\u81ea\u5df1\u90a3\u584a\u3001\u518d\u628a\u7d50\u679c\u5408\u8d77\u4f86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u2705 <strong>\u5fc5\u770b\u61c2<\/strong>\uff1aTP \u6578 = \u4f60\u8981\u7528\u5e7e\u5f35\u5361\u8dd1\u300c\u9019\u4e00\u500b\u300d\u6a21\u578b\u3002TP=1 \u5c31\u662f\u55ae\u5361\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\ud83d\udea9 <strong>\u7d05\u65d7\uff08\u5f8c\u9762\u8a73\u8ac7\uff09<\/strong>\uff1aTP \u6578\u5fc5\u9808\u80fd\u6574\u9664\u6a21\u578b\u7684 attention head \u6578\uff0c\u800c\u4e14\u901a\u5e38\u8a2d\u6210\u4f60<strong>\u5be6\u969b\u6709\u7684 GPU \u5f35\u6578<\/strong>\u3002AI \u5f88\u611b\u8907\u88fd\u5225\u4eba\u7684\u6307\u4ee4\uff0c\u628a <code>--tensor-parallel-size 4<\/code> \u8cbc\u5230\u4f60\u53ea\u6709 2 \u5f35\u5361\u7684\u6a5f\u5668\u4e0a\u2014\u2014\u555f\u52d5\u76f4\u63a5\u7206\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--max-num-seqs<\/code> vs <code>--max-num-batched-tokens<\/code>\uff1a\u5169\u500b\u5bb9\u91cf\u4e0a\u9650<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u5169\u500b\u6700\u5bb9\u6613\u641e\u6df7\uff0c\u7528\u4e00\u53e5\u8a71\u5206\u958b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><code>--max-num-seqs<\/code>\uff1a<strong>\u5e7e\u500b\u4eba<\/strong>\u80fd\u540c\u6642\u4e0a\u8eca\uff08\u8acb\u6c42\u6578\uff09\u3002<\/li>\n\n\n<li><code>--max-num-batched-tokens<\/code>\uff1a\u9019\u6279<strong>\u7e3d\u5171\u5e7e\u500b token<\/strong> \u7684\u8655\u7406\u91cf\uff08\u5de5\u4f5c\u91cf\uff09\u3002<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">vLLM \u6bcf\u500b step \u6703\u5728\u300c\u4e0d\u8d85\u904e\u9019\u5169\u500b\u4e0a\u9650\u300d\u7684\u524d\u63d0\u4e0b\uff0c\u76e1\u91cf\u628a\u5728\u6392\u968a\u7684\u8acb\u6c42\u585e\u9032\u6279\u6b21\u3002\u54ea\u500b\u5148\u649e\u5230\u4e0a\u9650\uff0c\u5c31\u7531\u54ea\u500b\u6c7a\u5b9a\u9019\u6279\u591a\u5927\u3002<\/p>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\"><span><code class=\"hljs\">\u60c5\u5883\uff1a\u4e00\u5806\u9577 prompt\uff08\u5404 4000 token\uff09\u6e67\u5165\n  max-num-batched-tokens = 8192 \u2192 \u4e00\u6279\u53ea\u585e\u5f97\u4e0b 2 \u500b\uff088192 \/ 4000 \u2248 2\uff09\n  \u5c31\u7b97 max-num-seqs = 256\uff0c\u4e5f\u8f2a\u4e0d\u5230\uff0ctoken \u4e0a\u9650\u5148\u5361\u4f4f\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u5982\u679c\u4f60\u770b\u5230<strong>\u541e\u5410\u4e0a\u4e0d\u53bb\u3001GPU \u4f7f\u7528\u7387\u537b\u6c92\u6eff<\/strong>\uff0c\u5f88\u591a\u6642\u5019\u662f <code>--max-num-batched-tokens<\/code> \u8a2d\u592a\u5c0f\uff0c\u6279\u6b21\u6839\u672c\u5582\u4e0d\u98fd GPU\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u91cf\u5316\u53d6\u6368\uff1a\u4ec0\u9ebc\u6642\u5019\u8a72\u58d3\u3001\u58d3\u4e86\u6703\u600e\u6a23<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cf\u5316\uff08quantization\uff09= \u628a\u6a21\u578b\u6b0a\u91cd\u5f9e 16-bit \u58d3\u6210 8-bit \u6216 4-bit\uff0c\u63db\u53d6\u66f4\u5c11\u986f\u5b58\u3001\u66f4\u9ad8\u4f75\u767c\u3002<\/p>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-3\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\">--quantization awq    <span class=\"hljs-comment\"># \u7528 AWQ\uff084-bit\uff09\u683c\u5f0f\u8f09\u5165<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-3\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u26a0\ufe0f <strong>\u8d85\u91cd\u8981\u524d\u63d0<\/strong>\uff1a<code>--quantization awq<\/code> \u4e0d\u662f\u300c\u5373\u6642\u5e6b\u4f60\u58d3\u300d\u3002\u5b83\u662f\u544a\u8a34 vLLM\u300c<strong>\u9019\u500b\u6a21\u578b\u6a94\u672c\u8eab\u5c31\u662f AWQ \u91cf\u5316\u904e\u7684\u7248\u672c<\/strong>\uff0c\u8acb\u7528 AWQ \u7684\u65b9\u5f0f\u8b80\u5b83\u300d\u3002\u4f60\u8f09\u5165\u7684\u6a21\u578b\u540d\u7a31\u5fc5\u9808\u662f\u91cf\u5316\u7248\uff0c\u4f8b\u5982 <code>Qwen\/Qwen2.5-7B-Instruct-AWQ<\/code>\uff0c\u4e0d\u662f\u539f\u7248\u3002\u9019\u662f\u65b0\u624b\u7b2c\u4e00\u865f\u5927\u5751\uff08\u7d05\u65d7\u5340\u7d30\u8b1b\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u53d6\u6368\u8868\uff1a\u7701\u4ec0\u9ebc\u3001\u8ce0\u4ec0\u9ebc<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead>\n<tr>\n<th>\u683c\u5f0f<\/th>\n<th>\u7cbe\u5ea6<\/th>\n<th>\u986f\u5b58<\/th>\n<th>\u9069\u5408\u5834\u666f<\/th>\n<th>\u4ee3\u50f9<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u4e0d\u91cf\u5316\uff08FP16\/BF16\uff09<\/td>\n<td>\u6700\u9ad8<\/td>\n<td>\u6700\u5403<\/td>\n<td>\u7cbe\u5ea6\u654f\u611f\u3001\u986f\u5b58\u5920<\/td>\n<td>\u4f75\u767c\u53d7\u9650\u65bc\u986f\u5b58<\/td>\n<\/tr>\n<tr>\n<td><strong>FP8<\/strong><\/td>\n<td>\u9ad8\uff08\u5e7e\u4e4e\u7121\u640d\uff09<\/td>\n<td>\u7701\u7d04\u4e00\u534a<\/td>\n<td>\u65b0\u5361\uff08H100\/Ada\uff09\u8ffd\u541e\u5410<\/td>\n<td>\u9700\u786c\u9ad4\u652f\u63f4 FP8<\/td>\n<\/tr>\n<tr>\n<td><strong>AWQ<\/strong>\uff084-bit\uff09<\/td>\n<td>\u4e2d\u9ad8<\/td>\n<td>\u7701\u5f88\u591a<\/td>\n<td>\u986f\u5b58\u5403\u7dca\u3001\u8981\u885d\u4f75\u767c<\/td>\n<td>\u7cbe\u5ea6\u7565\u964d\u3001\u9700 AWQ \u7248\u6a21\u578b<\/td>\n<\/tr>\n<tr>\n<td><strong>GPTQ<\/strong>\uff084-bit\uff09<\/td>\n<td>\u4e2d\u9ad8<\/td>\n<td>\u7701\u5f88\u591a<\/td>\n<td>\u540c\u4e0a\uff0c\u793e\u7fa4\u6a21\u578b\u591a<\/td>\n<td>\u540c\u4e0a<\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">\u4ec0\u9ebc\u6642\u5019\u300c\u503c\u5f97\u300d\u91cf\u5316<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li>\u2705 <strong>\u503c\u5f97<\/strong>\uff1a\u6a21\u578b\u585e\u4e0d\u9032\u4f60\u7684\u5361\u3001\u6216\u60f3\u5728\u540c\u4e00\u5f35\u5361\u4e0a\u670d\u52d9\u66f4\u591a\u4f75\u767c\uff08KV Cache \u6709\u66f4\u591a\u7a7a\u9593\uff09\u3002<\/li>\n\n\n<li>\u2705 <strong>\u503c\u5f97<\/strong>\uff1a\u4f60\u7684\u4efb\u52d9\u5c0d\u4e00\u5169\u5206\u7cbe\u5ea6\u4e0d\u654f\u611f\uff08\u804a\u5929\u3001\u6458\u8981\u3001\u5ba2\u670d\uff09\u3002<\/li>\n\n\n<li>\ud83e\udd14 <strong>\u518d\u60f3\u60f3<\/strong>\uff1a\u6578\u5b78\u3001\u7a0b\u5f0f\u78bc\u751f\u6210\u3001\u9700\u8981\u56b4\u8b39\u63a8\u7406\u2014\u2014\u5148\u7528\u5c0f\u6a23\u672c\u6bd4\u5c0d\u91cf\u5316\u524d\u5f8c\u7684\u7b54\u6848\u54c1\u8cea\u518d\u6c7a\u5b9a\u3002<\/li>\n\n\n<li>\u274c <strong>\u5225\u70ba\u7701\u800c\u7701<\/strong>\uff1a\u986f\u5b58\u660e\u660e\u5920\uff0c\u786c\u4e0a 4-bit \u53cd\u800c\u767d\u767d\u640d\u5931\u7cbe\u5ea6\uff0c\u541e\u5410\u63d0\u5347\u4e5f\u6709\u9650\u3002<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\ud83d\udccc <strong>\u77e5\u9053\u5c31\u597d<\/strong>\uff1aFP8 \u901a\u5e38\u662f H100 \/ L40S \u9019\u985e\u65b0\u5361\u624d\u5403\u5f97\u52d5\uff1bAWQ\/GPTQ \u662f\u793e\u7fa4\u6700\u5e38\u898b\u3001\u8001\u5361\u4e5f\u80fd\u8dd1\u7684 4-bit \u65b9\u6848\u3002\u4f60\u4e0d\u7528\u80cc\u7d30\u7bc0\uff0c\u8a18\u5f97\u300c\u91cf\u5316 = \u7701\u986f\u5b58\u63db\u4e00\u9ede\u7cbe\u5ea6\u300d\u5c31\u6293\u5f97\u4f4f\u65b9\u5411\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u58d3\u6e2c\uff1a\u7528\u6578\u5b57\u8aaa\u8a71\uff0c\u5225\u9760\u611f\u89ba<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u8abf\u53c3\u6700\u5fcc\u8af1\u300c\u611f\u89ba\u597d\u50cf\u8b8a\u5feb\u4e86\u300d\u3002vLLM \u5167\u5efa\u4e86\u58d3\u6e2c\u5de5\u5177\uff0c\u8b93\u4f60\u7528\u771f\u5be6\u6578\u5b57\u5224\u65b7\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6982\u5ff5\uff1a\u4e09\u500b\u4f60\u4e00\u5b9a\u8981\u770b\u61c2\u7684\u6307\u6a19<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\"><span><code class=\"hljs\">TTFT  (Time To First Token)  \u7b2c\u4e00\u500b\u5b57\u5410\u51fa\u4f86\u8981\u591a\u4e45 \u2192 \u4f7f\u7528\u8005\u300c\u7b49\u5f85\u611f\u300d\nTPOT  (Time Per Output Token) \u6bcf\u500b\u5f8c\u7e8c token \u9593\u9694 \u2192 \u6253\u5b57\u300c\u6d41\u66a2\u611f\u300d\nThroughput (tokens\/sec)       \u6574\u53f0\u6a5f\u6bcf\u79d2\u7e3d\u7522\u51fa \u2192 \u4f60\u7684\u300c\u7522\u80fd\u300d\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u53e5\u8a71\u8a18\u6cd5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>TTFT<\/strong> \u9ad8 = \u4f7f\u7528\u8005\u76ef\u8457\u7a7a\u767d\u756b\u9762\u4e7e\u7b49\uff08prefill \u968e\u6bb5\u6162\u6216\u6392\u968a\uff09\u3002<\/li>\n\n\n<li><strong>TPOT<\/strong> \u9ad8 = \u5b57\u4e00\u500b\u4e00\u500b\u6162\u6162\u8e66\uff08decode \u968e\u6bb5\u6162\uff0c\u6216\u6279\u6b21\u592a\u6eff\u5728\u6436\u8cc7\u6e90\uff09\u3002<\/li>\n\n\n<li><strong>Throughput<\/strong> \u4f4e = \u82b1\u540c\u6a23\u7684\u9322\u670d\u52d9\u7684\u4eba\u8b8a\u5c11\uff08\u6c92\u5582\u98fd GPU\uff09\u3002<\/li>\n\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u600e\u9ebc\u8dd1\uff08\u6982\u5ff5\u7248\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u65b0\u7248 vLLM \u63d0\u4f9b <code>vllm bench<\/code> \u5b50\u6307\u4ee4\uff1b\u820a\u4e00\u9ede\u7684\u7248\u672c\u662f <code>benchmark_serving.py<\/code> \u8173\u672c\u3002\u4f60\u4e0d\u7528\u80cc\u53c3\u6578\uff0c\u91cd\u9ede\u662f\u770b\u61c2\u5b83\u5728\u505a\u4ec0\u9ebc\uff1a<\/p>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-4\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\"><span class=\"hljs-comment\"># \u5148\u5728\u4e00\u500b\u7d42\u7aef\u628a server \u8dd1\u8d77\u4f86\uff08\u7b2c 1 \u7bc7\u6559\u904e\uff09<\/span>\nvllm serve Qwen\/Qwen2<span class=\"hljs-number\">.5<\/span><span class=\"hljs-number\">-7<\/span>B-Instruct --max-num-seqs <span class=\"hljs-number\">256<\/span>\n\n<span class=\"hljs-comment\"># \u53e6\u4e00\u500b\u7d42\u7aef\u6253\u58d3\u6e2c\uff1a\u6a21\u64ec\u591a\u500b\u4f75\u767c\u8acb\u6c42\u6253\u9032\u53bb<\/span>\nvllm bench serve \\\n  --model Qwen\/Qwen2<span class=\"hljs-number\">.5<\/span><span class=\"hljs-number\">-7<\/span>B-Instruct \\\n  --num-prompts <span class=\"hljs-number\">500<\/span> \\\n  --request-rate <span class=\"hljs-number\">20<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-4\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9019\u5728\u5e79\u561b<\/strong>\uff1a\u6a21\u64ec\u300c500 \u500b\u8acb\u6c42\u3001\u6bcf\u79d2\u7d04 20 \u500b\u300d\u704c\u9032\u4f60\u7684 server\uff0c\u8dd1\u5b8c\u5370\u51fa\u4e00\u5f35\u5831\u544a\uff1a\u5e73\u5747 TTFT\u3001TPOT\u3001\u4ee5\u53ca\u6574\u9ad4 throughput\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u770b\u61c2\u5831\u544a\uff0c\u6c7a\u5b9a\u5f80\u54ea\u8abf<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u62ff\u5230\u6578\u5b57\u5f8c\u7684\u6c7a\u7b56\u908f\u8f2f\uff08\u9019\u624d\u662f\u91cd\u9ede\uff09\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead>\n<tr>\n<th>\u89c0\u5bdf\u5230\u7684\u73fe\u8c61<\/th>\n<th>\u53ef\u80fd\u539f\u56e0<\/th>\n<th>\u8abf\u7684\u65b9\u5411<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Throughput \u4f4e\u3001GPU \u4f7f\u7528\u7387\u6c92\u6eff<\/td>\n<td>\u6279\u6b21\u5582\u4e0d\u98fd<\/td>\n<td>\u8abf\u9ad8 <code>--max-num-batched-tokens<\/code> \/ <code>--max-num-seqs<\/code><\/td>\n<\/tr>\n<tr>\n<td>TTFT \u96a8\u4f75\u767c\u66b4\u589e<\/td>\n<td>\u8acb\u6c42\u5728\u6392\u968a<\/td>\n<td>\u52a0\u5361\uff08TP\uff09\u6216\u91cf\u5316\u9a30\u51fa KV Cache \u7a7a\u9593<\/td>\n<\/tr>\n<tr>\n<td>\u9ad8\u4f75\u767c\u6642 OOM \/ \u8acb\u6c42\u88ab\u62d2<\/td>\n<td>\u986f\u5b58\u4e0d\u5920\u585e KV Cache<\/td>\n<td>\u91cf\u5316\u3001\u6216\u8abf\u9ad8 <code>--gpu-memory-utilization<\/code>\uff08\u7b2c 2 \u7bc7\uff09<\/td>\n<\/tr>\n<tr>\n<td>TPOT \u504f\u9ad8\u4f46\u541e\u5410\u4e5f\u9ad8<\/td>\n<td>\u6279\u6b21\u592a\u6eff\uff0c\u500b\u5225\u8acb\u6c42\u88ab\u7a00\u91cb<\/td>\n<td>\u8abf\u4f4e <code>--max-num-seqs<\/code> \u63db\u4f4e\u5ef6\u9072<\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\ud83c\udfaf <strong>\u5fc3\u6cd5<\/strong>\uff1a\u4e00\u6b21\u53ea\u52d5\u4e00\u500b\u53c3\u6578\uff0c\u91cd\u8dd1\u58d3\u6e2c\u6bd4\u8f03\u3002\u540c\u6642\u6539\u4e09\u500b\uff0c\u4f60\u6c38\u9060\u4e0d\u77e5\u9053\u662f\u54ea\u500b\u8d77\u7684\u4f5c\u7528\u2014\u2014\u9019\u8ddf\u4f60\u53eb AI \u4e00\u6b21\u6539\u4e00\u500b\u6a94\u6848\u518d\u9a57\u6536\uff0c\u662f\u540c\u4e00\u500b\u9053\u7406\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\ud83d\udea9 \u5be6\u6230\u7d05\u65d7\uff1aAI \u751f\u7684\u6307\u4ee4\u6700\u5e38\u898b\u7684\u5751<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">AI \u7d66\u4f60\u7684 <code>vllm serve<\/code> \u6307\u4ee4\u770b\u8d77\u4f86\u6c38\u9060\u5f88\u5c08\u696d\uff0c\u4ee5\u4e0b\u56db\u500b\u662f\u5b83\u6700\u5e38\u51fa\u5305\u3001\u800c\u4f60\u4e00\u773c\u5c31\u80fd\u6293\u7684\u5730\u65b9\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea9 \u7d05\u65d7 1\uff1aTP \u6578\u4e0d\u7b49\u65bc\u4f60\u7684 GPU \u5f35\u6578<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-5\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\"><span class=\"hljs-comment\"># AI \u6284\u4e86\u5225\u4eba 4 \u5361\u7684\u6307\u4ee4\uff0c\u8cbc\u5230\u4f60 2 \u5361\u7684\u6a5f\u5668<\/span>\nvllm serve some<span class=\"hljs-number\">-70<\/span>b-model --tensor-parallel-size <span class=\"hljs-number\">4<\/span>   <span class=\"hljs-comment\"># \ud83d\udea9 \u4f60\u53ea\u6709 2 \u5f35\u5361<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-5\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u75c7\u72c0<\/strong>\uff1a\u555f\u52d5\u5c31\u5831\u932f\uff0c\u985e\u4f3c <code>RuntimeError: ... world_size ... available devices 2<\/code>\u3002 <strong>\u600e\u9ebc\u8ddf AI \u8aaa<\/strong>\uff1a\u300c\u6211\u9019\u53f0\u53ea\u6709 2 \u5f35 GPU\uff0c\u628a tensor-parallel-size \u6539\u6210\u7b26\u5408\u6211\u7684\u786c\u9ad4\u300d\u3002 <strong>\u984d\u5916\u6aa2\u67e5<\/strong>\uff1aTP \u6578\u9084\u8981\u80fd\u6574\u9664\u6a21\u578b\u7684 attention head \u6578\uff08\u901a\u5e38 2\u30014\u30018 \u9019\u7a2e\u5f88\u5b89\u5168\uff09\u3002\u8a2d 3 \u9019\u7a2e\u602a\u6578\u5b57\u8981\u8b66\u89ba\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea9 \u7d05\u65d7 2\uff1a<code>--quantization awq<\/code> \u537b\u8f09\u5165\u539f\u7248\u6a21\u578b<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-6\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\"><span class=\"hljs-comment\"># \ud83d\udea9 \u6a21\u578b\u662f\u539f\u7248 FP16\uff0c\u537b\u53eb vLLM \u7528 awq \u8b80\u5b83<\/span>\nvllm serve Qwen\/Qwen2<span class=\"hljs-number\">.5<\/span><span class=\"hljs-number\">-7<\/span>B-Instruct --quantization awq\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-6\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u75c7\u72c0<\/strong>\uff1a\u8f09\u5165\u5931\u6557\uff0c\u6216\u5831\u300c\u627e\u4e0d\u5230\u91cf\u5316\u6b0a\u91cd \/ quant config\u300d\u3002 <strong>\u6839\u56e0<\/strong>\uff1a<code>--quantization awq<\/code> \u662f\u300c\u7528 AWQ \u65b9\u5f0f\u8b80\u4e00\u500b\u5df2\u7d93\u662f AWQ \u7684\u6a21\u578b\u300d\uff0c\u4e0d\u662f\u5373\u6642\u58d3\u7e2e\u3002 <strong>\u600e\u9ebc\u8ddf AI \u8aaa<\/strong>\uff1a\u300c\u8981\u561b\u63db\u6210\u91cf\u5316\u7248\u6a21\u578b\u540d\u7a31\uff08\u4f8b\u5982 <code>-AWQ<\/code> \u7d50\u5c3e\u90a3\u500b\uff09\uff0c\u8981\u561b\u62ff\u6389 <code>--quantization<\/code>\u300d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea9 \u7d05\u65d7 3\uff1a<code>--max-num-batched-tokens<\/code> \u8a2d\u592a\u5c0f\uff0c\u541e\u5410\u4e0a\u4e0d\u53bb<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-7\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\">vllm serve my-model --max-num-batched-tokens <span class=\"hljs-number\">512<\/span>   <span class=\"hljs-comment\"># \ud83d\udea9 \u592a\u5c0f\u4e86<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-7\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u75c7\u72c0<\/strong>\uff1a\u58d3\u6e2c\u6642 GPU \u4f7f\u7528\u7387\u4e0a\u4e0d\u53bb\u3001throughput \u5361\u5728\u4f4e\u9ede\uff0c\u4f46\u6c92\u6709\u5831\u932f\uff08\u6700\u9670\u96aa\uff0c\u56e0\u70ba\u300c\u770b\u8d77\u4f86\u80fd\u8dd1\u300d\uff09\u3002 <strong>\u6839\u56e0<\/strong>\uff1a\u6bcf\u500b step \u53ea\u80fd\u8655\u7406 512 token\uff0c\u9577 prompt \u9032\u4f86\u9023\u4e00\u500b\u90fd\u585e\u4e0d\u6eff\uff0cGPU \u5927\u91cf\u9592\u7f6e\u3002 <strong>\u600e\u9ebc\u8ddf AI \u8aaa<\/strong>\uff1a\u300c\u541e\u5410\u504f\u4f4e\u4f46 GPU \u6c92\u6eff\uff0c\u5e6b\u6211\u628a max-num-batched-tokens \u8abf\u5927\uff08\u4f8b\u5982 8192 \u6216\u66f4\u9ad8\uff09\u518d\u58d3\u6e2c\u6bd4\u8f03\u300d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea9 \u7d05\u65d7 4\uff1a\u91cf\u5316\u7701\u4e0b\u7684\u986f\u5b58\u6c92\u62ff\u53bb\u7528<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u91cf\u5316\u628a\u6b0a\u91cd\u58d3\u5c0f\u5f8c\uff0c\u591a\u51fa\u4f86\u7684\u986f\u5b58\u61c9\u8a72\u8b93 KV Cache \u5403\u6389\u3001\u670d\u52d9\u66f4\u591a\u4f75\u767c\u3002\u5982\u679c AI \u91cf\u5316\u4e86\u537b\u6c92\u52d5 <code>--max-num-seqs<\/code>\uff0c\u7b49\u65bc\u7701\u4e86\u9322\u6c92\u82b1\u51fa\u53bb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u600e\u9ebc\u8ddf AI \u8aaa<\/strong>\uff1a\u300c\u6211\u5df2\u7d93\u91cf\u5316\u7701\u4e86\u986f\u5b58\uff0c\u9806\u4fbf\u628a\u4f75\u767c\u4e0a\u9650\uff08max-num-seqs\uff09\u5f80\u4e0a\u8abf\uff0c\u628a\u7a7a\u9593\u7528\u8d77\u4f86\u300d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udea9 \u7d05\u65d7 5\uff1a\u53ea\u6191\u4e00\u6b21\u7d50\u679c\u5c31\u4e0b\u7d50\u8ad6<\/h3>\n\n\n<div class=\"cbc-code-wrapper\"><div class=\"cbc-code-bar\"><button class=\"cbc-copy-button\">Copy<\/button><\/div><pre class=\"wp-block-code\" aria-describedby=\"shcb-language-8\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\">\u300c\u6211\u628a max-num-seqs \u5f9e <span class=\"hljs-number\">128<\/span> \u8abf\u5230 <span class=\"hljs-number\">256<\/span>\uff0c\u611f\u89ba\u5feb\u591a\u4e86\u300d   <span class=\"hljs-comment\"># \ud83d\udea9 \u6c92\u6709\u6578\u64da\u3001\u6c92\u6709\u5c0d\u7167<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-8\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">PHP<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">php<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u6c92\u6709\u58d3\u6e2c\u6578\u5b57\u3001\u6c92\u6709\u56fa\u5b9a\u5176\u4ed6\u8b8a\u56e0\uff0c\u9019\u7a2e\u7d50\u8ad6\u4e0d\u53ef\u4fe1\u3002\u8981\u6709 before\/after \u7684 TTFT \/ throughput \u624d\u7b97\u6578\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">Vibe Coder \u9a57\u6536\u6aa2\u67e5\u9ede<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u62ff\u5230 AI \u7684\u8abf\u6821\u65b9\u6848\u6642\uff0c\u9010\u9805\u78ba\u8a8d\uff08\u505a\u5f97\u5230\u7684\u52d5\u4f5c\uff09\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>\u5c0d\u786c\u9ad4<\/strong>\uff1a<code>--tensor-parallel-size<\/code> \u7684\u6578\u5b57 = \u4f60\u6a5f\u5668\u5be6\u969b GPU \u5f35\u6578\u55ce\uff1f\n<ul>\n\n<li>\u52d5\u4f5c\uff1a\u8dd1 <code>nvidia-smi -L<\/code> \u770b\u6709\u5e7e\u5f35\u5361\uff0c\u6578\u5b57\u8981\u5c0d\u5f97\u4e0a\u3002<\/li>\n\n<\/ul>\n<\/li>\n\n\n<li><strong>\u5c0d\u6a21\u578b<\/strong>\uff1a\u6709 <code>--quantization<\/code> \u6642\uff0c\u6a21\u578b\u540d\u7a31\u662f\u91cf\u5316\u7248\u55ce\uff08\u901a\u5e38\u5e36 <code>AWQ<\/code>\/<code>GPTQ<\/code>\/<code>FP8<\/code> \u5b57\u6a23\uff09\uff1f<\/li>\n\n\n<li><strong>\u80fd\u8d77\u4f86<\/strong>\uff1aserver \u555f\u52d5 log \u6709\u6c92\u6709\u5370\u51fa\u985e\u4f3c <code>Application startup complete<\/code> \/ \u76e3\u807d\u57e0\u865f\uff0c\u800c\u4e0d\u662f traceback\uff1f\n<ul>\n\n<li>\u52d5\u4f5c\uff1a\u770b\u555f\u52d5\u8f38\u51fa\u6700\u5f8c\u5e7e\u884c\uff0c\u5361\u5728 error \u5c31\u662f\u6c92\u8d77\u4f86\u3002<\/li>\n\n<\/ul>\n<\/li>\n\n\n<li><strong>\u6709\u6578\u64da<\/strong>\uff1a\u8abf\u53c3\u524d\u5f8c\u90fd\u6709\u8dd1 <code>vllm bench<\/code>\uff08\u6216\u7b49\u50f9\u58d3\u6e2c\uff09\uff0c\u62ff\u5f97\u51fa TTFT \/ throughput \u7684 before\/after \u55ce\uff1f<\/li>\n\n\n<li><strong>\u4e00\u6b21\u4e00\u8b8a\u6578<\/strong>\uff1a\u9019\u6b21\u8abf\u6574\u53ea\u52d5\u4e86\u4e00\u500b\u53c3\u6578\u55ce\uff1f\uff08\u591a\u500b\u4e00\u8d77\u52d5 \u2192 \u8981\u6c42\u5206\u958b\u9a57\u8b49\uff09<\/li>\n\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u53ea\u8981\u6709\u4e00\u9805\u6253\u53c9\uff0c\u5c31\u628a\u5b83\u8cbc\u56de\u53bb\u554f AI\uff0c\u5225\u6025\u8457\u6536\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u770b\u4e0d\u61c2\u5c31\u9019\u6a23\u554f AI<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u628a\u6307\u4ee4\u6216\u58d3\u6e2c\u5831\u544a\u76f4\u63a5\u8cbc\u7d66 AI\uff0c\u5957\u7528\u9019\u4e9b\u63d0\u554f\u53e5\uff1a<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u7528\u767d\u8a71\u89e3\u91cb\u9019\u4e32 vllm serve \u53c3\u6578\u6bcf\u4e00\u500b\u5728\u5e79\u561b\uff0c\u6211\u8981\u5224\u65b7\u9069\u4e0d\u9069\u5408\u6211\u7684\u786c\u9ad4\u3002\u6211\u6709 __ \u5f35 __ \u986f\u5361\u3002\u300d<\/p>\n<\/blockquote>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u9019\u662f\u6211\u7684 vllm bench \u58d3\u6e2c\u5831\u544a\uff08\u8cbc\u4e0a\uff09\uff0cthroughput \u504f\u4f4e\u4f46 GPU \u4f7f\u7528\u7387\u6c92\u6eff\uff0c\u6700\u53ef\u80fd\u662f\u54ea\u500b\u53c3\u6578\u8981\u8abf\uff1f\u4e00\u6b21\u53ea\u7d66\u6211\u4e00\u500b\u5efa\u8b70\u3002\u300d<\/p>\n<\/blockquote>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u6211\u60f3\u5728\u540c\u4e00\u5f35\u5361\u4e0a\u670d\u52d9\u66f4\u591a\u4f75\u767c\uff0c\u91cf\u5316(AWQ\/GPTQ\/FP8)\u54ea\u7a2e\u9069\u5408\u6211\u9019\u5f35\u5361\uff1f\u5404\u81ea\u7684\u7cbe\u5ea6\u4ee3\u50f9\u662f\u4ec0\u9ebc\uff1f\u300d<\/p>\n<\/blockquote>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u5e6b\u6211\u6aa2\u67e5\uff1a\u9019\u500b &#8211;quantization \u8a2d\u5b9a\u8ddf\u6211\u8f09\u5165\u7684\u6a21\u578b\u540d\u7a31\u662f\u76f8\u5bb9\u7684\u55ce\uff1f\u300d<\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u672c\u7bc7\u5c0f\u7d50<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>continuous batching<\/strong> \u662f vLLM \u9ad8\u541e\u5410\u7684\u5f15\u64ce\uff0c\u81ea\u52d5\u904b\u4f5c\uff0c\u4f60\u53ea\u8981\u9935\u5c0d\u53c3\u6578\u628a\u5b83\u5582\u98fd\u3002<\/li>\n\n\n<li>\u4e94\u500b\u6838\u5fc3\u65cb\u9215\uff1a<code>--tensor-parallel-size<\/code>\uff08\u5e7e\u5f35\u5361\u8dd1\u4e00\u500b\u6a21\u578b\uff09\u3001<code>--max-num-seqs<\/code>\uff08\u5e7e\u500b\u4eba\u4e0a\u8eca\uff09\u3001<code>--max-num-batched-tokens<\/code>\uff08\u4e00\u6279\u7e3d\u5de5\u4f5c\u91cf\uff09\u3001<code>--quantization<\/code>\uff08\u58d3\u6b0a\u91cd\u63db\u986f\u5b58\uff09\u3001<code>--dtype<\/code>\uff08\u653e auto \u5c31\u597d\uff09\u3002<\/li>\n\n\n<li><strong>\u91cf\u5316 = \u7701\u986f\u5b58\/\u63d0\u4f75\u767c \u63db\u4e00\u9ede\u7cbe\u5ea6<\/strong>\uff1b\u986f\u5b58\u5403\u7dca\u6216\u8981\u885d\u4f75\u767c\u624d\u503c\u5f97\uff0c\u7cbe\u5ea6\u654f\u611f\u4efb\u52d9\u5148\u5c0f\u6a23\u672c\u9a57\u8b49\u3002<\/li>\n\n\n<li><strong>\u8abf\u53c3\u9760\u6578\u64da<\/strong>\uff1a\u770b TTFT \/ TPOT \/ throughput \u4e09\u500b\u6307\u6a19\uff0c\u4e00\u6b21\u52d5\u4e00\u500b\u53c3\u6578\u91cd\u58d3\u6e2c\u3002<\/li>\n\n\n<li>\u8a8d\u5f97 5 \u500b\u7d05\u65d7\uff1aTP \u5c0d\u4e0d\u4e0a\u5361\u6578\u3001\u91cf\u5316\u683c\u5f0f\u5c0d\u4e0d\u4e0a\u6a21\u578b\u3001batched-tokens \u592a\u5c0f\u3001\u7701\u4e86\u986f\u5b58\u6c92\u7528\u3001\u6c92\u6578\u64da\u5c31\u4e0b\u7d50\u8ad6\u3002<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff08#04\uff0c\u7cfb\u5217\u6700\u7d42\u7bc7\uff09\u6703\u628a\u9019\u4e9b\u4e32\u8d77\u4f86\uff0c\u8ac7\u4e0a\u7dda\u5f8c\u7684\u7a69\u5b9a\u6027\u8207\u76e3\u63a7\u2014\u2014\u7576\u534a\u591c\u8acb\u6c42\u91cf\u7206\u4e86\uff0c\u4f60\u600e\u9ebc\u5f9e\u6307\u6a19\u770b\u51fa\u662f\u649e\u5230\u54ea\u500b\u4e0a\u9650\u3002<\/p>\n\n\n\n<div id=\"quiz-a3280bb1\">\n<style>\n#quiz-a3280bb1 .badge { display: inline-block; padding: 4px 10px; border-radius: 12px; font-size: 0.75em; font-weight: bold; margin-left: 8px; }#quiz-a3280bb1 .badge.scenario { background: #ebf8ff; color: #2b6cb0; }#quiz-a3280bb1 .badge.debug { background: #fef3c7; color: #92400e; }#quiz-a3280bb1 .question-adv { background: white; border-radius: 10px; padding: 24px; margin: 20px 0; box-shadow: 0 2px 8px rgba(0,0,0,0.08); border-left: 4px solid #ed8936; }#quiz-a3280bb1 .question-adv h3 { margin-top: 0; color: #2d3748; }#quiz-a3280bb1 .code-block { background: #1e1e1e; color: #d4d4d4; padding: 14px; border-radius: 6px; font-family: \"Menlo\",\"Monaco\",\"Courier New\",monospace; font-size: 0.9em; overflow-x: auto; margin: 12px 0; white-space: pre-wrap; }#quiz-a3280bb1 .options-adv { list-style: none; padding: 0; }#quiz-a3280bb1 .options-adv li { padding: 14px 16px; margin: 10px 0; border: 2px solid #e2e8f0; border-radius: 8px; cursor: pointer; transition: all 0.2s; }#quiz-a3280bb1 .options-adv li:hover:not(.disabled) { border-color: #ed8936; background: #fffaf0; }#quiz-a3280bb1 .options-adv li.disabled { cursor: default; opacity: 0.7; }#quiz-a3280bb1 .options-adv li.show-answer { border-color: #48bb78; background: #f0fff4; }#quiz-a3280bb1 .options-adv li.wrong { border-color: #fc8181; background: #fff5f5; }#quiz-a3280bb1 .feedback-adv { margin-top: 16px; padding: 14px; border-radius: 8px; display: none; line-height: 1.6; }#quiz-a3280bb1 .feedback-adv.correct { background: #f0fff4; color: #276749; display: block; }#quiz-a3280bb1 .feedback-adv.wrong { background: #fff5f5; color: #c53030; display: block; }#quiz-a3280bb1 .feedback-adv \n      .score-adv { text-align: center; font-size: 1.4em; padding: 24px; background: white; border-radius: 10px; margin-top: 24px; box-shadow: 0 2px 8px rgba(0,0,0,0.08); }#quiz-a3280bb1 .intro { background: #ebf8ff; padding: 16px; border-radius: 8px; margin-bottom: 20px; color: #2c5282; }\n<\/style>\n<h1>\u9032\u968e\u6e2c\u9a57\uff1avLLM \u69a8\u4e7e\u541e\u5410\u91cf<\/h1>\n    <div class=\"intro\">\n      <strong>\u6e2c\u9a57\u76ee\u6a19\uff1a<\/strong>\u9a57\u8b49\u4f60\u662f\u5426\u80fd\u5728\u5be6\u969b\u60c5\u5883\u4e2d\u61c9\u7528\u6240\u5b78\u3002<br \/>\n      \u5171 5 \u984c\uff0c\u5305\u542b\u60c5\u5883\u984c\u8207\u932f\u8aa4\u8a3a\u65b7\u984c\u3002\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv1\">\n      <h3>1. \u4f60\u7684\u670d\u52d9\u5e38\u6536\u5230\u300c\u4e00\u5806\u9577 prompt\uff08\u5404\u7d04 4000 token\uff09\u300d\u6e67\u5165\uff0c\u767c\u73fe\u541e\u5410\u4e0a\u4e0d\u53bb\u3001GPU \u537b\u6c92\u6eff\u8f09\u3002\u76ee\u524d\u8a2d\u5b9a\u662f <code>--max-num-seqs 256 --max-num-batched-tokens 8192<\/code>\u3002\u6700\u8a72\u5148\u52d5\u7684\u662f\uff1f <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <div class=\"code-block\">&#8211;max-num-seqs 256           # \u4f75\u767c\u4e0a\u9650\uff08\u5e7e\u500b\u4eba\u4e0a\u8eca\uff09\n&#8211;max-num-batched-tokens 8192  # \u4e00\u6279\u7e3d token \u91cf\uff088192 \/ 4000 \u2248 2\uff09<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(1, 'A', 'B')\">A. \u628a max-num-seqs \u8abf\u5230 512\uff0c\u8b93\u66f4\u591a\u4eba\u4e0a\u8eca<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'B', 'B')\">B. \u8abf\u9ad8 max-num-batched-tokens\uff0c\u56e0\u70ba token \u4e0a\u9650\u5148\u5361\u4f4f\u3001\u6bcf\u6279\u53ea\u585e\u5f97\u4e0b\u7d04 2 \u500b\u9577\u8acb\u6c42<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'C', 'B')\">C. \u52a0\u4e0a &#8211;quantization awq \u5c31\u6703\u81ea\u52d5\u8b8a\u5feb<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'D', 'B')\">D. \u628a tensor-parallel-size \u8abf\u6210 1<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv1\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv2\">\n      <h3>2. \u4f60\u8981\u5728\u540c\u4e00\u5f35\u5361\u4e0a\u670d\u52d9\u66f4\u591a\u4f75\u767c\u4f7f\u7528\u8005\uff0c\u6c7a\u5b9a\u6539\u7528\u91cf\u5316\u6a21\u578b\u3002\u4ee5\u4e0b\u54ea\u500b\u505a\u6cd5\u6700\u6b63\u78ba\uff1f <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(2, 'A', 'C')\">A. \u4fdd\u6301\u539f\u7248\u6a21\u578b\uff0c\u53ea\u52a0\u4e0a &#8211;quantization awq \u5373\u53ef<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'B', 'C')\">B. \u63db\u91cf\u5316\u7248\u6a21\u578b\u5f8c\uff0c\u7dad\u6301\u539f\u672c\u7684 max-num-seqs \u4e0d\u52d5<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'C', 'C')\">C. \u63db\u6210\u91cf\u5316\u7248\u6a21\u578b\uff08\u540d\u7a31\u5e36 AWQ\uff09\uff0c\u4e26\u628a\u7701\u4e0b\u7684\u986f\u5b58\u62ff\u53bb\u8abf\u9ad8 max-num-seqs<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'D', 'C')\">D. \u986f\u5b58\u660e\u660e\u5920\u4e5f\u786c\u4e0a 4-bit\uff0c\u53cd\u6b63\u8d8a\u7701\u8d8a\u597d<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv2\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv3\">\n      <h3>3. \u4f60\u8981\u9a57\u6536 AI \u7684\u8abf\u6821\u65b9\u6848\uff0c\u4e14\u9700\u8981\u300c\u6709\u6578\u64da\u300d\u624d\u80fd\u6536\u3002\u4e0b\u5217\u54ea\u500b\u505a\u6cd5\u6700\u7b26\u5408\u6587\u7ae0\u5fc3\u6cd5\uff1f <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(3, 'A', 'D')\">A. \u540c\u6642\u8abf max-num-seqs\u3001batched-tokens\u3001\u91cf\u5316\u4e09\u9805\uff0c\u4e00\u6b21\u5230\u4f4d<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'B', 'D')\">B. \u6191\u624b\u52d5\u6253\u5e7e\u53e5\u8a71\u300c\u611f\u89ba\u8b8a\u5feb\u4e86\u300d\u5c31\u6536<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'C', 'D')\">C. \u53ea\u770b server \u6709\u6c92\u6709\u555f\u52d5\u6210\u529f\u5c31\u7b97\u901a\u904e<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'D', 'D')\">D. \u4e00\u6b21\u53ea\u52d5\u4e00\u500b\u53c3\u6578\uff0c\u7528 vllm bench \u8dd1\u51fa before\/after \u7684 TTFT \u8207 throughput \u6bd4\u8f03<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv3\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv4\">\n      <h3>4. AI \u7d66\u4f60\u9019\u4e32\u6307\u4ee4\uff0c\u8cbc\u5230\u4f60\u53ea\u6709 2 \u5f35 GPU \u7684\u6a5f\u5668\u4e0a\u8dd1\uff0c\u555f\u52d5\u5c31\u5831\u932f\u3002\u6700\u53ef\u80fd\u7684\u539f\u56e0\u662f\uff1f <span class=\"badge debug\">\u932f\u8aa4\u8a3a\u65b7<\/span><\/h3>\n      <div class=\"code-block\">$ vllm serve some-70b-model &#8211;tensor-parallel-size 4\nRuntimeError: &#8230; world_size &#8230; available devices 2<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(4, 'A', 'B')\">A. \u6a21\u578b\u6a94\u6848\u640d\u6bc0\uff0c\u9700\u8981\u91cd\u65b0\u4e0b\u8f09<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'B', 'B')\">B. tensor-parallel-size \u8a2d 4 \u4f46\u6a5f\u5668\u53ea\u6709 2 \u5f35\u5361\uff0cTP \u6578\u5c0d\u4e0d\u4e0a\u5be6\u969b GPU \u5f35\u6578<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'C', 'B')\">C. \u5fd8\u8a18\u52a0 &#8211;quantization \u53c3\u6578<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'D', 'B')\">D. max-num-batched-tokens \u8a2d\u592a\u5c0f<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv4\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv5\">\n      <h3>5. \u4e0b\u9762\u9019\u4e32\u6307\u4ee4\u8f09\u5165\u5931\u6557\uff0c\u5831\u300c\u627e\u4e0d\u5230\u91cf\u5316\u6b0a\u91cd \/ quant config\u300d\u3002\u554f\u984c\u51fa\u5728\u54ea\uff1f <span class=\"badge debug\">\u932f\u8aa4\u8a3a\u65b7<\/span><\/h3>\n      <div class=\"code-block\">$ vllm serve Qwen\/Qwen2.5-7B-Instruct &#8211;quantization awq\n# \u932f\u8aa4\uff1a\u627e\u4e0d\u5230\u91cf\u5316\u6b0a\u91cd \/ quant config<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(5, 'A', 'C')\">A. &#8211;quantization \u61c9\u8a72\u5beb\u6210 gptq \u624d\u5c0d<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'B', 'C')\">B. \u986f\u5b58\u4e0d\u8db3\uff0c\u9700\u8981\u8abf\u9ad8 &#8211;gpu-memory-utilization<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'C', 'C')\">C. \u8f09\u5165\u7684\u662f\u539f\u7248 FP16 \u6a21\u578b\uff0c\u537b\u53eb vLLM \u7528 AWQ \u8b80\u5b83\uff1b\u61c9\u63db\u6210 AWQ \u91cf\u5316\u7248\u6a21\u578b\u6216\u62ff\u6389 &#8211;quantization<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'D', 'C')\">D. tensor-parallel-size \u6c92\u8a2d\u5b9a<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv5\"><\/div>\n    <\/div>\n\n    <div class=\"score-adv\" id=\"scoreAdv\" style=\"display:none;\"><\/div>\n<script>\n\n      let answeredAdv = 0, correctAdv = 0;\n      const totalAdv = 5;\n      const explanationsAdv = {\n        1: \"\u7576 prompt \u5f88\u9577\u6642\uff0cmax-num-batched-tokens \u9019\u500b\u300c\u7e3d\u5de5\u4f5c\u91cf\u300d\u4e0a\u9650\u6703\u5148\u649e\u5230\uff1a8192 \/ 4000 \u2248 2\uff0c\u4e00\u6279\u53ea\u585e\u5f97\u4e0b\u7d04 2 \u500b\uff0c\u5c31\u7b97 max-num-seqs \u958b\u5230 256 \u4e5f\u8f2a\u4e0d\u5230\u3002\u6240\u4ee5\u8981\u8abf\u9ad8 batched-tokens \u628a\u6279\u6b21\u5582\u98fd\u3002\u55ae\u8abf seqs \u6c92\u7528\u3001\u91cf\u5316\u4e0d\u6703\u300c\u81ea\u52d5\u8b8a\u5feb\u300d\u3001TP \u4e5f\u975e\u6b64\u554f\u984c\u6839\u56e0\u3002\",\n        2: \"--quantization awq \u4e0d\u662f\u5373\u6642\u58d3\u7e2e\uff0c\u5fc5\u9808\u63db\u6210\u672c\u8eab\u5c31\u662f AWQ \u7684\u91cf\u5316\u7248\u6a21\u578b\uff08\u540d\u7a31\u901a\u5e38\u5e36 AWQ\uff09\u3002\u800c\u4e14\u91cf\u5316\u7701\u4e0b\u7684\u986f\u5b58\u8981\u62ff\u53bb\u8abf\u9ad8 max-num-seqs \u624d\u7b97\u628a\u597d\u8655\u7528\u51fa\u4f86\u3002\u986f\u5b58\u5920\u537b\u786c\u91cf\u5316\u53ea\u6703\u767d\u640d\u7cbe\u5ea6\u3002\",\n        3: \"\u6587\u7ae0\u5fc3\u6cd5\u662f\u300c\u4e00\u6b21\u53ea\u52d5\u4e00\u500b\u53c3\u6578\u3001\u7528\u6578\u64da\u8aaa\u8a71\u300d\u3002\u8981\u6709 vllm bench \u7684 before\/after\uff08TTFT\u3001throughput\uff09\u624d\u53ef\u4fe1\u3002\u540c\u6642\u6539\u591a\u9805\u7121\u6cd5\u6b78\u56e0\u3001\u6191\u611f\u89ba\u4e0d\u53ef\u4fe1\u3001\u53ea\u770b\u80fd\u5426\u555f\u52d5\u4e0d\u4ee3\u8868\u541e\u5410\u6709\u6539\u5584\u3002\",\n        4: \"\u932f\u8aa4\u8a0a\u606f world_size \/ available devices 2 \u660e\u767d\u6307\u51fa\uff1a\u4f60\u5ba3\u544a\u8981\u7528 4 \u5f35\u5361\uff08TP=4\uff09\u4f46\u5be6\u969b\u53ea\u6709 2 \u5f35\u3002\u7528 nvidia-smi -L \u78ba\u8a8d\u5361\u6578\uff0c\u628a tensor-parallel-size \u6539\u6210\u7b26\u5408\u786c\u9ad4\u7684\u503c\u5373\u53ef\u3002\",\n        5: \"--quantization awq \u662f\u300c\u7528 AWQ \u65b9\u5f0f\u8b80\u4e00\u500b\u5df2\u7d93\u662f AWQ \u7684\u6a21\u578b\u300d\uff0c\u539f\u7248 FP16 \u6a21\u578b\u88e1\u6c92\u6709\u91cf\u5316\u6b0a\u91cd\uff0c\u81ea\u7136\u627e\u4e0d\u5230 quant config\u3002\u89e3\u6cd5\uff1a\u63db\u6210 -AWQ \u7d50\u5c3e\u7684\u91cf\u5316\u7248\u6a21\u578b\uff0c\u6216\u76f4\u63a5\u62ff\u6389 --quantization\u3002\u9019\u8ddf gptq\u3001\u986f\u5b58\u3001TP \u90fd\u7121\u95dc\u3002\"\n      };\n      function checkAnswerAdv(qNum, selected, correctAns) {\n        const question = document.getElementById(\"qAdv\" + qNum);\n        if (question.classList.contains(\"answered\")) return;\n        question.classList.add(\"answered\");\n        answeredAdv++;\n        const options = question.querySelectorAll(\".options-adv li\");\n        options.forEach((opt) => {\n          opt.classList.add(\"disabled\");\n          if (opt.textContent.startsWith(correctAns + \".\")) opt.classList.add(\"show-answer\");\n        });\n        const feedback = document.getElementById(\"fbAdv\" + qNum);\n        if (selected === correctAns) {\n          correctAdv++;\n          feedback.className = \"feedback-adv correct\";\n          feedback.innerHTML = \"\u2713 \u6b63\u78ba\uff01\" + explanationsAdv[qNum];\n        } else {\n          feedback.className = \"feedback-adv wrong\";\n          feedback.innerHTML = \"\u2717 \u932f\u8aa4\u3002\u6b63\u78ba\u7b54\u6848\u662f \" + correctAns + \"\u3002\" + explanationsAdv[qNum];\n          options.forEach((opt) => { if (opt.textContent.startsWith(selected + \".\")) opt.classList.add(\"wrong\"); });\n        }\n        if (answeredAdv === totalAdv) {\n          const scoreDiv = document.getElementById(\"scoreAdv\");\n          scoreDiv.style.display = \"block\";\n          let msg = \"\";\n          if (correctAdv === totalAdv) msg = \"\ud83c\udfaf \u5b8c\u7f8e\uff01\u4f60\u5df2\u7d93\u6e96\u5099\u597d\u5be6\u6230\u4e86\uff01\";\n          else if (correctAdv >= Math.ceil(totalAdv * 0.8)) msg = \"\ud83d\udcaa \u5f88\u68d2\uff01\u518d\u8907\u7fd2\u4e00\u4e0b\u932f\u7684\u984c\u76ee\";\n          else if (correctAdv >= Math.ceil(totalAdv * 0.6)) msg = \"\ud83d\udcd6 \u5efa\u8b70\u91cd\u8b80\u76f8\u95dc\u6bb5\u843d\u5f8c\u518d\u8a66\u4e00\u6b21\";\n          else msg = \"\ud83d\udd04 \u5efa\u8b70\u91cd\u65b0\u95b1\u8b80\u6587\u4ef6\uff0c\u7279\u5225\u6ce8\u610f\u64cd\u4f5c\u6d41\u7a0b\";\n          scoreDiv.innerHTML = \"\u6e2c\u9a57\u5b8c\u6210\uff01\u5f97\u5206\uff1a\" + correctAdv + \" \/ \" + totalAdv + \"<br>\" + msg;\n        }\n      }\n    \n<\/script>\n<\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u6e2c\u9a57\uff1avLLM \u69a8\u4e7e\u541e\u5410\u91cf \u5171 5 \u984c\uff0c\u9ede\u9078\u7b54\u6848\u5f8c\u6703\u7acb\u5373\u986f\u793a\u7d50\u679c 1. \u95dc\u65bc continuous batch [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"rolpb_likes":0,"footnotes":""},"categories":[858],"tags":[867,847,869,868,365],"class_list":["post-839","post","type-post","status-publish","format-standard","hentry","category-vllm-","tag-continuous-batching","tag-vllm","tag-869","tag-868","tag-365"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":7}},"_links":{"self":[{"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/posts\/839","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/comments?post=839"}],"version-history":[{"count":0,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/posts\/839\/revisions"}],"wp:attachment":[{"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/media?parent=839"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/categories?post=839"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/tags?post=839"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}