{"id":838,"date":"2026-07-12T01:15:11","date_gmt":"2026-07-11T17:15:11","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%9102-kv-cache-%e8%88%87-pagedattention%ef%bc%9a%e8%a8%98%e6%86%b6%e9%ab%94%e5%8e%bb%e5%93%aa%e4%ba%86%e3%80%81%e5%8f%83%e6%95%b8\/"},"modified":"2026-07-12T01:15:11","modified_gmt":"2026-07-11T17:15:11","slug":"%e3%80%90vllm-%e5%b7%a5%e7%a8%8b%e5%b8%ab%e5%af%a6%e6%88%b0%e3%80%9102-kv-cache-%e8%88%87-pagedattention%ef%bc%9a%e8%a8%98%e6%86%b6%e9%ab%94%e5%8e%bb%e5%93%aa%e4%ba%86%e3%80%81%e5%8f%83%e6%95%b8","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%9102-kv-cache-%e8%88%87-pagedattention%ef%bc%9a%e8%a8%98%e6%86%b6%e9%ab%94%e5%8e%bb%e5%93%aa%e4%ba%86%e3%80%81%e5%8f%83%e6%95%b8\/","title":{"rendered":"\u3010vLLM \u5de5\u7a0b\u5e2b\u5be6\u6230\u3011#02 KV Cache \u8207 PagedAttention\uff1a\u8a18\u61b6\u9ad4\u53bb\u54ea\u4e86\u3001\u53c3\u6578\u600e\u9ebc\u8abf"},"content":{"rendered":"\n<div id=\"quiz-14549a14\">\n<style>\n#quiz-14549a14 .question { background: white; border-radius: 8px; padding: 20px; margin: 20px 0; box-shadow: 0 2px 4px rgba(0,0,0,0.1); }#quiz-14549a14 .question h3 { margin-top: 0; color: #333; }#quiz-14549a14 .options { list-style: none; padding: 0; }#quiz-14549a14 .options li { padding: 12px 15px; margin: 8px 0; border: 2px solid #ddd; border-radius: 6px; cursor: pointer; transition: all 0.2s; }#quiz-14549a14 .options li:hover:not(.disabled) { border-color: #4caf50; background: #f0fff0; }#quiz-14549a14 .options li.correct { border-color: #4caf50; background: #e8f5e9; }#quiz-14549a14 .options li.wrong { border-color: #f44336; background: #ffebee; }#quiz-14549a14 .options li.disabled { cursor: default; opacity: 0.7; }#quiz-14549a14 .options li.show-answer { border-color: #4caf50; background: #e8f5e9; }#quiz-14549a14 .feedback { margin-top: 15px; padding: 12px; border-radius: 6px; display: none; }#quiz-14549a14 .feedback.correct { background: #e8f5e9; color: #2e7d32; display: block; }#quiz-14549a14 .feedback.wrong { background: #ffebee; color: #c62828; display: block; }#quiz-14549a14 .score { text-align: center; font-size: 1.5em; padding: 20px; background: white; border-radius: 8px; margin-top: 20px; }#quiz-14549a14 .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\uff1aKV Cache \u8207 PagedAttention<\/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. \u6839\u64da\u6587\u7ae0\uff0c\u4e00\u9846 GPU \u8dd1 LLM \u6642\uff0c\u300c\u6703\u96a8\u8457\u8acb\u6c42\u8b8a\u591a\u3001\u5e8f\u5217\u8b8a\u9577\u800c\u81a8\u8139\u300d\u3001\u5403\u6389\u5927\u534a\u8a18\u61b6\u9ad4\u7684\u4e3b\u89d2\u662f\u4ec0\u9ebc\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(1, 'A', 'C')\">A. \u6a21\u578b\u6b0a\u91cd<\/li>\n        <li onclick=\"checkAnswer(1, 'B', 'C')\">B. CUDA context \u8207\u6846\u67b6\u66ab\u5b58<\/li>\n        <li onclick=\"checkAnswer(1, 'C', 'C')\">C. KV Cache<\/li>\n        <li onclick=\"checkAnswer(1, 'D', 'C')\">D. block-size \u8a2d\u5b9a\u503c<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb1\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q2\">\n      <h3>2. \u6587\u7ae0\u7528\u300c\u4fbf\u5229\u8cbc\u7b46\u8a18\u300d\u6bd4\u55bb KV Cache\uff0c\u4e3b\u8981\u662f\u70ba\u4e86\u8aaa\u660e\u5b83\u7684\u4ec0\u9ebc\u4f5c\u7528\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(2, 'A', 'A')\">A. \u5b58\u8d77\u7b97\u904e\u7684\u4e2d\u9593\u7d50\u679c\uff0c\u907f\u514d\u6bcf\u5410\u4e00\u500b\u5b57\u90fd\u628a\u524d\u9762\u91cd\u7b97\u4e00\u904d<\/li>\n        <li onclick=\"checkAnswer(2, 'B', 'A')\">B. \u628a\u6a21\u578b\u6b0a\u91cd\u58d3\u7e2e\u4ee5\u7bc0\u7701\u786c\u789f\u7a7a\u9593<\/li>\n        <li onclick=\"checkAnswer(2, 'C', 'A')\">C. \u8a18\u9304\u4f7f\u7528\u8005\u7684\u6b77\u53f2\u5c0d\u8a71\u4ee5\u4fbf\u65e5\u5f8c\u67e5\u8a62<\/li>\n        <li onclick=\"checkAnswer(2, 'D', 'A')\">D. \u52a0\u5bc6\u8acb\u6c42\u5167\u5bb9\u4ee5\u4fdd\u8b77\u96b1\u79c1<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb2\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q3\">\n      <h3>3. PagedAttention \u501f\u7528\u4e86\u4f5c\u696d\u7cfb\u7d71\u7684\u54ea\u500b\u8001\u62db\u6578\u4f86\u7ba1\u7406 KV Cache\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(3, 'A', 'B')\">A. \u591a\u57f7\u884c\u7dd2\u6392\u7a0b<\/li>\n        <li onclick=\"checkAnswer(3, 'B', 'B')\">B. \u5206\u9801\uff08\u628a\u8a18\u61b6\u9ad4\u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u9801\/block \u7ba1\u7406\uff09<\/li>\n        <li onclick=\"checkAnswer(3, 'C', 'B')\">C. \u6a94\u6848\u7cfb\u7d71\u5feb\u53d6<\/li>\n        <li onclick=\"checkAnswer(3, 'D', 'B')\">D. \u5783\u573e\u56de\u6536\uff08GC\uff09<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb3\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q4\">\n      <h3>4. \u770b\u5230\u4ee5\u4e0b\u555f\u52d5\u65e5\u8a8c\uff0c\u9019\u884c\u4e3b\u8981\u5728\u544a\u8a34\u4f60\u4ec0\u9ebc\uff1f<\/h3>\n      <div class=\"code-block\">Maximum concurrency for 8,192 tokens per request: 9.04x<\/div>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(4, 'A', 'D')\">A. \u6a21\u578b\u6b0a\u91cd\u4f54\u4e86 9.04GB<\/li>\n        <li onclick=\"checkAnswer(4, 'B', 'D')\">B. \u6bcf\u500b\u8acb\u6c42\u6700\u591a\u80fd\u751f\u6210 9 \u500b token<\/li>\n        <li onclick=\"checkAnswer(4, 'C', 'D')\">C. block-size \u88ab\u8a2d\u6210 9<\/li>\n        <li onclick=\"checkAnswer(4, 'D', 'D')\">D. \u82e5\u6bcf\u500b\u8acb\u6c42\u90fd\u5403\u6eff 8,192 token\uff0c\u5927\u7d04\u80fd\u540c\u6642\u670d\u52d9 9 \u500b\u8acb\u6c42<\/li>\n      <\/ul>\n      <div class=\"feedback\" id=\"fb4\"><\/div>\n    <\/div>\n\n    <div class=\"question\" id=\"q5\">\n      <h3>5. \u4f9d\u6587\u7ae0\u7684 OOM \u6392\u67e5\u9806\u5e8f\uff0c\u9047\u5230\u8a18\u61b6\u9ad4\u4e0d\u8db3\u6642\u300c\u6700\u6709\u6548\u3001\u61c9\u8a72\u6700\u5148\u52d5\u300d\u7684\u53c3\u6578\u662f\u54ea\u4e00\u500b\uff1f<\/h3>\n      <ul class=\"options\">\n        <li onclick=\"checkAnswer(5, 'A', 'B')\">A. &#8211;block-size<\/li>\n        <li onclick=\"checkAnswer(5, 'B', 'B')\">B. &#8211;max-model-len<\/li>\n        <li onclick=\"checkAnswer(5, 'C', 'B')\">C. &#8211;max-num-seqs<\/li>\n        <li onclick=\"checkAnswer(5, 'D', 'B')\">D. \u76f4\u63a5\u63db\u66f4\u5c0f\u7684\u6a21\u578b<\/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;\n      let correct = 0;\n      const total = 5;\n      const explanations = {\n        1: \"\u6a21\u578b\u6b0a\u91cd\u662f\u56fa\u5b9a\u7684\uff08\u6a21\u578b\u591a\u5927\u5c31\u591a\u5927\uff09\uff0c\u771f\u6b63\u6703\u81a8\u8139\u4e26\u5403\u6389\u5927\u534a\u8a18\u61b6\u9ad4\u7684\u662f KV Cache\u2014\u2014\u8acb\u6c42\u8d8a\u591a\u3001\u5e8f\u5217\u8d8a\u9577\uff0c\u7b46\u8a18\u8d8a\u539a\u3002\",\n        2: \"KV Cache \u5c31\u662f\u628a\u6bcf\u500b token \u7b97\u904e\u7684\u6ce8\u610f\u529b\u4e2d\u9593\u7d50\u679c\u5b58\u8d77\u4f86\uff0c\u4e0b\u6b21\u76f4\u63a5\u62ff\u4f86\u7528\uff0c\u907f\u514d\u91cd\u8907\u8a08\u7b97\uff0c\u5e8f\u5217\u8d8a\u9577\u7b46\u8a18\u8d8a\u539a\u3002\",\n        3: \"PagedAttention \u5b78\u4f5c\u696d\u7cfb\u7d71\u7684\u5206\u9801\uff1a\u628a KV Cache \u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684 block\uff0c\u7528\u591a\u5c11\u7d66\u591a\u5c11\u3001\u53ef\u4ee5\u6563\u8457\u653e\uff0c\u6e1b\u5c11\u9810\u7559\u6d6a\u8cbb\u548c\u788e\u7247\u3002\",\n        4: \"\u9019\u884c\u662f\u5bb9\u91cf\u5100\u8868\u677f\uff1a\u628a KV cache \u7e3d\u91cf\u9664\u4ee5\u6bcf\u500b\u8acb\u6c42\u7684 max-model-len\uff0c\u4f30\u7b97\u5927\u7d04\u80fd\u540c\u6642\u670d\u52d9\u5e7e\u500b\u8acb\u6c42\uff0874,096 \u00f7 8,192 \u2248 9.04\uff09\u3002\",\n        5: \"\u6392\u67e5\u9806\u5e8f\u662f\uff1a\u5148\u964d max-model-len\uff08\u6700\u6709\u6548\uff0c\u76f4\u63a5\u7e2e\u5c0f\u6bcf\u4efd\u7b46\u8a18\u4e0a\u9650\uff09\uff0c\u518d\u964d gpu-memory-utilization\uff0c\u518d\u964d max-num-seqs\uff0c\u6700\u5f8c\u624d\u63db\u5c0f\u6a21\u578b\u6216\u91cf\u5316\u3002\"\n      };\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 2 \u7bc7\uff0c\u5171 4 \u7bc7\uff5c\u96e3\u5ea6\uff1aL2-\u9032\u968e<br>\u524d\u7f6e\uff1a\u5df2\u8b80\u7b2c 1 \u7bc7\uff0c\u80fd\u7528 <code>vllm serve<\/code> \u628a\u6a21\u578b\u8dd1\u8d77\u4f86\uff0c\u77e5\u9053 token\u3001context length \u662f\u4ec0\u9ebc\u3002<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b2c 1 \u7bc7\u4f60\u5df2\u7d93\u628a\u6a21\u578b\u670d\u52d9\u8d77\u4f86\u4e86\u3002\u9019\u4e00\u7bc7\u8981\u56de\u7b54\u4e00\u500b\u4f60\u9072\u65e9\u6703\u9047\u5230\u7684\u554f\u984c\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\u6211\u7684 GPU \u660e\u660e\u6709 24GB\uff0c\u6a21\u578b\u6a94\u6848\u624d 14GB\uff0c\u70ba\u4ec0\u9ebc\u4e00\u8dd1\u8d77\u4f86\u5c31\u8aaa\u8a18\u61b6\u9ad4\u4e0d\u5920 \/ OOM\uff1f\u5269\u4e0b\u7684\u8a18\u61b6\u9ad4\u5230\u5e95\u88ab\u8ab0\u5403\u6389\u4e86\uff1f\u300d<\/p>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b54\u6848\u7684\u4e3b\u89d2\u5c31\u662f\u9019\u7bc7\u7684\u5169\u500b\u540d\u8a5e\uff1a<strong>KV Cache<\/strong> \u548c <strong>PagedAttention<\/strong>\u3002\u641e\u61c2\u5b83\u5011\uff0c\u4f60\u5c31\u80fd\u770b\u61c2 vLLM \u555f\u52d5\u6642\u90a3\u4e00\u9577\u4e32\u65e5\u8a8c\uff0c\u4e5f\u80fd\u81ea\u5df1\u52d5\u624b\u8abf\u53c3\u6578\u63db\u4f75\u767c\u3001\u907f\u958b OOM\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u7bc7\u4e0d\u63a8\u5c0e\u4efb\u4f55\u6578\u5b78\uff0c\u6211\u5011\u7528\u5de5\u7a0b\u5e2b\u7684\u76f4\u89ba\u4f86\u7406\u89e3\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<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>KV Cache<\/strong>\uff1a\u628a\u6a21\u578b\u7b97\u904e\u7684 token\u300c\u7b46\u8a18\u300d\u5b58\u8d77\u4f86\uff0c\u907f\u514d\u6bcf\u6b21\u90fd\u91cd\u7b97\u2014\u2014\u5e8f\u5217\u8d8a\u9577\uff0c\u7b46\u8a18\u8d8a\u539a\uff0c\u5403\u8d8a\u591a\u8a18\u61b6\u9ad4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>PagedAttention<\/strong>\uff1a\u50cf\u4f5c\u696d\u7cfb\u7d71\u7ba1\u8a18\u61b6\u9ad4\u90a3\u6a23\uff0c\u628a\u9019\u5806\u7b46\u8a18\u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u5c0f\u584a\u5206\u9801\u7ba1\u7406\uff0c\u4e0d\u518d\u4e00\u6b21\u9810\u7559\u4e00\u5927\u584a\u9023\u7e8c\u7a7a\u9593\uff0c\u6240\u4ee5\u585e\u5f97\u4e0b\u66f4\u591a\u8acb\u6c42\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">KV Cache \u662f\u4ec0\u9ebc\uff1a\u6a21\u578b\u7684\u300c\u8349\u7a3f\u7b46\u8a18\u300d<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5927\u578b\u8a9e\u8a00\u6a21\u578b\u662f\u4e00\u500b token \u4e00\u500b token \u5f80\u5916\u5410\u5b57\u7684\u3002\u6bcf\u5410\u4e00\u500b\u65b0\u5b57\uff0c\u5b83\u90fd\u8981\u300c\u56de\u982d\u770b\u300d\u524d\u9762\u6240\u6709\u7684\u5b57\uff0c\u624d\u77e5\u9053\u63a5\u4e0b\u4f86\u8a72\u8b1b\u4ec0\u9ebc\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u554f\u984c\u4f86\u4e86\uff1a\u5982\u679c\u6bcf\u5410\u4e00\u500b\u5b57\u90fd\u8981\u628a\u524d\u9762\u5168\u90e8\u91cd\u7b97\u4e00\u6b21\uff0c\u8d8a\u5230\u5f8c\u9762\u8d8a\u6162\uff0c\u7b97\u529b\u5168\u6d6a\u8cbb\u5728\u91cd\u8907\u52de\u52d5\u4e0a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u4ee5\u6a21\u578b\u505a\u4e86\u4e00\u4ef6\u5f88\u76f4\u89ba\u7684\u4e8b\uff1a<strong>\u628a\u524d\u9762\u6bcf\u500b token \u7b97\u904e\u7684\u4e2d\u9593\u7d50\u679c\u5b58\u8d77\u4f86\uff0c\u4e0b\u6b21\u76f4\u63a5\u62ff\u4f86\u7528\u3002<\/strong> \u9019\u4efd\u5b58\u8d77\u4f86\u7684\u4e2d\u9593\u7d50\u679c\uff0c\u5c31\u53eb KV Cache\uff08K \u548c V \u662f Key\/Value \u7684\u7e2e\u5beb\uff0c\u4f60\u4e0d\u7528\u7ba1\u5b83\u7684\u6578\u5b78\u610f\u7fa9\uff0c\u628a\u5b83\u60f3\u6210\u300c\u6bcf\u500b token \u7684\u6ce8\u610f\u529b\u7b46\u8a18\u300d\u5c31\u597d\uff09\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528\u767d\u8a71\u985e\u6bd4\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\">\u6c92\u6709 KV Cache\uff1a\n  \u5beb\u7b2c 100 \u500b\u5b57\u6642\uff0c\u628a\u524d\u9762 99 \u500b\u5b57\u5f9e\u982d\u8b80\u4e00\u904d \u2192 \u6162\n\n\u6709 KV Cache\uff1a\n  \u524d\u9762 99 \u500b\u5b57\u7684\u91cd\u9ede\u65e9\u5c31\u8a18\u5728\u4fbf\u5229\u8cbc\u4e0a\u4e86 \u2192 \u76f4\u63a5\u770b\u4fbf\u5229\u8cbc \u2192 \u5feb\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u95dc\u9375\u76f4\u89ba\uff1a\u9019\u4efd\u7b46\u8a18\u6703\u96a8\u8457\u5e8f\u5217\u8b8a\u9577\u800c\u8b8a\u539a\u3002<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li>\u4f60\u7684 prompt \u8d8a\u9577 \u2192 \u7b46\u8a18\u8d8a\u539a<\/li>\n\n\n<li>\u6a21\u578b\u751f\u6210\u7684\u5167\u5bb9\u8d8a\u9577 \u2192 \u7b46\u8a18\u8d8a\u539a<\/li>\n\n\n<li>\u540c\u6642\u670d\u52d9\u7684\u8acb\u6c42\u8d8a\u591a \u2192 \u6bcf\u500b\u8acb\u6c42\u5404\u6709\u4e00\u4efd\u7b46\u8a18\uff0c\u52a0\u8d77\u4f86\u8d8a\u539a<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u5c31\u662f\u70ba\u4ec0\u9ebc GPU \u8a18\u61b6\u9ad4\u6703\u88ab\u5403\u6389\u4e00\u5927\u534a\u2014\u2014<strong>\u4e0d\u662f\u6a21\u578b\u6b0a\u91cd\u5403\u7684\uff0c\u662f\u9019\u4e9b\u7b46\u8a18\u5403\u7684\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u8a18\u61b6\u9ad4\u5927\u6982\u600e\u9ebc\u5206\u914d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u9846\u8dd1 LLM \u7684 GPU\uff0c\u8a18\u61b6\u9ad4\u5927\u81f4\u5206\u4e09\u584a\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\">GPU \u8a18\u61b6\u9ad4\n\u251c\u2500\u2500 \u6a21\u578b\u6b0a\u91cd         \u2190 \u56fa\u5b9a\u7684\uff0c\u6a21\u578b\u591a\u5927\u5c31\u591a\u5927\uff08\u4f8b\u5982 14GB\uff09\n\u251c\u2500\u2500 KV Cache        \u2190 \u6703\u81a8\u8139\u7684\uff0c\u8acb\u6c42\u8d8a\u591a\u3001\u5e8f\u5217\u8d8a\u9577\u5403\u8d8a\u591a  \u2190 \u4e3b\u89d2\n\u2514\u2500\u2500 \u5176\u4ed6\u96dc\u9805         \u2190 \u6846\u67b6\u672c\u8eab\u3001\u66ab\u5b58\u7a7a\u9593\u7b49\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u4f60\u7684 24GB GPU \u8dd1 14GB \u7684\u6a21\u578b\uff0c\u6263\u6389\u6b0a\u91cd\u548c\u96dc\u9805\uff0c\u5269\u4e0b\u7684\u300c\u624d\u662f\u300d\u80fd\u62ff\u4f86\u653e KV Cache \u7684\u7a7a\u9593\u3002vLLM \u6703\u628a\u9019\u584a\u7a7a\u9593\u62ff\u53bb\u5207\u6210\u4e00\u683c\u4e00\u683c\u7684 block\uff08\u4e0b\u4e00\u6bb5\u8b1b\uff09\uff0c\u80fd\u585e\u5e7e\u683c\uff0c\u5c31\u6c7a\u5b9a\u4e86\u4f60\u80fd\u540c\u6642\u670d\u52d9\u591a\u5c11\u8acb\u6c42\u3001\u6bcf\u500b\u8acb\u6c42\u80fd\u591a\u9577\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">PagedAttention\uff1a\u50cf\u4f5c\u696d\u7cfb\u7d71\u4e00\u6a23\u7ba1\u7b46\u8a18<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u77e5\u9053 KV Cache \u6703\u81a8\u8139\u4e4b\u5f8c\uff0c\u4e0b\u4e00\u500b\u554f\u984c\u662f\uff1a<strong>\u600e\u9ebc\u7ba1\u7406\u9019\u5806\u6703\u81a8\u8139\u7684\u7b46\u8a18\uff1f<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u820a\u505a\u6cd5\u7684\u6d6a\u8cbb<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728 PagedAttention \u51fa\u73fe\u524d\uff0c\u50b3\u7d71\u505a\u6cd5\u662f\uff1a\u4e00\u500b\u8acb\u6c42\u9032\u4f86\uff0c\u5c31\u5148\u300c\u9810\u7559\u300d\u4e00\u5927\u584a\u9023\u7e8c\u8a18\u61b6\u9ad4\uff0c\u9810\u8a2d\u5b83\u53ef\u80fd\u6703\u9577\u5230\u6700\u5927\u9577\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u554f\u984c\u662f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li>\u5927\u90e8\u5206\u8acb\u6c42\u6839\u672c\u7528\u4e0d\u5230\u6700\u5927\u9577\u5ea6\uff0c\u9810\u7559\u7684\u7a7a\u9593\u5c31\u9019\u6a23\u7a7a\u8457\u6d6a\u8cbb\uff08\u5167\u90e8\u788e\u7247\uff09<\/li>\n\n\n<li>\u5269\u4e0b\u7684\u96f6\u788e\u7a7a\u9593\u6771\u4e00\u584a\u897f\u4e00\u584a\uff0c\u6e4a\u4e0d\u51fa\u4e0b\u4e00\u500b\u8acb\u6c42\u8981\u7684\u300c\u9023\u7e8c\u300d\u5927\u584a\uff08\u5916\u90e8\u788e\u7247\uff09<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u7d50\u679c\u5c31\u662f\uff1a\u8a18\u61b6\u9ad4\u660e\u660e\u9084\u6709\uff0c\u537b\u585e\u4e0d\u9032\u65b0\u8acb\u6c42\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">PagedAttention \u7684\u9ede\u5b50<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vLLM \u7684\u4f5c\u8005\u501f\u7528\u4e86\u4e00\u500b\u4f5c\u696d\u7cfb\u7d71\u7ba1\u4e86\u5e7e\u5341\u5e74\u8a18\u61b6\u9ad4\u7684\u8001\u62db\u6578\uff1a<strong>\u5206\u9801\uff08paging\uff09<\/strong>\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u696d\u7cfb\u7d71\u4e0d\u6703\u7d66\u7a0b\u5f0f\u4e00\u6574\u584a\u9023\u7e8c\u7684\u5be6\u9ad4\u8a18\u61b6\u9ad4\uff0c\u800c\u662f\u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u300c\u9801\u300d\uff0c\u7a0b\u5f0f\u8981\u591a\u5c11\u7d66\u591a\u5c11\u9801\uff0c\u9019\u4e9b\u9801\u5728\u5be6\u9ad4\u8a18\u61b6\u9ad4\u88e1\u53ef\u4ee5\u662f\u6563\u7684\uff0c\u9760\u4e00\u5f35\u8868\u5c0d\u61c9\u8d77\u4f86\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">PagedAttention \u5c0d KV Cache \u505a\u4e00\u6a21\u4e00\u6a23\u7684\u4e8b\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\">\u50b3\u7d71\u505a\u6cd5\uff1a\n  \u8acb\u6c42A \u2593\u2593\u2593\u2593\u2593\u2593\u2593\u2593\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591  \u2190 \u9810\u7559\u4e00\u5927\u584a\uff0c\u5f8c\u9762\u4e00\u534a\u7a7a\u8457\u6d6a\u8cbb\n  \u8acb\u6c42B \u2593\u2593\u2593\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591\u2591  \u2190 \u53c8\u9810\u7559\u4e00\u5927\u584a\n\nPagedAttention\uff1a\n  \u628a\u8a18\u61b6\u9ad4\u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684\u5c0f\u683c\u5b50\uff08block\uff09\n  \u250c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u252c\u2500\u2500\u2510\n  \u2502A \u2502A \u2502A \u2502B \u2502A \u2502B \u2502  \u2502  \u2502  \u2190 \u8ab0\u8981\u7528\u5c31\u7d66\u4e00\u683c\uff0c\u7528\u5b8c\u518d\u7d66\n  \u2514\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2534\u2500\u2500\u2518\n  \u683c\u5b50\u53ef\u4ee5\u6563\u8457\u653e\uff0c\u9760\u4e00\u5f35\u8868\u4e32\u8d77\u4f86 \u2192 \u5e7e\u4e4e\u4e0d\u6d6a\u8cbb\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u4f60\u53ea\u8981\u8a18\u4f4f\u4e09\u500b\u597d\u8655\uff1a<\/strong><\/p>\n\n\n\n<ol class=\"wp-block-list\">\n\n<li><strong>\u4e0d\u9810\u7559\u3001\u7528\u591a\u5c11\u7d66\u591a\u5c11<\/strong> \u2192 \u8a18\u61b6\u9ad4\u6d6a\u8cbb\u5927\u5e45\u6e1b\u5c11<\/li>\n\n\n<li><strong>\u788e\u7247\u5e7e\u4e4e\u6d88\u5931<\/strong> \u2192 \u540c\u6a23\u7684 GPU \u80fd\u585e\u4e0b\u66f4\u591a\u8acb\u6c42\uff08\u66f4\u9ad8\u4f75\u767c\uff09<\/li>\n\n\n<li><strong>\u683c\u5b50\u53ef\u4ee5\u5171\u4eab<\/strong> \u2192 \u5169\u500b\u8acb\u6c42\u5982\u679c\u958b\u982d\u4e00\u6a23\uff08\u4f8b\u5982\u540c\u4e00\u6bb5 system prompt\uff09\uff0c\u53ef\u4ee5\u5171\u7528\u540c\u4e00\u6279\u683c\u5b50\uff08\u9019\u53eb prefix \u5171\u4eab\uff0c\u662f\u7b2c 4 \u7bc7\u7684\u4e3b\u89d2\uff0c\u9019\u88e1\u5148\u57cb\u500b\u4f0f\u7b46\uff09<\/li>\n\n<\/ol>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u300c\u56fa\u5b9a\u5927\u5c0f\u7684\u5c0f\u683c\u5b50\u300d\uff0c\u5728 vLLM \u88e1\u5c31\u53eb <strong>block<\/strong>\uff0c\u5927\u5c0f\u7531 <code>--block-size<\/code> \u63a7\u5236\uff08\u9810\u8a2d\u901a\u5e38\u662f 16\uff0c\u8868\u793a\u4e00\u683c\u88dd 16 \u500b token \u7684\u7b46\u8a18\uff09\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\u61c2 vLLM \u555f\u52d5\u65e5\u8a8c<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7406\u89e3\u4e86\u4e0a\u9762\u7684\u6982\u5ff5\uff0c\u73fe\u5728\u4f86\u770b vLLM \u555f\u52d5\u6642\u5370\u51fa\u4f86\u7684\u90a3\u5e7e\u884c\u95dc\u9375\u65e5\u8a8c\u3002\u9019\u662f\u4f60<strong>\u6bcf\u6b21\u4e0a\u7dda\u90fd\u8a72\u6383\u4e00\u773c<\/strong>\u7684\u8cc7\u8a0a\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\">INFO ... Memory profiling results: ... available_kv_cache_memory=8.42GiB\nINFO ... GPU KV cache size: 74,096 tokens\nINFO ... Maximum concurrency for 8,192 tokens per request: 9.04x\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\">\u4e00\u884c\u4e00\u884c\u7ffb\u8b6f\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-1\" data-shcb-language-name=\"JavaScript\" data-shcb-language-slug=\"javascript\"><span><code class=\"hljs language-javascript\">available_kv_cache_memory=<span class=\"hljs-number\">8.42<\/span>GiB\n  \u2192 \u6263\u6389\u6a21\u578b\u6b0a\u91cd\u548c\u96dc\u9805\u5f8c\uff0c\u9084\u6709 <span class=\"hljs-number\">8.42<\/span>GB \u62ff\u4f86\u653e KV Cache \u7b46\u8a18\n\nGPU KV cache size: <span class=\"hljs-number\">74<\/span>,<span class=\"hljs-number\">096<\/span> tokens\n  \u2192 \u9019 <span class=\"hljs-number\">8.42<\/span>GB \u63db\u7b97\u4e0b\u4f86\uff0c\u7e3d\u5171\u80fd\u5b58 <span class=\"hljs-number\">74<\/span>,<span class=\"hljs-number\">096<\/span> \u500b token \u7684\u7b46\u8a18\n  \u2192 \uff08\u6709\u4e9b\u7248\u672c\u6703\u5370\u6210 <span class=\"hljs-string\">\"# GPU blocks: 4631\"<\/span> \u4e4b\u985e\uff0cblock \u6578 \u00d7 block-size \u5c31\u662f\u9019\u500b token \u6578\uff09\n\nMaximum concurrency <span class=\"hljs-keyword\">for<\/span> <span class=\"hljs-number\">8<\/span>,<span class=\"hljs-number\">192<\/span> tokens per request: <span class=\"hljs-number\">9.04<\/span>x\n  \u2192 \u5982\u679c\u6bcf\u500b\u8acb\u6c42\u90fd\u5403\u6eff <span class=\"hljs-number\">8<\/span>,<span class=\"hljs-number\">192<\/span> token\uff08= \u4f60\u8a2d\u7684 max-model-len\uff09\uff0c\n    \u90a3\u5927\u7d04\u80fd\u540c\u6642\u585e <span class=\"hljs-number\">9<\/span> \u500b\u8acb\u6c42\n  \u2192 <span class=\"hljs-number\">74<\/span>,<span class=\"hljs-number\">096<\/span> \u00f7 <span class=\"hljs-number\">8<\/span>,<span class=\"hljs-number\">192<\/span> \u2248 <span class=\"hljs-number\">9.04<\/span>\n<\/code><\/span><small class=\"shcb-language\" id=\"shcb-language-1\"><span class=\"shcb-language__label\">Code language:<\/span> <span class=\"shcb-language__name\">JavaScript<\/span> <span class=\"shcb-language__paren\">(<\/span><span class=\"shcb-language__slug\">javascript<\/span><span class=\"shcb-language__paren\">)<\/span><\/small><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u9019\u4e09\u884c\u5c31\u662f\u4f60\u7684\u5bb9\u91cf\u5100\u8868\u677f\u3002<\/strong> \u89ba\u5f97\u4f75\u767c\u4e0d\u5920\uff1f\u5c31\u662f\u8981\u60f3\u8fa6\u6cd5\u628a <code>GPU KV cache size<\/code> \u9019\u500b\u6578\u5b57\u505a\u5927\uff0c\u6216\u628a\u6bcf\u500b\u8acb\u6c42\u7684 <code>max-model-len<\/code> \u505a\u5c0f\u3002\u600e\u9ebc\u8abf\uff0c\u770b\u4e0b\u4e00\u6bb5\u3002<\/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\">\u8a8d\u5f97\u5c31\u597d\uff1a\u4e0d\u540c vLLM \u7248\u672c\u9019\u5e7e\u884c\u7684\u7528\u5b57\u6703\u7565\u6709\u4e0d\u540c\uff08\u6709\u7684\u5370 <code># GPU blocks<\/code>\u3001\u6709\u7684\u5370 <code>GPU KV cache size<\/code>\uff09\uff0c\u4f46\u610f\u601d\u90fd\u662f\u300c\u7e3d\u5171\u80fd\u5b58\u591a\u5c11\u7b46\u8a18\u300d\u548c\u300c\u80fd\u540c\u6642\u670d\u52d9\u5e7e\u500b\u8acb\u6c42\u300d\u3002\u6293\u9019\u5169\u500b\u6578\u5b57\u5c31\u5c0d\u4e86\u3002<\/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\">\u5be6\u52d9\u8abf\u53c3\uff1a\u56db\u500b\u4f60\u6703\u52d5\u5230\u7684\u53c3\u6578<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ee5\u4e0b\u56db\u500b\u662f\u4f60 90% \u60c5\u6cc1\u6703\u7528\u5230\u7684\u53c3\u6578\u3002\u5148\u770b\u5b83\u5011\u5404\u81ea\u5728\u5e79\u561b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead>\n<tr>\n<th>\u53c3\u6578<\/th>\n<th>\u767d\u8a71\u610f\u601d<\/th>\n<th>\u8abf\u5927\u6703\u600e\u6a23<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><code>--gpu-memory-utilization<\/code><\/td>\n<td>\u5141\u8a31 vLLM \u7528\u6389 GPU \u8a18\u61b6\u9ad4\u7684\u6bd4\u4f8b\uff08\u9810\u8a2d 0.9\uff09<\/td>\n<td>\u66f4\u591a\u7a7a\u9593\u7d66 KV Cache\uff0c\u4f46\u592a\u8caa\u5fc3\u6703 OOM<\/td>\n<\/tr>\n<tr>\n<td><code>--max-model-len<\/code><\/td>\n<td>\u55ae\u4e00\u8acb\u6c42\u6700\u9577\u80fd\u5230\u591a\u5c11 token\uff08prompt + \u751f\u6210\uff09<\/td>\n<td>\u6bcf\u500b\u8acb\u6c42\u4e0a\u9650\u66f4\u9ad8\uff0c\u4f46\u5403\u66f4\u591a\u7b46\u8a18\u7a7a\u9593<\/td>\n<\/tr>\n<tr>\n<td><code>--max-num-seqs<\/code><\/td>\n<td>\u6700\u591a\u540c\u6642\u8655\u7406\u5e7e\u500b\u8acb\u6c42<\/td>\n<td>\u4f75\u767c\u66f4\u9ad8\uff0c\u4f46\u6bcf\u500b\u8acb\u6c42\u5206\u5230\u7684\u7a7a\u9593\u8b8a\u5c11<\/td>\n<\/tr>\n<tr>\n<td><code>--block-size<\/code><\/td>\n<td>\u4e00\u683c block \u88dd\u5e7e\u500b token \u7684\u7b46\u8a18\uff08\u9810\u8a2d 16\uff09<\/td>\n<td>\u901a\u5e38\u4e0d\u7528\u52d5\uff0c\u7528\u9810\u8a2d\u5c31\u597d<\/td>\n<\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--gpu-memory-utilization<\/code>\uff1a\u7d66 vLLM \u591a\u5927\u7684\u7897<\/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=\"CSS\" data-shcb-language-slug=\"css\"><span><code class=\"hljs language-css\"><span class=\"hljs-selector-tag\">vllm<\/span> <span class=\"hljs-selector-tag\">serve<\/span> <span class=\"hljs-selector-tag\">my-model<\/span> <span class=\"hljs-selector-tag\">--gpu-memory-utilization<\/span> 0<span class=\"hljs-selector-class\">.9<\/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\">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>\u7ffb\u8b6f<\/strong>\uff1a\u300c\u4f60\u6700\u591a\u53ef\u4ee5\u7528\u6389\u9019\u5f35 GPU 90% \u7684\u8a18\u61b6\u9ad4\u300d\u3002\u5269\u4e0b\u7684 10% \u7559\u7d66\u7cfb\u7d71\u548c\u5b89\u5168\u7de9\u885d\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li>\u8abf<strong>\u9ad8<\/strong>\uff08\u4f8b\u5982 0.95\uff09\u2192 \u66f4\u591a\u7a7a\u9593\u653e KV Cache \u2192 \u4f75\u767c\u66f4\u9ad8\uff0c\u4f46<strong>\u592a\u8caa\u5fc3\u5bb9\u6613 OOM<\/strong>\uff0c\u5c24\u5176\u9019\u5f35\u5361\u4e0a\u9084\u6709\u5225\u7684\u7a0b\u5f0f\u5728\u8dd1\u6642<\/li>\n\n\n<li>\u8abf<strong>\u4f4e<\/strong>\uff08\u4f8b\u5982 0.8\uff09\u2192 \u6bd4\u8f03\u5b89\u5168\uff0c\u4f46\u80fd\u585e\u7684\u8acb\u6c42\u8b8a\u5c11<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u662f\u4f60<strong>\u8a18\u61b6\u9ad4\u4e0d\u5920\u6642\u7b2c\u4e00\u500b\u8a72\u6aa2\u67e5<\/strong>\u7684\u53c3\u6578\uff1a\u662f\u4e0d\u662f\u8a2d\u592a\u9ad8\uff0c\u628a\u7cfb\u7d71\u9700\u8981\u7684\u9918\u88d5\u4e5f\u5403\u6389\u4e86\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--max-model-len<\/code>\uff1a\u6bcf\u500b\u8acb\u6c42\u7684\u9577\u5ea6\u4e0a\u9650<\/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\">vllm serve my-model --max-model-len 8192\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7ffb\u8b6f<\/strong>\uff1a\u300c\u6bcf\u500b\u8acb\u6c42\uff08prompt + \u751f\u6210\uff09\u6700\u9577 8192 \u500b token\uff0c\u8d85\u904e\u5c31\u62d2\u7d55\u300d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u53c3\u6578\u5c0d\u8a18\u61b6\u9ad4\u7684\u5f71\u97ff\u5f88\u76f4\u63a5\uff1a\u56de\u982d\u770b\u525b\u525b\u7684\u65e5\u8a8c\uff0c<code>\u6700\u5927\u4f75\u767c = KV cache \u7e3d\u91cf \u00f7 max-model-len<\/code>\u3002\u6240\u4ee5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><code>max-model-len<\/code> \u8a2d<strong>\u5c0f<\/strong> \u2192 \u6bcf\u500b\u8acb\u6c42\u4f54\u7684\u7a7a\u9593\u4e0a\u9650\u8b8a\u5c0f \u2192 \u80fd\u585e\u66f4\u591a\u8acb\u6c42<\/li>\n\n\n<li><code>max-model-len<\/code> \u8a2d<strong>\u5927<\/strong> \u2192 \u652f\u63f4\u8d85\u9577 prompt\uff0c\u4f46\u4f75\u767c\u76f4\u63a5\u88ab\u58d3\u4e0b\u4f86<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5be6\u52d9\u5fc3\u6cd5<\/strong>\uff1a\u4e0d\u8981\u7121\u8166\u8a2d\u6210\u6a21\u578b\u652f\u63f4\u7684\u6700\u5927\u503c\uff08\u6709\u4e9b\u6a21\u578b\u5ba3\u7a31 128K\uff09\u3002\u4f60\u7684\u7528\u6236\u5be6\u969b\u7528\u4e0d\u5230\u90a3\u9ebc\u9577\uff0c\u8a2d\u592a\u5927\u53ea\u662f\u767d\u767d\u58d3\u4f4e\u4f75\u767c\u3001\u9084\u53ef\u80fd\u4e00\u555f\u52d5\u5c31 OOM\u3002<strong>\u8a2d\u6210\u4f60\u5be6\u969b\u9700\u8981\u7684\u9577\u5ea6\u5c31\u597d\u3002<\/strong><\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--max-num-seqs<\/code>\uff1a\u540c\u6642\u670d\u52d9\u5e7e\u500b\u8acb\u6c42<\/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\">vllm serve my-model --max-num-seqs 256\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7ffb\u8b6f<\/strong>\uff1a\u300c\u6700\u591a\u540c\u6642\u8655\u7406 256 \u500b\u8acb\u6c42\uff0c\u591a\u7684\u6392\u968a\u7b49\u300d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u662f\u4f60<strong>\u7528\u4f75\u767c\u63db\u7a69\u5b9a<\/strong>\u7684\u65cb\u9215\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li>\u8abf<strong>\u9ad8<\/strong> \u2192 \u541e\u5410\u91cf\u66f4\u9ad8\uff0c\u4f46\u6bcf\u500b\u8acb\u6c42\u53ef\u7528\u7684 KV Cache \u7a7a\u9593\u88ab\u7a00\u91cb\uff0c\u9577\u8acb\u6c42\u53ef\u80fd\u88ab\u4e2d\u9014\u6436\u4f54\uff08preemption\uff09\u800c\u8b8a\u6162<\/li>\n\n\n<li>\u8abf<strong>\u4f4e<\/strong> \u2192 \u6bcf\u500b\u8acb\u6c42\u66f4\u5bec\u88d5\u3001\u66f4\u7a69\uff0c\u4f46\u6574\u9ad4\u541e\u5410\u4e0b\u964d<\/li>\n\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\"><code>--block-size<\/code>\uff1a\u5148\u5225\u52d5\u5b83<\/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\">vllm serve my-model --block-size 16\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u7ffb\u8b6f<\/strong>\uff1a\u300c\u4e00\u683c block \u88dd 16 \u500b token \u7684\u7b46\u8a18\u300d\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u662f PagedAttention \u90a3\u500b\u300c\u56fa\u5b9a\u5927\u5c0f\u683c\u5b50\u300d\u7684\u5927\u5c0f\u3002<strong>\u65b0\u624b\u76f4\u63a5\u7528\u9810\u8a2d\u5c31\u597d<\/strong>\uff0c\u9664\u975e\u4f60\u5728\u505a\u9032\u968e\u6548\u80fd\u8abf\u6821\uff0c\u5426\u5247\u52d5\u5b83\u901a\u5e38\u5f0a\u5927\u65bc\u5229\u3002\u8a8d\u5f97\u5b83\u662f\u4ec0\u9ebc\u3001\u77e5\u9053\u5b83\u8ddf PagedAttention \u6709\u95dc\uff0c\u5c31\u5920\u4e86\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">OOM \u4e86\u600e\u9ebc\u8fa6\uff1a\u6392\u67e5\u9806\u5e8f<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><code>torch.OutOfMemoryError: CUDA out of memory<\/code> \u662f\u4f60\u6700\u5e38\u649e\u5230\u7684\u932f\u8aa4\u3002\u6309\u9019\u500b\u9806\u5e8f\u8655\u7406\uff0c\u4e0d\u8981\u4e82\u69cd\u6253\u9ce5\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\">1. \u5148\u964d --max-model-len\n   \u2192 \u6700\u6709\u6548\u3002\u6bcf\u500b\u8acb\u6c42\u7684\u7b46\u8a18\u4e0a\u9650\u76f4\u63a5\u8b8a\u5c0f\uff0c\u99ac\u4e0a\u7701\u4e00\u5927\u584a\n   \u2192 \u4f8b\u5982\u5f9e 32768 \u964d\u5230 8192\n\n2. \u518d\u964d --gpu-memory-utilization\n   \u2192 \u5982\u679c\u662f\u300c\u4e00\u555f\u52d5\u5c31 OOM\u300d\u6216\u9019\u5f35\u5361\u6709\u5225\u7684\u7a0b\u5f0f\uff0c\n     \u5f9e 0.9 \u964d\u5230 0.85 \/ 0.8 \u7559\u591a\u9ede\u9918\u88d5\n\n3. \u7136\u5f8c\u964d --max-num-seqs\n   \u2192 \u5982\u679c\u662f\u300c\u8dd1\u4e00\u9663\u5b50\u3001\u4f75\u767c\u885d\u9ad8\u624d OOM\u300d\uff0c\u628a\u540c\u6642\u8acb\u6c42\u6578\u58d3\u4e0b\u4f86\n\n4. \u90fd\u4e0d\u884c \u2192 \u63db\u66f4\u5c0f\u7684\u6a21\u578b \u6216 \u958b\u91cf\u5316\u7248\u672c\uff08\u4f8b\u5982 AWQ\/GPTQ\uff09\n   \u2192 \u6b0a\u91cd\u672c\u8eab\u5c31\u4f54\u592a\u591a\uff0c\u524d\u4e09\u6b65\u6551\u4e0d\u56de\u4f86\n<\/code><\/span><\/pre><\/div>\n\n\n<p class=\"wp-block-paragraph\"><strong>\u5224\u65b7\u6280\u5de7<\/strong>\uff1a\u770b OOM \u662f<strong>\u4ec0\u9ebc\u6642\u5019\u767c\u751f<\/strong>\u7684\u2014\u2014<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>\u4e00\u555f\u52d5\u5c31 OOM<\/strong> \u2192 \u6b0a\u91cd + \u4fdd\u7559\u7a7a\u9593\u5c31\u8d85\u4e86\uff0c\u52d5 <code>gpu-memory-utilization<\/code> \u6216\u63db\u5c0f\u6a21\u578b<\/li>\n\n\n<li><strong>\u8dd1\u4e00\u9663\u5b50\u3001\u6d41\u91cf\u4e0a\u4f86\u624d OOM<\/strong> \u2192 \u662f KV Cache \u81a8\u8139\u6490\u7206\u7684\uff0c\u52d5 <code>max-model-len<\/code> \u6216 <code>max-num-seqs<\/code><\/li>\n\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u70ba\u4ec0\u9ebc vLLM \u80fd\u670d\u52d9\u66f4\u591a\u8acb\u6c42<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7e5e\u4e86\u4e00\u5708\uff0c\u56de\u5230\u6700\u958b\u59cb\u7684\u76f4\u89ba\u3002\u540c\u4e00\u5f35 GPU\uff0c\u70ba\u4ec0\u9ebc vLLM \u80fd\u6bd4\u571f\u6cd5\u7149\u92fc\u585e\u4e0b\u66f4\u591a\u4f75\u767c\u8acb\u6c42\uff1f<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u70ba PagedAttention \u628a KV Cache \u9019\u5806\u300c\u6703\u81a8\u8139\u7684\u7b46\u8a18\u300d\u7ba1\u5f97\u53c8\u7701\u53c8\u7dca\uff1a\u4e0d\u9810\u7559\u3001\u4e0d\u6d6a\u8cbb\u3001\u788e\u7247\u5e7e\u4e4e\u70ba\u96f6\uff0c\u7528\u591a\u5c11\u7d66\u591a\u5c11\u3002\u7701\u4e0b\u4f86\u7684\u6bcf\u4e00\u683c\u8a18\u61b6\u9ad4\uff0c\u90fd\u80fd\u62ff\u53bb\u670d\u52d9\u4e0b\u4e00\u500b\u8acb\u6c42\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u800c\u90a3\u500b\u300c\u683c\u5b50\u53ef\u4ee5\u5171\u4eab\u300d\u7684\u4f0f\u7b46\u2014\u2014\u7576\u5f88\u591a\u8acb\u6c42\u90fd\u7528\u540c\u4e00\u6bb5\u958b\u982d\uff08\u4f8b\u5982\u540c\u4e00\u500b system prompt\u3001\u540c\u4e00\u4efd few-shot \u7bc4\u4f8b\uff09\u6642\uff0c\u9019\u6279\u7b46\u8a18\u53ea\u5b58\u4e00\u4efd\u3001\u5927\u5bb6\u5171\u7528\u2014\u2014\u5c31\u662f <strong>prefix \u5171\u4eab \/ prefix caching<\/strong>\uff0c\u80fd\u518d\u9032\u4e00\u6b65\u7701\u8a18\u61b6\u9ad4\u3001\u52a0\u901f\u56de\u61c9\u3002\u9019\u662f\u7b2c 4 \u7bc7\u7684\u4e3b\u984c\uff0c\u5230\u6642\u5019\u518d\u5c55\u958b\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 \u7d05\u65d7\uff1a\u9019\u4e9b\u8a2d\u5b9a\u770b\u5230\u8981\u8b66\u89ba<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d05\u65d7 1\uff1a<code>--gpu-memory-utilization<\/code> \u8a2d\u5230 0.98 \u751a\u81f3 1.0<\/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-3\" data-shcb-language-name=\"CSS\" data-shcb-language-slug=\"css\"><span><code class=\"hljs language-css\"><span class=\"hljs-selector-tag\">vllm<\/span> <span class=\"hljs-selector-tag\">serve<\/span> <span class=\"hljs-selector-tag\">my-model<\/span> <span class=\"hljs-selector-tag\">--gpu-memory-utilization<\/span> 0<span class=\"hljs-selector-class\">.99<\/span>   # \ud83d\udea9 \u6c92\u7559\u9918\u88d5\uff0c\u96a8\u6642\u6703 <span class=\"hljs-selector-tag\">OOM<\/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\">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>\u70ba\u4ec0\u9ebc\u5371\u96aa<\/strong>\uff1aGPU \u8a18\u61b6\u9ad4\u4e0d\u662f\u53ea\u6709 vLLM \u5728\u7528\uff0c\u6846\u67b6\u66ab\u5b58\u3001CUDA context\u3001\u5cf0\u503c\u6ce2\u52d5\u90fd\u9700\u8981\u9918\u88d5\u3002\u8a2d\u592a\u6eff\uff0c\u5e73\u5e38\u6c92\u4e8b\uff0c\u4e00\u9047\u6d41\u91cf\u5c16\u5cf0\u5c31 OOM \u639b\u6389\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8ddf AI \u9019\u6a23\u8aaa<\/strong>\uff1a\u300c\u9019\u500b <code>--gpu-memory-utilization 0.99<\/code> \u592a\u8caa\u5fc3\u4e86\uff0c\u6539\u6210\u4fdd\u5b88\u7684 0.9 \u4e26\u89e3\u91cb\u70ba\u4ec0\u9ebc\u8981\u7559\u9918\u88d5\u3002\u300d<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d05\u65d7 2\uff1a<code>--max-model-len<\/code> \u7121\u8166\u62c9\u5230\u6a21\u578b\u6700\u5927\u503c<\/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-4\" data-shcb-language-name=\"PHP\" data-shcb-language-slug=\"php\"><span><code class=\"hljs language-php\">vllm serve my-model --max-model-len <span class=\"hljs-number\">131072<\/span>   <span class=\"hljs-comment\"># \ud83d\udea9 \u7528\u4e0d\u5230\u537b\u628a\u4f75\u767c\u58d3\u5230\u5269 1~2 \u500b<\/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>\u70ba\u4ec0\u9ebc\u5371\u96aa<\/strong>\uff1a\u628a\u9577\u5ea6\u4e0a\u9650\u8a2d\u6210 128K\uff0c\u4ee3\u8868 vLLM \u8981\u70ba\u300c\u842c\u4e00\u6709\u8d85\u9577\u8acb\u6c42\u300d\u9810\u7559\u898f\u5283\u7a7a\u9593\uff0c\u7d50\u679c\u4f75\u767c\u76f4\u63a5\u5d29\u5230\u5269\u500b\u4f4d\u6578\uff0c\u751a\u81f3\u4e00\u555f\u52d5\u5c31 OOM\u3002\u591a\u6578\u5834\u666f\u6839\u672c\u7528\u4e0d\u5230\u9019\u9ebc\u9577\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8ddf AI \u9019\u6a23\u8aaa<\/strong>\uff1a\u300c\u6211\u5be6\u969b\u8acb\u6c42\u6700\u9577\u5927\u6982 X token\uff0c\u628a <code>--max-model-len<\/code> \u8a2d\u6210\u8cbc\u8fd1\u5be6\u969b\u9700\u6c42\u7684\u503c\uff0c\u4e0d\u8981\u7528\u6a21\u578b\u4e0a\u9650\uff0c\u4e26\u8aaa\u660e\u9019\u6a23\u5c0d\u4f75\u767c\u7684\u5f71\u97ff\u3002\u300d<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u7d05\u65d7 3\uff1a\u53ea\u52a0\u53c3\u6578\u537b\u4e0d\u770b\u555f\u52d5\u65e5\u8a8c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u70ba\u4ec0\u9ebc\u5371\u96aa<\/strong>\uff1a\u8abf\u5b8c\u53c3\u6578\u5982\u679c\u4e0d\u56de\u982d\u770b <code>Maximum concurrency<\/code> \u548c <code>GPU KV cache size<\/code> \u9019\u5169\u884c\uff0c\u4f60\u6839\u672c\u4e0d\u77e5\u9053\u9019\u6b21\u8abf\u6574\u5230\u5e95\u8b93\u4f75\u767c\u8b8a\u9ad8\u9084\u8b8a\u4f4e\uff0c\u7b49\u65bc\u778e\u8abf\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u8ddf AI \u9019\u6a23\u8aaa<\/strong>\uff1a\u300c\u5e6b\u6211\u628a vLLM \u555f\u52d5\u65e5\u8a8c\u88e1\u8ddf KV cache\u3001max concurrency \u6709\u95dc\u7684\u884c\u627e\u51fa\u4f86\uff0c\u89e3\u91cb\u9019\u6b21\u53c3\u6578\u6539\u52d5\u8b93\u6578\u5b57\u5f80\u54ea\u500b\u65b9\u5411\u8b8a\u4e86\u3002\u300d<\/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<ul class=\"wp-block-list\">\n\n<li>[ ] \u555f\u52d5 vLLM \u5f8c\uff0c\u5728\u65e5\u8a8c\u88e1\u627e\u5230 <code>GPU KV cache size<\/code>\uff08\u6216 <code># GPU blocks<\/code>\uff09\u548c <code>Maximum concurrency<\/code> \u9019\u5169\u884c\uff0c\u8b80\u51fa\u300c\u7e3d\u5171\u80fd\u5b58\u5e7e\u500b token\u300d\u548c\u300c\u5927\u6982\u80fd\u540c\u6642\u670d\u52d9\u5e7e\u500b\u8acb\u6c42\u300d\u3002<\/li>\n\n\n<li>[ ] \u628a <code>--max-model-len<\/code> \u5f9e\u5927\u503c\u6539\u6210\u5c0f\u503c\uff08\u4f8b\u5982 32768 \u2192 8192\uff09\u91cd\u555f\u4e00\u6b21\uff0c\u89c0\u5bdf <code>Maximum concurrency<\/code> \u90a3\u884c\u7684\u500d\u6578\u662f\u4e0d\u662f\u8b8a\u5927\u4e86\u3002<\/li>\n\n\n<li>[ ] \u6545\u610f\u628a <code>--gpu-memory-utilization<\/code> \u8a2d 0.99 \u8a66\u8dd1\uff0c\u770b\u662f\u5426\u66f4\u5bb9\u6613 OOM \u6216\u555f\u52d5\u5931\u6557\uff0c\u7406\u89e3\u300c\u7559\u9918\u88d5\u300d\u7684\u610f\u7fa9\uff08\u6e2c\u5b8c\u8a18\u5f97\u6539\u56de 0.9\uff09\u3002<\/li>\n\n\n<li>[ ] \u554f AI\uff1a\u300c\u4ee5\u6211\u7684 GPU \u8a18\u61b6\u9ad4\u5927\u5c0f\u548c\u9019\u500b\u6a21\u578b\uff0c<code>--max-model-len<\/code> \u548c <code>--max-num-seqs<\/code> \u5efa\u8b70\u8a2d\u591a\u5c11\uff1f\u8acb\u89e3\u91cb\u4f60\u600e\u9ebc\u5f9e KV cache \u7a7a\u9593\u63a8\u7b97\u7684\u3002\u300d<\/li>\n\n\n<li>[ ] \u649e\u5230 OOM \u6642\uff0c\u5148\u78ba\u8a8d\u932f\u8aa4\u662f\u300c\u4e00\u555f\u52d5\u300d\u9084\u662f\u300c\u6d41\u91cf\u4e0a\u4f86\u5f8c\u300d\u767c\u751f\u7684\uff0c\u518d\u5c0d\u7167\u6392\u67e5\u9806\u5e8f\u6c7a\u5b9a\u5148\u52d5\u54ea\u500b\u53c3\u6578\u3002<\/li>\n\n<\/ul>\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<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u300c\u6211\u5728\u8dd1 vLLM\uff0c\u8cbc\u4e0a\u555f\u52d5\u65e5\u8a8c\u3002\u8acb\u7528\u767d\u8a71\u89e3\u91cb <code>GPU KV cache size<\/code>\u3001<code>Maximum concurrency<\/code>\u3001<code>available_kv_cache_memory<\/code> \u9019\u5e7e\u884c\u5404\u4ee3\u8868\u4ec0\u9ebc\uff0c\u6211\u662f\u521d\u5b78\u8005\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\u7684 GPU \u662f __GB\uff0c\u8981\u8dd1 __ \u9019\u500b\u6a21\u578b\uff0c\u5be6\u969b\u8acb\u6c42\u6700\u9577\u5927\u7d04 __ token\u3002\u8acb\u5e6b\u6211\u5efa\u8b70 <code>--gpu-memory-utilization<\/code>\u3001<code>--max-model-len<\/code>\u3001<code>--max-num-seqs<\/code> \u8a72\u8a2d\u591a\u5c11\uff0c\u4e26\u89e3\u91cb\u6bcf\u500b\u503c\u600e\u9ebc\u5f71\u97ff\u8a18\u61b6\u9ad4\u548c\u4f75\u767c\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\u9047\u5230 CUDA out of memory\uff0c\u932f\u8aa4\u767c\u751f\u5728 __\uff08\u4e00\u555f\u52d5 \/ \u8dd1\u4e00\u9663\u5b50\u5f8c\uff09\u3002\u8acb\u6309\u6392\u67e5\u9806\u5e8f\u544a\u8a34\u6211\u5148\u8abf\u54ea\u500b vLLM \u53c3\u6578\u3001\u70ba\u4ec0\u9ebc\u3002\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\">\u5c0f\u7d50<\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n\n<li><strong>KV Cache<\/strong> = \u6a21\u578b\u7b97\u904e token \u7684\u7b46\u8a18\uff0c\u5e8f\u5217\u8d8a\u9577\u3001\u8acb\u6c42\u8d8a\u591a\u5403\u8d8a\u591a\u8a18\u61b6\u9ad4\uff0c\u5b83\u624d\u662f\u5403\u6389 GPU \u8a18\u61b6\u9ad4\u7684\u4e3b\u89d2\uff08\u4e0d\u662f\u6b0a\u91cd\uff09\u3002<\/li>\n\n\n<li><strong>PagedAttention<\/strong> = \u5b78\u4f5c\u696d\u7cfb\u7d71\u5206\u9801\uff0c\u628a KV Cache \u5207\u6210\u56fa\u5b9a\u5927\u5c0f\u7684 block \u7ba1\u7406\uff0c\u4e0d\u9810\u7559\u3001\u5c11\u788e\u7247\u3001\u53ef\u5171\u4eab \u2192 \u540c\u4e00\u5f35\u5361\u585e\u66f4\u591a\u8acb\u6c42\u3002<\/li>\n\n\n<li>\u555f\u52d5\u65e5\u8a8c\u770b\u5169\u500b\u6578\u5b57\uff1a<strong>KV cache \u7e3d\u91cf<\/strong> \u548c <strong>max concurrency<\/strong>\uff0c\u9019\u662f\u4f60\u7684\u5bb9\u91cf\u5100\u8868\u677f\u3002<\/li>\n\n\n<li>\u56db\u500b\u5e38\u7528\u53c3\u6578\uff1a<code>--gpu-memory-utilization<\/code>\uff08\u7897\u591a\u5927\uff09\u3001<code>--max-model-len<\/code>\uff08\u6bcf\u4efd\u591a\u9577\uff09\u3001<code>--max-num-seqs<\/code>\uff08\u540c\u6642\u5e7e\u4efd\uff09\u3001<code>--block-size<\/code>\uff08\u683c\u5b50\u5927\u5c0f\uff0c\u5148\u5225\u52d5\uff09\u3002<\/li>\n\n\n<li>OOM \u6392\u67e5\u9806\u5e8f\uff1a\u5148\u964d <code>max-model-len<\/code> \u2192 \u518d\u964d <code>gpu-memory-utilization<\/code> \u2192 \u518d\u964d <code>max-num-seqs<\/code> \u2192 \u6700\u5f8c\u63db\u5c0f\u6a21\u578b \/ \u91cf\u5316\u3002<\/li>\n\n\n<li>\u300c\u683c\u5b50\u53ef\u4ee5\u5171\u4eab\u300d\u7684 prefix caching\uff0c\u7559\u5f85\u7b2c 4 \u7bc7\u3002<\/li>\n\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u4e00\u7bc7\uff08#03\uff09\u6211\u5011\u6703\u628a\u670d\u52d9\u63a8\u5f97\u66f4\u9060\uff0c\u770b\u66f4\u591a\u5be6\u6230\u5834\u666f\u7684\u8abf\u6821\u3002<\/p>\n\n\n\n<div id=\"quiz-587eaf54\">\n<style>\n#quiz-587eaf54 .badge { display: inline-block; padding: 4px 10px; border-radius: 12px; font-size: 0.75em; font-weight: bold; margin-left: 8px; }#quiz-587eaf54 .badge.scenario { background: #ebf8ff; color: #2b6cb0; }#quiz-587eaf54 .badge.debug { background: #fef3c7; color: #92400e; }#quiz-587eaf54 .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-587eaf54 .question-adv h3 { margin-top: 0; color: #2d3748; }#quiz-587eaf54 .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-587eaf54 .options-adv { list-style: none; padding: 0; }#quiz-587eaf54 .options-adv li { padding: 14px 16px; margin: 10px 0; border: 2px solid #e2e8f0; border-radius: 8px; cursor: pointer; transition: all 0.2s; }#quiz-587eaf54 .options-adv li:hover:not(.disabled) { border-color: #ed8936; background: #fffaf0; }#quiz-587eaf54 .options-adv li.disabled { cursor: default; opacity: 0.7; }#quiz-587eaf54 .options-adv li.show-answer { border-color: #48bb78; background: #f0fff4; }#quiz-587eaf54 .options-adv li.wrong { border-color: #fc8181; background: #fff5f5; }#quiz-587eaf54 .feedback-adv { margin-top: 16px; padding: 14px; border-radius: 8px; display: none; line-height: 1.6; }#quiz-587eaf54 .feedback-adv.correct { background: #f0fff4; color: #276749; display: block; }#quiz-587eaf54 .feedback-adv.wrong { background: #fff5f5; color: #c53030; display: block; }#quiz-587eaf54 .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-587eaf54 .intro { background: #ebf8ff; padding: 16px; border-radius: 8px; margin-bottom: 20px; color: #2c5282; }\n<\/style>\n<h1>\u9032\u968e\u6e2c\u9a57\uff1aKV Cache \u8207 PagedAttention<\/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. \u60c5\u5883\u5224\u65b7 <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <div class=\"code-block\">\u4f60\u7684 GPU \u6709 24GB\uff0c\u8dd1\u4e00\u500b 14GB \u7684\u6a21\u578b\u3002\u555f\u52d5\u65e5\u8a8c\u986f\u793a\uff1a\nMaximum concurrency for 32,768 tokens per request: 1.5x\n\u4f60\u767c\u73fe\u4f75\u767c\u592a\u4f4e\uff0c\u670d\u52d9\u5e38\u5e38\u6392\u968a\u3002\u4f60\u300c\u77e5\u9053\u7528\u6236\u5be6\u969b\u7684 prompt \u5f88\u5c11\u8d85\u904e 8K token\u300d\u3002\n\u6700\u8a72\u512a\u5148\u505a\u7684\u8abf\u6574\u662f\uff1f<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(1, 'A', 'B')\">A. \u628a &#8211;gpu-memory-utilization \u5f9e 0.9 \u62c9\u5230 0.99<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'B', 'B')\">B. \u628a &#8211;max-model-len \u5f9e 32768 \u964d\u5230 8192<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'C', 'B')\">C. \u628a &#8211;block-size \u5f9e 16 \u6539\u6210 32<\/li>\n        <li onclick=\"checkAnswerAdv(1, 'D', 'B')\">D. \u7acb\u523b\u63db\u4e00\u500b\u66f4\u5c0f\u7684\u6a21\u578b<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv1\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv2\">\n      <h3>2. \u60c5\u5883\u5224\u65b7 <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <div class=\"code-block\">\u4f60\u7684\u670d\u52d9\u6709\u4e00\u6bb5\u56fa\u5b9a\u7684\u9577 system prompt\uff0c\u6bcf\u500b\u8acb\u6c42\u958b\u982d\u90fd\u4e00\u6a21\u4e00\u6a23\u3002\n\u540c\u4e8b\u8aaa\uff1a\u300c\u9019\u7a2e\u60c5\u6cc1 vLLM \u6709\u8fa6\u6cd5\u8b93\u9019\u4e9b\u76f8\u540c\u958b\u982d\u7684 KV Cache \u4e0d\u7528\u5404\u5b58\u4e00\u4efd\u3002\u300d\n\u9019\u6307\u7684\u662f\u6587\u7ae0\u9810\u544a\u7684\u54ea\u500b\u6a5f\u5236\uff08\u7b2c 4 \u7bc7\u4e3b\u984c\uff09\uff1f<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(2, 'A', 'C')\">A. \u63d0\u9ad8 gpu-memory-utilization<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'B', 'C')\">B. \u628a block-size \u8abf\u5927<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'C', 'C')\">C. prefix \u5171\u4eab \/ prefix caching\uff08\u76f8\u540c\u958b\u982d\u7684 block \u53ef\u5171\u7528\uff09<\/li>\n        <li onclick=\"checkAnswerAdv(2, 'D', 'C')\">D. \u964d\u4f4e max-num-seqs<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv2\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv3\">\n      <h3>3. \u60c5\u5883\u5224\u65b7 <span class=\"badge scenario\">\u60c5\u5883\u984c<\/span><\/h3>\n      <div class=\"code-block\">\u4f60\u8abf\u5b8c\u53c3\u6578\u91cd\u555f vLLM\uff0c\u60f3\u78ba\u8a8d\u9019\u6b21\u6539\u52d5\u300c\u5230\u5e95\u8b93\u80fd\u540c\u6642\u670d\u52d9\u7684\u8acb\u6c42\u6578\u8b8a\u591a\u9084\u8b8a\u5c11\u300d\u3002\n\u4f60\u61c9\u8a72\u53bb\u65e5\u8a8c\u88e1\u770b\u54ea\u4e9b\u884c\u4f86\u5224\u65b7\uff0c\u800c\u4e0d\u662f\u6191\u611f\u89ba\uff1f<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(3, 'A', 'A')\">A. GPU KV cache size\uff08\u6216 # GPU blocks\uff09\u8207 Maximum concurrency \u9019\u5169\u884c<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'B', 'A')\">B. \u53ea\u770b\u6a21\u578b\u6b0a\u91cd\u7684\u8f09\u5165\u6642\u9593<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'C', 'A')\">C. \u770b CUDA \u7248\u672c\u865f\u90a3\u4e00\u884c<\/li>\n        <li onclick=\"checkAnswerAdv(3, 'D', 'A')\">D. \u770b API \u76e3\u807d\u7684 port \u90a3\u4e00\u884c<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv3\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv4\">\n      <h3>4. \u932f\u8aa4\u8a3a\u65b7 <span class=\"badge debug\">\u932f\u8aa4\u8a3a\u65b7<\/span><\/h3>\n      <div class=\"code-block\"># AI \u5e6b\u4f60\u7522\u751f\u7684\u555f\u52d5\u6307\u4ee4\nvllm serve my-model \\\n  &#8211;gpu-memory-utilization 0.99 \\\n  &#8211;max-model-len 131072\n\n# \u73fe\u8c61\uff1a\u5e73\u5e38\u6c92\u4e8b\uff0c\u4e00\u9047\u5230\u6d41\u91cf\u5c16\u5cf0\u5c31\u6574\u500b\u670d\u52d9 OOM \u5d29\u6f70\n\u9019\u6bb5\u8a2d\u5b9a\u6700\u4e3b\u8981\u7684\u554f\u984c\u662f\u4ec0\u9ebc\uff1f<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(4, 'A', 'D')\">A. \u6c92\u8a2d &#8211;block-size \u5c0e\u81f4\u5206\u9801\u5931\u6548<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'B', 'D')\">B. \u6a21\u578b\u540d\u7a31\u5beb\u932f<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'C', 'D')\">C. &#8211;max-model-len \u8a2d\u592a\u5c0f\uff0c\u9577\u8acb\u6c42\u88ab\u62d2\u7d55<\/li>\n        <li onclick=\"checkAnswerAdv(4, 'D', 'D')\">D. \u8a18\u61b6\u9ad4\u7528\u5230 0.99 \u6c92\u7559\u9918\u88d5\uff0c\u52a0\u4e0a max-model-len \u62c9\u5230 128K \u58d3\u57ae\u7a7a\u9593\uff0c\u5c16\u5cf0\u6642 KV Cache \u6490\u7206\u800c OOM<\/li>\n      <\/ul>\n      <div class=\"feedback-adv\" id=\"fbAdv4\"><\/div>\n    <\/div>\n\n    <div class=\"question-adv\" id=\"qAdv5\">\n      <h3>5. \u932f\u8aa4\u8a3a\u65b7 <span class=\"badge debug\">\u932f\u8aa4\u8a3a\u65b7<\/span><\/h3>\n      <div class=\"code-block\">\u5169\u4f4d\u5de5\u7a0b\u5e2b\u90fd\u9047\u5230 CUDA out of memory\uff0c\u75c7\u72c0\u4e0d\u540c\uff1a\n\n\u7532\uff1avllm serve \u4e00\u555f\u52d5\u3001\u9084\u6c92\u6709\u4efb\u4f55\u8acb\u6c42\u9032\u4f86\u5c31\u76f4\u63a5 OOM\n\u4e59\uff1a\u670d\u52d9\u8dd1\u5f97\u597d\u597d\u7684\uff0c\u4f46\u540c\u6642\u6e67\u5165\u5927\u91cf\u9577\u8acb\u6c42\u6642\u624d OOM\n\n\u6309\u6587\u7ae0\u7684\u5224\u65b7\u6280\u5de7\uff0c\u4ed6\u5011\u8a72\u512a\u5148\u52d5\u7684\u53c3\u6578\u5206\u5225\u662f\uff1f<\/div>\n      <ul class=\"options-adv\">\n        <li onclick=\"checkAnswerAdv(5, 'A', 'C')\">A. \u7532\u4e59\u90fd\u5148\u964d max-num-seqs<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'B', 'C')\">B. \u7532\u964d max-num-seqs\uff1b\u4e59\u964d gpu-memory-utilization<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'C', 'C')\">C. \u7532\u52d5 gpu-memory-utilization \u6216\u63db\u5c0f\u6a21\u578b\uff1b\u4e59\u52d5 max-model-len \u6216 max-num-seqs<\/li>\n        <li onclick=\"checkAnswerAdv(5, 'D', 'C')\">D. \u7532\u4e59\u90fd\u53ea\u80fd\u63db\u66f4\u5c0f\u7684\u6a21\u578b\uff0c\u5225\u7121\u4ed6\u6cd5<\/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;\n      let correctAdv = 0;\n      const totalAdv = 5;\n      const explanationsAdv = {\n        1: \"\u6700\u6709\u6548\u4e14\u8a72\u512a\u5148\u52d5\u7684\u662f <code>--max-model-len<\/code>\uff1a\u65e2\u7136\u7528\u6236\u7528\u4e0d\u5230 32K\uff0c\u628a\u4e0a\u9650\u964d\u5230 8192\uff0c\u6700\u5927\u4f75\u767c\u76f4\u63a5\u7ffb\u597d\u5e7e\u500d\u3002A \u628a\u8a18\u61b6\u9ad4\u7528\u5230 0.99 \u53cd\u800c\u66f4\u6613 OOM\uff1bblock-size \u901a\u5e38\u4e0d\u7528\u52d5\uff1b\u9084\u6c92\u5230\u8981\u63db\u6a21\u578b\u7684\u5730\u6b65\u3002\",\n        2: \"\u7576\u591a\u500b\u8acb\u6c42\u5171\u7528\u540c\u4e00\u6bb5\u958b\u982d\uff08\u5982\u540c\u4e00\u4efd system prompt\uff09\uff0cPagedAttention \u7684 block \u53ef\u4ee5\u5171\u4eab\u3001\u53ea\u5b58\u4e00\u4efd\uff0c\u5c31\u662f prefix \u5171\u4eab \/ prefix caching\uff0c\u6587\u7ae0\u9810\u544a\u70ba\u7b2c 4 \u7bc7\u4e3b\u984c\u3002\",\n        3: \"<code>GPU KV cache size<\/code>\uff08\u80fd\u5b58\u591a\u5c11 token \u7684\u7b46\u8a18\uff09\u548c <code>Maximum concurrency<\/code>\uff08\u5927\u7d04\u80fd\u540c\u6642\u670d\u52d9\u5e7e\u500b\u8acb\u6c42\uff09\u5c31\u662f\u5bb9\u91cf\u5100\u8868\u677f\uff0c\u8abf\u5b8c\u53c3\u6578\u8981\u56de\u982d\u6838\u5c0d\u9019\u5169\u500b\u6578\u5b57\u5f80\u54ea\u500b\u65b9\u5411\u8b8a\u3002\",\n        4: \"\u5169\u500b\u8caa\u5fc3\u8a2d\u5b9a\u758a\u52a0\uff1autilization 0.99 \u4e0d\u7559\u9918\u88d5\u3001max-model-len \u53c8\u62c9\u5230 128K \u4f54\u7528\u5927\u91cf\u898f\u5283\u7a7a\u9593\u3002\u5e73\u5e38\u6490\u5f97\u4f4f\uff0c\u5c16\u5cf0\u6642 KV Cache \u81a8\u8139\u5c31\u628a\u8a18\u61b6\u9ad4\u6490\u7206 OOM\u3002\u61c9\u964d\u56de 0.9 \u4e26\u628a max-model-len \u8a2d\u6210\u8cbc\u8fd1\u5be6\u969b\u9700\u6c42\u3002\",\n        5: \"\u5224\u65b7\u6280\u5de7\u770b OOM \u767c\u751f\u6642\u6a5f\uff1a\u4e00\u555f\u52d5\u5c31 OOM \u4ee3\u8868\u6b0a\u91cd+\u4fdd\u7559\u7a7a\u9593\u5c31\u8d85\u4e86\uff0c\u52d5 gpu-memory-utilization \u6216\u63db\u5c0f\u6a21\u578b\uff1b\u6d41\u91cf\u4e0a\u4f86\u624d OOM \u662f KV Cache \u81a8\u8139\u6490\u7206\uff0c\u52d5 max-model-len \u6216 max-num-seqs\u3002\"\n      };\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\uff1aKV Cache \u8207 PagedAttention \u5171 5 \u984c\uff0c\u9ede\u9078\u7b54\u6848\u5f8c\u6703\u7acb\u5373\u986f\u793a\u7d50\u679c 1. \u6839\u64da\u6587\u7ae0 [&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":[865,863,866,864,847],"class_list":["post-838","post","type-post","status-publish","format-standard","hentry","category-vllm-","tag-gpu","tag-kv-cache","tag-oom","tag-pagedattention","tag-vllm"],"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\/838","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=838"}],"version-history":[{"count":0,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/posts\/838\/revisions"}],"wp:attachment":[{"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/media?parent=838"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/categories?post=838"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.itsmygo.uk\/wordpress\/wp-json\/wp\/v2\/tags?post=838"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}