{"id":112,"date":"2026-04-23T17:23:49","date_gmt":"2026-04-23T09:23:49","guid":{"rendered":"https:\/\/www.bojinhu.xyz\/?p=112"},"modified":"2026-04-23T17:23:49","modified_gmt":"2026-04-23T09:23:49","slug":"vllm%e9%83%a8%e7%bd%b2%e6%9c%ac%e5%9c%b0%e6%a8%a1%e5%9e%8b%e7%9a%84%e4%b8%80%e4%ba%9b%e5%b0%8f%e5%ae%9e%e9%aa%8c","status":"publish","type":"post","link":"https:\/\/www.bojinhu.xyz\/index.php\/2026\/04\/23\/vllm%e9%83%a8%e7%bd%b2%e6%9c%ac%e5%9c%b0%e6%a8%a1%e5%9e%8b%e7%9a%84%e4%b8%80%e4%ba%9b%e5%b0%8f%e5%ae%9e%e9%aa%8c\/","title":{"rendered":"vllm\u90e8\u7f72\u672c\u5730\u6a21\u578b\u7684\u4e00\u4e9b\u5c0f\u5b9e\u9a8c"},"content":{"rendered":"<h1 id=\"\u6211\u5728\u672c\u5730\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668\u4e0a\u505a\u4e86\u4e00\u6b21-vllm-\u90e8\u7f72\u6d4b\u8bd5\u54ea\u4e9b\u53c2\u6570\u771f\u7684\u6709\u7528\">\u6211\u5728\u672c\u5730\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668\u4e0a\u505a\u4e86\u4e00\u6b21vLLM \u90e8\u7f72\u6d4b\u8bd5\uff1a\u54ea\u4e9b\u53c2\u6570\u771f\u7684\u6709\u7528\uff1f<\/h1>\n<p>\u6700\u8fd1\u6211\u5728\u672c\u5730\u4e00\u53f0<strong>\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668<\/strong>\u4e0a\u505a\u4e86\u4e00\u8f6e vLLM<br \/>\n\u90e8\u7f72\u6d4b\u8bd5\u3002\u76ee\u7684\u5e76\u4e0d\u662f\u8dd1\u4e00\u4e2a\u201c\u6700\u597d\u770b\u201d\u7684<br \/>\nbenchmark\uff0c\u800c\u662f\u60f3\u56de\u7b54\u4e00\u4e2a\u66f4\u5b9e\u9645\u7684\u95ee\u9898\uff1a<\/p>\n<blockquote><p>\u5728\u81ea\u5df1\u7684\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u5927\u6a21\u578b\u65f6\uff0cvLLM<br \/>\n\u7684\u4e0d\u540c\u53c2\u6570\u914d\u7f6e\uff0c\u5230\u5e95\u4f1a\u5bf9\u771f\u5b9e\u63a8\u7406\u6027\u80fd\u5e26\u6765\u591a\u5927\u5f71\u54cd\uff1f<\/p><\/blockquote>\n<p>\u8fd9\u7c7b\u95ee\u9898\u6211\u4e4b\u524d\u5176\u5b9e\u4e00\u76f4\u662f\u201c\u77e5\u9053\u6709\u8fd9\u4e9b\u53c2\u6570\u201d\uff0c\u4f46\u5e76\u6ca1\u6709\u771f\u6b63\u7cfb\u7edf\u5730\u6d4b\u8fc7\u3002\u6bd4\u5982\uff1a<code>optimization_level<\/code><br \/>\n\u8981\u5f00\u5230\u51e0\u7ea7\uff1f<code>performance_mode<\/code> \u9009 throughput \u8fd8\u662f<br \/>\ninteractivity\uff1f<code>YaRN<\/code>\u3001<code>MTP<\/code>\u3001<code>expert parallel<\/code><br \/>\n\u8fd9\u4e9b\u5f00\u5173\u5230\u5e95\u503c\u4e0d\u503c\u5f97\u52a8\uff1f<\/p>\n<p>\u6240\u4ee5\u8fd9\u6b21\u6211\u5e72\u8106\u628a\u5b83\u4eec\u62c6\u5f00\u505a\u4e86\u4e00\u8f6e\u6d4b\u8bd5\uff0c\u6574\u7406\u6210\u8fd9\u7bc7\u5b9e\u9a8c\u8bb0\u5f55\u3002\u6574\u7bc7\u6587\u7ae0\u66f4\u504f\u5411<strong>\u4e2a\u4eba\u5b9e\u8df5\u603b\u7ed3<\/strong>\uff0c\u9002\u5408\u548c\u6211\u4e00\u6837\u60f3\u5728\u672c\u5730\u670d\u52a1\u5668\u4e0a\u4f53\u9a8c<br \/>\nvLLM \u90e8\u7f72\u3001\u8c03\u53c2\u548c\u6027\u80fd\u5206\u6790\u7684\u4eba\u53c2\u8003\u3002<\/p>\n<hr \/>\n<h2 id=\"\u4e00\u6d4b\u8bd5\u80cc\u666f\">\u4e00\u3001\u6d4b\u8bd5\u80cc\u666f<\/h2>\n<p>\u8fd9\u6b21\u6d4b\u8bd5\u7684\u573a\u666f\u975e\u5e38\u660e\u786e\uff1a<\/p>\n<ul>\n<li><strong>\u90e8\u7f72\u65b9\u5f0f<\/strong>\uff1a\u672c\u5730\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668<\/li>\n<li><strong>\u63a8\u7406\u6846\u67b6<\/strong>\uff1avLLM<\/li>\n<li><strong>\u6d4b\u8bd5\u76ee\u6807<\/strong>\uff1a\u6bd4\u8f83\u4e0d\u540c\u90e8\u7f72\u53c2\u6570\u7ec4\u5408\u5bf9\u541e\u5410\u3001\u9996 token<br \/>\n\u5ef6\u8fdf\u3001\u5355 token \u751f\u6210\u65f6\u95f4\u548c\u7aef\u5230\u7aef\u65f6\u5ef6\u7684\u5f71\u54cd<\/li>\n<li><strong>\u5173\u6ce8\u91cd\u70b9<\/strong>\uff1a\n<ul>\n<li>\u4f4e\u5e76\u53d1\u573a\u666f\u4e0b\uff0c\u600e\u6837\u7684\u914d\u7f6e\u66f4\u9002\u5408\u201c\u4ea4\u4e92\u5f0f\u4f7f\u7528\u201d<\/li>\n<li>\u9ad8\u5e76\u53d1\u573a\u666f\u4e0b\uff0c\u600e\u6837\u7684\u914d\u7f6e\u66f4\u9002\u5408\u201c\u538b\u69a8\u541e\u5410\u201d<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c\u8fd9\u6b21\u6d4b\u8bd5\u4e0d\u662f\u60f3\u8bc1\u660e\u67d0\u4e2a\u53c2\u6570\u201c\u7406\u8bba\u4e0a\u5148\u8fdb\u201d\uff0c\u800c\u662f\u60f3\u56de\u7b54\u4e00\u4e2a\u5de5\u7a0b\u95ee\u9898\uff1a<\/p>\n<blockquote><p>\u5982\u679c\u6211\u624b\u91cc\u53ea\u6709\u4e00\u53f0\u672c\u5730\u53cc\u5361\u673a\u5668\uff0c\u60f3\u628a vLLM<br \/>\n\u90e8\u7f72\u5f97\u66f4\u5b9e\u7528\uff0c\u6211\u5e94\u8be5\u4f18\u5148\u5173\u6ce8\u54ea\u4e9b\u53c2\u6570\uff1f<\/p><\/blockquote>\n<hr \/>\n<h2 id=\"\u4e8c\u6d4b\u8bd5\u73af\u5883\u4e0e\u65b9\u6cd5\">\u4e8c\u3001\u6d4b\u8bd5\u73af\u5883\u4e0e\u65b9\u6cd5<\/h2>\n<h3 id=\"\u6d4b\u8bd5\u73af\u5883\">1. \u6d4b\u8bd5\u73af\u5883<\/h3>\n<p>\u4ece\u5b9e\u9a8c\u8bb0\u5f55\u6765\u770b\uff0c\u672c\u6b21\u6d4b\u8bd5\u4f7f\u7528\u7684\u662f\u4e00\u5957\u56fa\u5b9a\u7684\u90e8\u7f72\u73af\u5883\uff1a<\/p>\n<ul>\n<li><strong>\u5355\u8282\u70b9\u53cc\u5361<\/strong><\/li>\n<li><code>tensor_parallel_size = 2<\/code><\/li>\n<li><code>max_model_len = 81920<\/code><\/li>\n<li>\u8f93\u5165\u957f\u5ea6\uff1a<code>512<\/code><\/li>\n<li>\u8f93\u51fa\u957f\u5ea6\uff1a<code>256<\/code><\/li>\n<li>\u6e29\u5ea6\uff1a<code>0<\/code><\/li>\n<li>\u6bcf\u7ec4\u6d4b\u8bd5\u8bf7\u6c42\u6570\uff1a<code>100<\/code><\/li>\n<\/ul>\n<p>\u6d4b\u8bd5\u4e2d\u4f7f\u7528\u4e86\u4e24\u7c7b\u5e76\u53d1\u573a\u666f\uff1a<\/p>\n<ul>\n<li><strong>\u4f4e\u5e76\u53d1<\/strong>\uff1a<code>max_concurrency = 2<\/code><\/li>\n<li><strong>\u9ad8\u5e76\u53d1<\/strong>\uff1a<code>max_concurrency = 100<\/code><\/li>\n<\/ul>\n<p>\u6211\u628a\u5b83\u4eec\u7406\u89e3\u6210\u4e24\u79cd\u5178\u578b\u4f7f\u7528\u65b9\u5f0f\uff1a<\/p>\n<ul>\n<li><strong>\u4f4e\u5e76\u53d1<\/strong>\u66f4\u50cf\u672c\u5730\u4ea4\u4e92\u3001\u5c11\u91cf\u7528\u6237\u6216\u81ea\u5df1\u8c03\u8bd5\u65f6\u7684\u573a\u666f\u3002<\/li>\n<li><strong>\u9ad8\u5e76\u53d1<\/strong>\u66f4\u50cf\u538b\u6d4b\u3001\u6279\u91cf\u8bf7\u6c42\u6216\u8005\u591a\u4eba\u5171\u7528\u670d\u52a1\u65f6\u7684\u573a\u666f\u3002<\/li>\n<\/ul>\n<h3 id=\"\u5bf9\u6bd4\u7684\u914d\u7f6e\u9879\">2. \u5bf9\u6bd4\u7684\u914d\u7f6e\u9879<\/h3>\n<p>\u8fd9\u6b21\u4e00\u5171\u6d4b\u8bd5\u4e86 <strong>10 \u7ec4\u914d\u7f6e<\/strong>\uff0c\u6bcf\u7ec4\u5728<strong>\u4f4e\u5e76\u53d1 \/<br \/>\n\u9ad8\u5e76\u53d1<\/strong>\u4e24\u4e2a\u573a\u666f\u4e0b\u5404\u8dd1 <strong>2 \u6b21\u91cd\u590d\u5b9e\u9a8c<\/strong>\uff0c\u603b\u5171<br \/>\n<strong>40 \u6761\u7ed3\u679c<\/strong>\u3002<\/p>\n<p>\u4e3b\u8981\u5bf9\u6bd4\u7684\u53d8\u91cf\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><code>optimization_level = 0 \/ 1 \/ 2 \/ 3<\/code><\/li>\n<li><code>performance_mode = throughput \/ balanced \/ interactivity<\/code><\/li>\n<li>\u662f\u5426\u5f00\u542f <code>expert parallel<\/code><\/li>\n<li>\u662f\u5426\u5f00\u542f <code>YaRN<\/code><\/li>\n<li>\u662f\u5426\u5f00\u542f <code>MTP<\/code><\/li>\n<li>\u662f\u5426\u8bbe\u7f6e <code>cp_kv_cache_interleave_size = 16<\/code><\/li>\n<\/ul>\n<h3 id=\"\u5173\u6ce8\u7684\u6307\u6807\">3. \u5173\u6ce8\u7684\u6307\u6807<\/h3>\n<p>\u672c\u6b21\u4e3b\u8981\u770b\u4ee5\u4e0b\u51e0\u4e2a\u6307\u6807\uff1a<\/p>\n<ul>\n<li><strong>request throughput<\/strong>\uff1a\u8bf7\u6c42\u541e\u5410<\/li>\n<li><strong>output throughput<\/strong>\uff1a\u8f93\u51fa token \u541e\u5410<\/li>\n<li><strong>TTFT<\/strong>\uff08Time To First Token\uff09\uff1a\u9996 token \u5ef6\u8fdf<\/li>\n<li><strong>TPOT<\/strong>\uff08Time Per Output Token\uff09\uff1a\u5e73\u5747\u5355 token<br \/>\n\u751f\u6210\u8017\u65f6<\/li>\n<li><strong>E2EL<\/strong>\uff08End-to-End Latency\uff09\uff1a\u7aef\u5230\u7aef\u65f6\u5ef6<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u6307\u6807\u7ec4\u5408\u5728\u4e00\u8d77\uff0c\u57fa\u672c\u80fd\u53cd\u6620\u4e00\u5957\u63a8\u7406\u914d\u7f6e\u5230\u5e95\u662f\u504f\u201c\u4ea4\u4e92\u4f53\u9a8c\u201d\u8fd8\u662f\u504f\u201c\u541e\u5410\u6548\u7387\u201d\u3002<\/p>\n<hr \/>\n<h2 id=\"\u4e09\u6574\u4f53\u7ed3\u8bba\u5148\u8bf4\">\u4e09\u3001\u6574\u4f53\u7ed3\u8bba\u5148\u8bf4<\/h2>\n<p>\u5982\u679c\u53ea\u8ba9\u6211\u5148\u8bf4\u7ed3\u8bba\uff0c\u6211\u4f1a\u603b\u7ed3\u6210\u4e0b\u9762\u51e0\u53e5\u8bdd\uff1a<\/p>\n<ol type=\"1\">\n<li><strong><code>optimization_level<\/code> \u4ece 0 \u5347\u5230<br \/>\n2\/3\uff0c\u6027\u80fd\u63d0\u5347\u975e\u5e38\u660e\u663e\uff0cOpt0 \u57fa\u672c\u4e0d\u5efa\u8bae\u4f7f\u7528\u3002<\/strong><\/li>\n<li><strong>\u4f4e\u5e76\u53d1\u573a\u666f\u4e0b\uff0c\u4fdd\u7559 MTP \u66f4\u597d\uff1b\u5173\u95ed MTP<br \/>\n\u4f1a\u660e\u663e\u62c9\u4f4e\u541e\u5410\u5e76\u62c9\u957f\u7aef\u5230\u7aef\u65f6\u5ef6\u3002<\/strong><\/li>\n<li><strong>\u9ad8\u5e76\u53d1\u573a\u666f\u4e0b\uff0c\u5173\u95ed MTP<br \/>\n\u53cd\u800c\u662f\u8fd9\u6279\u6d4b\u8bd5\u91cc\u8868\u73b0\u6700\u597d\u7684\u914d\u7f6e\u3002<\/strong><\/li>\n<li><strong><code>YaRN<\/code>\u3001<code>performance_mode<\/code>\u3001<code>expert parallel<\/code><br \/>\n\u7684\u5f71\u54cd\u90fd\u5b58\u5728\uff0c\u4f46\u8fdc\u5c0f\u4e8e <code>optimization_level<\/code> \u548c<br \/>\n<code>MTP<\/code> \u5f00\u5173\u3002<\/strong><\/li>\n<li><strong>Opt2 \u548c Opt3<br \/>\n\u5df2\u7ecf\u5f88\u63a5\u8fd1\uff0c\u5f88\u591a\u60c5\u51b5\u4e0b\u771f\u6b63\u51b3\u5b9a\u7ed3\u679c\u7684\u4e0d\u662f\u201c\u518d\u5fae\u8c03\u4e00\u70b9<br \/>\nmode\u201d\uff0c\u800c\u662f\u6709\u6ca1\u6709\u8e29\u5230\u5173\u952e\u53c2\u6570\u3002<\/strong><\/li>\n<\/ol>\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u8fd9\u6b21\u5b9e\u9a8c\u6700\u6709\u4ef7\u503c\u7684\u5730\u65b9\u5e76\u4e0d\u662f\u53d1\u73b0\u4e86\u67d0\u4e2a\u201c\u795e\u5947\u53c2\u6570\u201d\uff0c\u800c\u662f\u8ba9\u6211\u66f4\u6e05\u695a\u5730\u770b\u5230\uff1a<\/p>\n<blockquote><p>\u5728 vLLM<br \/>\n\u672c\u5730\u90e8\u7f72\u91cc\uff0c\u771f\u6b63\u4e00\u9524\u5b9a\u97f3\u7684\u914d\u7f6e\u9879\u5176\u5b9e\u6ca1\u6709\u90a3\u4e48\u591a\uff0c\u6700\u5173\u952e\u7684\u5c31\u90a3\u51e0\u4e2a\u3002<\/p><\/blockquote>\n<hr \/>\n<h2 id=\"\u56db\u4f4e\u5e76\u53d1\u573a\u666f\u66f4\u50cf\u771f\u5b9e\u4ea4\u4e92\u4f7f\u7528\">\u56db\u3001\u4f4e\u5e76\u53d1\u573a\u666f\uff1a\u66f4\u50cf\u771f\u5b9e\u4ea4\u4e92\u4f7f\u7528<\/h2>\n<p>\u5148\u770b\u4f4e\u5e76\u53d1\uff0c\u4e5f\u5c31\u662f <code>max_concurrency = 2<\/code> \u7684\u60c5\u51b5\u3002<\/p>\n<p>\u5728\u8fd9\u4e2a\u573a\u666f\u91cc\uff0c<strong>Opt3<\/strong> \u662f\u4e00\u4e2a\u5f88\u5f3a\u7684\u57fa\u51c6\u914d\u7f6e\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\uff1a<strong>1.035 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\uff1a<strong>264.879 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 TTFT\uff1a<strong>169.859 ms<\/strong><\/li>\n<li>\u5e73\u5747 E2EL\uff1a<strong>1930.785 ms<\/strong><\/li>\n<\/ul>\n<h3 id=\"opt0-\u548c-opt1-\u7684\u5dee\u8ddd\u975e\u5e38\u660e\u663e\">1. Opt0 \u548c Opt1 \u7684\u5dee\u8ddd\u975e\u5e38\u660e\u663e<\/h3>\n<p>\u6700\u5dee\u7684\u914d\u7f6e\u662f <strong>Opt0<\/strong>\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\u53ea\u6709 <strong>0.169 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\u53ea\u6709 <strong>43.373 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 E2EL \u9ad8\u8fbe <strong>11796.860 ms<\/strong><\/li>\n<\/ul>\n<p>\u548c Opt3 \u76f8\u6bd4\uff0cOpt0\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\u4e0b\u964d\u7ea6 <strong>83.7%<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\u4e0b\u964d\u7ea6 <strong>83.6%<\/strong><\/li>\n<li>\u7aef\u5230\u7aef\u65f6\u5ef6\u589e\u52a0\u7ea6 <strong>511%<\/strong><\/li>\n<\/ul>\n<p>\u8fd9\u4e2a\u7ed3\u679c\u5176\u5b9e\u5f88\u76f4\u89c2\uff1a<strong>optimization_level<br \/>\n\u4e0d\u5f00\u8d77\u6765\uff0c\u5f88\u591a\u4f18\u5316\u6839\u672c\u6ca1\u5403\u5230<\/strong>\uff0c\u6700\u7ec8\u4f1a\u8ba9\u6574\u5957\u670d\u52a1\u4ece\u201c\u53ef\u7528\u201d\u53d8\u6210\u201c\u975e\u5e38\u7b28\u91cd\u201d\u3002<\/p>\n<p>Opt1 \u867d\u7136\u6bd4 Opt0 \u597d\u5f88\u591a\uff0c\u4f46\u548c Opt3 \u4ecd\u7136\u6709\u660e\u663e\u5dee\u8ddd\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410 <strong>0.733 req\/s<\/strong><\/li>\n<li>\u5e73\u5747 E2EL <strong>2724.515 ms<\/strong><\/li>\n<\/ul>\n<p>\u76f8\u6bd4 Opt3\uff0c\u541e\u5410\u4ecd\u4f4e\u7ea6 <strong>29%<\/strong>\uff0c\u7aef\u5230\u7aef\u65f6\u5ef6\u9ad8\u7ea6<br \/>\n<strong>41%<\/strong>\u3002\u6240\u4ee5\u5982\u679c\u673a\u5668\u548c\u6a21\u578b\u5141\u8bb8\uff0c<strong>Opt1<br \/>\n\u4e5f\u4e0d\u7b97\u4e00\u4e2a\u7406\u60f3\u9009\u62e9<\/strong>\u3002<\/p>\n<h3 id=\"opt2-\u548c-opt3-\u5df2\u7ecf\u975e\u5e38\u63a5\u8fd1\">2. Opt2 \u548c Opt3 \u5df2\u7ecf\u975e\u5e38\u63a5\u8fd1<\/h3>\n<p>Opt2 \u7684\u7ed3\u679c\u662f\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410 <strong>1.026 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410 <strong>262.748 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 E2EL <strong>1945.348 ms<\/strong><\/li>\n<\/ul>\n<p>\u548c Opt3 \u76f8\u6bd4\uff0c\u5dee\u8ddd\u90fd\u5728 <strong>1%<br \/>\n\u5de6\u53f3<\/strong>\u3002\u8fd9\u8bf4\u660e\u5728\u4f4e\u5e76\u53d1\u573a\u666f\u4e0b\uff0c<strong>Opt2<br \/>\n\u57fa\u672c\u5df2\u7ecf\u5403\u5230\u4e86\u5927\u90e8\u5206\u6027\u80fd\u6536\u76ca<\/strong>\uff0cOpt3<br \/>\n\u66f4\u591a\u662f\u8fdb\u4e00\u6b65\u6253\u78e8\uff0c\u800c\u4e0d\u662f\u672c\u8d28\u6027\u8dc3\u5347\u3002<\/p>\n<h3 id=\"\u5173\u95ed-mtp-\u4f1a\u660e\u663e\u53d8\u5dee\">3. \u5173\u95ed MTP \u4f1a\u660e\u663e\u53d8\u5dee<\/h3>\n<p>\u4f4e\u5e76\u53d1\u4e0b\uff0c\u6211\u6700\u5173\u6ce8\u7684\u4e00\u4e2a\u73b0\u8c61\u662f\uff1a<strong>\u5173\u95ed MTP<br \/>\n\u4f1a\u660e\u663e\u4f24\u5bb3\u6027\u80fd\u3002<\/strong><\/p>\n<p>\u5173\u95ed MTP \u540e\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\uff1a<strong>0.798 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\uff1a<strong>204.313 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 E2EL\uff1a<strong>2505.243 ms<\/strong><\/li>\n<\/ul>\n<p>\u76f8\u5bf9 Opt3\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\u4e0b\u964d\u7ea6 <strong>22.9%<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\u4e0b\u964d\u7ea6 <strong>22.9%<\/strong><\/li>\n<li>\u5e73\u5747 E2EL \u589e\u52a0\u7ea6 <strong>29.8%<\/strong><\/li>\n<\/ul>\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u5728\u4f4e\u5e76\u53d1\u4ea4\u4e92\u573a\u666f\u91cc\uff0c<strong>MTP<br \/>\n\u662f\u6709\u660e\u663e\u6b63\u6536\u76ca\u7684<\/strong>\u3002\u5982\u679c\u4f60\u7684\u4e3b\u8981\u4f7f\u7528\u65b9\u5f0f\u662f\u81ea\u5df1\u5728\u672c\u5730\u4ea4\u4e92\u3001\u5c11\u91cf\u7528\u6237\u8c03\u7528\uff0c\u6216\u8005\u66f4\u91cd\u89c6\u54cd\u5e94\u4f53\u9a8c\uff0c\u90a3\u4e48\u6211\u4f1a\u503e\u5411\u4e8e<strong>\u4fdd\u7559<br \/>\nMTP<\/strong>\u3002<\/p>\n<h3 id=\"\u5176\u4ed6\u53c2\u6570\u5f71\u54cd\u76f8\u5bf9\u6709\u9650\">4. \u5176\u4ed6\u53c2\u6570\u5f71\u54cd\u76f8\u5bf9\u6709\u9650<\/h3>\n<p>\u5269\u4e0b\u51e0\u4e2a\u53d8\u91cf\u2014\u2014<code>YaRN<\/code>\u3001<code>Balanced \u6a21\u5f0f<\/code>\u3001<code>Interactivity \u6a21\u5f0f<\/code>\u3001\u5173\u95ed<br \/>\n<code>expert parallel<\/code>\u2014\u2014\u5728\u4f4e\u5e76\u53d1\u91cc\u53d8\u5316\u90fd\u4e0d\u5927\u3002<\/p>\n<p>\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li><strong>\u5173\u95ed YaRN<\/strong>\uff1a\u51e0\u4e4e\u548c Opt3 \u6301\u5e73<\/li>\n<li><strong>Balanced \u6a21\u5f0f<\/strong>\uff1a\u548c Opt3 \u57fa\u672c\u540c\u7ea7<\/li>\n<li><strong>Interactivity \u6a21\u5f0f<\/strong>\uff1a\u6ca1\u6709\u51fa\u73b0\u7279\u522b\u7a81\u51fa\u7684\u4f18\u52bf<\/li>\n<li><strong>\u5173\u95ed EP<\/strong>\uff1a\u751a\u81f3\u7565\u597d\u4e00\u70b9\uff0c\u4f46\u5e45\u5ea6\u5f88\u5c0f<\/li>\n<\/ul>\n<p>\u8fd9\u8bf4\u660e\u5728\u4f4e\u5e76\u53d1\u573a\u666f\u4e0b\uff0c\u771f\u6b63\u5f71\u54cd\u4f53\u9a8c\u7684\u4e0d\u662f\u8fd9\u4e9b\u201c\u7ec6\u5206 mode\u201d\uff0c\u800c\u662f\uff1a<\/p>\n<ol type=\"1\">\n<li>\u4f60\u7684\u4f18\u5316\u7b49\u7ea7\u662f\u4e0d\u662f\u592a\u4f4e\uff1b<\/li>\n<li>\u4f60\u6709\u6ca1\u6709\u628a MTP \u5173\u6389\u3002<\/li>\n<\/ol>\n<hr \/>\n<h2 id=\"\u4e94\u9ad8\u5e76\u53d1\u573a\u666f\u66f4\u63a5\u8fd1\u538b\u6d4b\u4e0e\u5171\u4eab\u670d\u52a1\">\u4e94\u3001\u9ad8\u5e76\u53d1\u573a\u666f\uff1a\u66f4\u63a5\u8fd1\u538b\u6d4b\u4e0e\u5171\u4eab\u670d\u52a1<\/h2>\n<p>\u518d\u770b\u9ad8\u5e76\u53d1\uff0c\u4e5f\u5c31\u662f <code>max_concurrency = 100<\/code> \u7684\u60c5\u51b5\u3002<\/p>\n<p>Opt3 \u5728\u9ad8\u5e76\u53d1\u4e0b\u7684\u57fa\u51c6\u7ed3\u679c\u662f\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\uff1a<strong>6.497 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\uff1a<strong>1663.330 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 TTFT\uff1a<strong>4670.809 ms<\/strong><\/li>\n<li>\u5e73\u5747 E2EL\uff1a<strong>14881.842 ms<\/strong><\/li>\n<\/ul>\n<p>\u8fd9\u4e2a\u573a\u666f\u4e0b\uff0c\u6700\u6709\u610f\u601d\u7684\u5730\u65b9\u6765\u4e86\u3002<\/p>\n<h3 id=\"\u5173\u95ed-mtp-\u53cd\u800c\u6210\u4e86\u6700\u4f18\u914d\u7f6e\">1. \u5173\u95ed MTP \u53cd\u800c\u6210\u4e86\u6700\u4f18\u914d\u7f6e<\/h3>\n<p>\u5728\u9ad8\u5e76\u53d1\u6d4b\u8bd5\u91cc\uff0c<strong>\u5173\u95ed MTP<\/strong> \u7684\u7ed3\u679c\u6700\u597d\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\uff1a<strong>6.974 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\uff1a<strong>1785.471 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 TTFT\uff1a<strong>4078.795 ms<\/strong><\/li>\n<li>\u5e73\u5747 E2EL\uff1a<strong>14302.512 ms<\/strong><\/li>\n<\/ul>\n<p>\u76f8\u5bf9 Opt3\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\u63d0\u5347\u7ea6 <strong>7.3%<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\u63d0\u5347\u7ea6 <strong>7.3%<\/strong><\/li>\n<li>\u5e73\u5747 TTFT \u964d\u4f4e\u7ea6 <strong>12.7%<\/strong><\/li>\n<li>\u5e73\u5747 E2EL \u964d\u4f4e\u7ea6 <strong>3.9%<\/strong><\/li>\n<\/ul>\n<p>\u8fd9\u4e2a\u7ed3\u8bba\u548c\u4f4e\u5e76\u53d1\u6b63\u597d\u76f8\u53cd\u3002<\/p>\n<p>\u6211\u7684\u7406\u89e3\u662f\uff1a\u5728\u9ad8\u5e76\u53d1\u538b\u529b\u4e0b\uff0c\u7cfb\u7edf\u5df2\u7ecf\u4e0d\u518d\u662f\u201c\u5355\u4e2a\u8bf7\u6c42\u662f\u5426\u66f4\u5feb\u201d\u8fd9\u4e48\u7b80\u5355\uff0c\u800c\u662f\u8981\u8003\u8651\u6574\u4f53\u8c03\u5ea6\u3001\u8d44\u6e90\u7ade\u4e89\u548c\u541e\u5410\u6548\u7387\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c<strong>MTP<br \/>\n\u5e26\u6765\u7684\u989d\u5916\u673a\u5236\u672a\u5fc5\u603b\u662f\u51c0\u6536\u76ca<\/strong>\uff0c\u751a\u81f3\u53ef\u80fd\u62d6\u7d2f\u6574\u4f53\u5e76\u53d1\u6548\u7387\u3002<\/p>\n<p>\u6240\u4ee5\u5982\u679c\u4f60\u7684\u4f7f\u7528\u76ee\u6807\u66f4\u63a5\u8fd1\uff1a<\/p>\n<ul>\n<li>\u591a\u8bf7\u6c42\u6279\u91cf\u8dd1<\/li>\n<li>\u591a\u4eba\u5171\u4eab\u63a8\u7406\u670d\u52a1<\/li>\n<li>\u60f3\u5c3d\u53ef\u80fd\u69a8\u51fa\u541e\u5410<\/li>\n<\/ul>\n<p>\u90a3\u4e48\u8fd9\u7ec4\u6570\u636e\u7ed9\u51fa\u7684\u5efa\u8bae\u5f88\u660e\u786e\uff1a<strong>\u53ef\u4ee5\u8ba4\u771f\u8003\u8651\u5173\u95ed MTP<br \/>\n\u6d4b\u4e00\u8f6e\u3002<\/strong><\/p>\n<h3 id=\"opt0-\u4f9d\u7136\u660e\u663e\u843d\u540e\">2. Opt0 \u4f9d\u7136\u660e\u663e\u843d\u540e<\/h3>\n<p>\u9ad8\u5e76\u53d1\u4e0b Opt0 \u4e5f\u8fd8\u662f\u6700\u5dee\u7684\u4e00\u6863\uff1a<\/p>\n<ul>\n<li>\u8bf7\u6c42\u541e\u5410\uff1a<strong>5.379 req\/s<\/strong><\/li>\n<li>\u8f93\u51fa\u541e\u5410\uff1a<strong>1377.060 tok\/s<\/strong><\/li>\n<li>\u5e73\u5747 E2EL\uff1a<strong>17076.570 ms<\/strong><\/li>\n<\/ul>\n<p>\u867d\u7136\u9ad8\u5e76\u53d1\u4e0b\u5b83\u6ca1\u6709\u50cf\u4f4e\u5e76\u53d1\u90a3\u6837\u201c\u5d29\u5f97\u7279\u522b\u5938\u5f20\u201d\uff0c\u4f46\u76f8\u5bf9 Opt3 \u4f9d\u65e7\uff1a<\/p>\n<ul>\n<li>\u541e\u5410\u4e0b\u964d\u7ea6 <strong>17.2%<\/strong><\/li>\n<li>\u7aef\u5230\u7aef\u65f6\u5ef6\u589e\u52a0\u7ea6 <strong>14.7%<\/strong><\/li>\n<\/ul>\n<p>\u6240\u4ee5\u6211\u7684\u7ed3\u8bba\u8fd8\u662f\u4e00\u6837\uff1a<strong>Opt0 \u4e0d\u63a8\u8350\u3002<\/strong><\/p>\n<h3 id=\"opt1-opt2-opt3-\u5dee\u8ddd\u5e76\u4e0d\u5927\">3. Opt1 \/ Opt2 \/ Opt3 \u5dee\u8ddd\u5e76\u4e0d\u5927<\/h3>\n<p>\u9ad8\u5e76\u53d1\u4e0b\uff0cOpt1\u3001Opt2\u3001Opt3 \u4e09\u8005\u975e\u5e38\u63a5\u8fd1\uff1a<\/p>\n<ul>\n<li><strong>Opt1<\/strong>\uff1a6.523 req\/s<\/li>\n<li><strong>Opt2<\/strong>\uff1a6.581 req\/s<\/li>\n<li><strong>Opt3<\/strong>\uff1a6.497 req\/s<\/li>\n<\/ul>\n<p>Opt2 \u76f8\u6bd4 Opt3 \u53ea\u9ad8\u4e86\u7ea6 <strong>1.3%<\/strong>\uff0c\u800c Opt1 \u4e5f\u53ea\u9ad8\u4e86\u7ea6<br \/>\n<strong>0.4%<\/strong>\u3002\u8fd9\u8bf4\u660e\u5728\u9ad8\u5e76\u53d1\u573a\u666f\u91cc\uff0c<strong>optimization_level<br \/>\n\u4ece 1 \u5230 3 \u7684\u6536\u76ca\u5df2\u7ecf\u6ca1\u6709\u4ece 0 \u5230 1<br \/>\n\u90a3\u4e48\u5938\u5f20<\/strong>\uff0c\u7cfb\u7edf\u66f4\u50cf\u662f\u5728\u4e00\u4e2a\u9ad8\u4f4d\u5e73\u53f0\u4e0a\u5fae\u8c03\u3002<\/p>\n<h3 id=\"yarnperformance_modeep-\u7684\u5f71\u54cd\u4f9d\u7136\u8f83\u5c0f\">4.<br \/>\nYaRN\u3001performance_mode\u3001EP \u7684\u5f71\u54cd\u4f9d\u7136\u8f83\u5c0f<\/h3>\n<p>\u9ad8\u5e76\u53d1\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>\u5173\u95ed YaRN<\/strong>\uff1a\u6bd4 Opt3 \u7565\u597d\u4e00\u70b9<\/li>\n<li><strong>Balanced \/ Interactivity \u6a21\u5f0f<\/strong>\uff1a\u90fd\u6bd4 Opt3<br \/>\n\u5c0f\u5e45\u6539\u5584<\/li>\n<li><strong>\u5173\u95ed EP<\/strong>\uff1a\u7565\u5dee\u4e00\u70b9<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u53d8\u5316\u6574\u4f53\u90fd\u5728\u5c0f\u8303\u56f4\u5185\u6ce2\u52a8\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c<strong>\u5173\u95ed MTP<br \/>\n\u5e26\u6765\u7684\u5dee\u5f02\u6700\u6e05\u6670\uff0c\u4e5f\u6700\u6709\u5de5\u7a0b\u610f\u4e49\u3002<\/strong><\/p>\n<hr \/>\n<h2 id=\"\u516d\u8fd9\u6b21\u6d4b\u8bd5\u8ba9\u6211\u771f\u6b63\u8bb0\u4f4f\u7684\u51e0\u70b9\">\u516d\u3001\u8fd9\u6b21\u6d4b\u8bd5\u8ba9\u6211\u771f\u6b63\u8bb0\u4f4f\u7684\u51e0\u70b9<\/h2>\n<p>\u505a\u5b8c\u8fd9\u8f6e\u5b9e\u9a8c\u4ee5\u540e\uff0c\u6211\u89c9\u5f97\u81ea\u5df1\u5bf9 vLLM<br \/>\n\u7684\u7406\u89e3\u6bd4\u4e4b\u524d\u6e05\u6670\u5f88\u591a\uff0c\u5c24\u5176\u662f\u4e0b\u9762\u51e0\u70b9\u3002<\/p>\n<h3 id=\"\u4e0d\u662f\u6240\u6709\u53c2\u6570\u90fd\u503c\u5f97\u540c\u7b49\u5173\u6ce8\">1. \u4e0d\u662f\u6240\u6709\u53c2\u6570\u90fd\u503c\u5f97\u540c\u7b49\u5173\u6ce8<\/h3>\n<p>\u5728\u6ca1\u6709\u7cfb\u7edf\u6d4b\u8bd5\u4e4b\u524d\uff0c\u5f88\u5bb9\u6613\u5bf9\u4e00\u5806\u53c2\u6570\u90fd\u6709\u201c\u662f\u4e0d\u662f\u8be5\u8c03\u4e00\u4e0b\u201d\u7684\u51b2\u52a8\u3002\u4f46\u8fd9\u8f6e\u7ed3\u679c\u544a\u8bc9\u6211\uff1a<\/p>\n<ul>\n<li>\u771f\u6b63\u5173\u952e\u7684\uff0c\u4f18\u5148\u770b<br \/>\n<strong><code>optimization_level<\/code><\/strong><\/li>\n<li>\u5176\u6b21\u8981\u91cd\u70b9\u5173\u6ce8 <strong><code>MTP<\/code><\/strong><\/li>\n<li><code>performance_mode<\/code>\u3001<code>YaRN<\/code>\u3001<code>expert parallel<\/code><br \/>\n\u66f4\u9002\u5408\u5728\u524d\u4e24\u8005\u7a33\u5b9a\u540e\u505a\u7ec6\u8c03<\/li>\n<\/ul>\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u8c03\u53c2\u4e0d\u662f\u201c\u628a\u6240\u6709\u5f00\u5173\u90fd\u8bd5\u4e00\u904d\u201d\uff0c\u800c\u662f\u5e94\u8be5\u5148\u627e\u51fa\u771f\u6b63\u9ad8\u5f71\u54cd\u7684\u53d8\u91cf\u3002<\/p>\n<h3 id=\"\u540c\u4e00\u4e2a\u53c2\u6570\u5728\u4e0d\u540c\u8d1f\u8f7d\u4e0b\u53ef\u80fd\u7ed3\u8bba\u76f8\u53cd\">2.<br \/>\n\u540c\u4e00\u4e2a\u53c2\u6570\uff0c\u5728\u4e0d\u540c\u8d1f\u8f7d\u4e0b\u53ef\u80fd\u7ed3\u8bba\u76f8\u53cd<\/h3>\n<p>MTP \u662f\u8fd9\u6b21\u6700\u5178\u578b\u7684\u4f8b\u5b50\uff1a<\/p>\n<ul>\n<li><strong>\u4f4e\u5e76\u53d1\u4e0b\uff0c\u5f00\u7740\u66f4\u597d<\/strong><\/li>\n<li><strong>\u9ad8\u5e76\u53d1\u4e0b\uff0c\u5173\u6389\u66f4\u597d<\/strong><\/li>\n<\/ul>\n<p>\u8fd9\u4ef6\u4e8b\u5f88\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u63d0\u9192\u6211\uff1a<\/p>\n<blockquote><p>\u63a8\u7406\u90e8\u7f72\u91cc\u7684\u201c\u6700\u4f18\u914d\u7f6e\u201d\u4e0d\u662f\u7edd\u5bf9\u7684\uff0c\u800c\u662f\u548c\u4f60\u7684\u4f7f\u7528\u573a\u666f\u7ed1\u5b9a\u7684\u3002<\/p><\/blockquote>\n<p>\u5982\u679c\u4e0d\u533a\u5206\u573a\u666f\uff0c\u53ea\u770b\u5355\u4e00 benchmark\uff0c\u5f88\u5bb9\u6613\u5f97\u51fa\u9519\u8bef\u7ed3\u8bba\u3002<\/p>\n<h3 id=\"\u672c\u5730\u90e8\u7f72\u5b66\u4e60\u7684\u4ef7\u503c\u4e0d\u53ea\u662f\u8dd1\u8d77\u6765\">3.<br \/>\n\u672c\u5730\u90e8\u7f72\u5b66\u4e60\u7684\u4ef7\u503c\uff0c\u4e0d\u53ea\u662f\u8dd1\u8d77\u6765<\/h3>\n<p>\u6211\u8fd9\u6b21\u505a\u8fd9\u8f6e\u6d4b\u8bd5\uff0c\u6700\u5927\u7684\u6536\u83b7\u5176\u5b9e\u4e0d\u662f\u201c\u77e5\u9053\u54ea\u4e2a\u914d\u7f6e\u6700\u5feb\u201d\uff0c\u800c\u662f\u66f4\u5177\u4f53\u5730\u611f\u53d7\u5230\uff1a<\/p>\n<ul>\n<li>\u90e8\u7f72\u53c2\u6570\u4f1a\u600e\u6837\u5f71\u54cd\u670d\u52a1\u884c\u4e3a<\/li>\n<li>\u541e\u5410\u548c\u65f6\u5ef6\u4e4b\u95f4\u600e\u4e48\u6743\u8861<\/li>\n<li>\u4e0d\u540c\u8d1f\u8f7d\u6a21\u5f0f\u4e0b\uff0c\u7cfb\u7edf\u74f6\u9888\u4f1a\u600e\u4e48\u53d8\u5316<\/li>\n<\/ul>\n<p>\u8fd9\u6bd4\u5355\u7eaf\u628a\u6a21\u578b\u542f\u52a8\u6210\u529f\u66f4\u6709\u4ef7\u503c\u3002\u56e0\u4e3a\u771f\u6b63\u505a\u5de5\u7a0b\u65f6\uff0c\u6700\u540e\u62fc\u7684\u5f80\u5f80\u4e0d\u662f\u201c\u80fd\u4e0d\u80fd\u8dd1\u201d\uff0c\u800c\u662f\u201c\u80fd\u4e0d\u80fd\u89e3\u91ca\u4e3a\u4ec0\u4e48\u8fd9\u6837\u914d\u201d\u3002<\/p>\n<hr \/>\n<h2 id=\"\u9644\u672c\u6b21\u5b9e\u9a8c\u4e2d\u6211\u6700\u5173\u5fc3\u7684\u51e0\u4e2a\u7ed3\u8bba\u7b80\u7248\">\u9644\uff1a\u672c\u6b21\u5b9e\u9a8c\u4e2d\u6211\u6700\u5173\u5fc3\u7684\u51e0\u4e2a\u7ed3\u8bba\uff08\u7b80\u7248\uff09<\/h2>\n<ul>\n<li><strong>Opt0 \u4e0d\u63a8\u8350<\/strong>\uff0c\u4f4e\u5e76\u53d1\u548c\u9ad8\u5e76\u53d1\u90fd\u660e\u663e\u843d\u540e\u3002<\/li>\n<li><strong>Opt2 \/ Opt3<br \/>\n\u5df2\u7ecf\u5f88\u63a5\u8fd1<\/strong>\uff0c\u4e0d\u662f\u6240\u6709\u573a\u666f\u90fd\u9700\u8981\u6267\u7740\u4e8e\u62c9\u6ee1\u3002<\/li>\n<li><strong>\u4f4e\u5e76\u53d1\u4fdd\u7559 MTP \u66f4\u597d<\/strong>\u3002<\/li>\n<li><strong>\u9ad8\u5e76\u53d1\u5173\u95ed MTP \u66f4\u597d<\/strong>\u3002<\/li>\n<li><code>YaRN<\/code>\u3001<code>Balanced \/ Interactivity mode<\/code>\u3001<code>expert parallel<\/code><br \/>\n\u6709\u5f71\u54cd\uff0c\u4f46\u4e0d\u662f\u7b2c\u4e00\u4f18\u5148\u7ea7\u3002<\/li>\n<\/ul>\n<p>\u5982\u679c\u4f60\u4e5f\u5728\u505a\u7c7b\u4f3c\u7684\u672c\u5730 vLLM<br \/>\n\u90e8\u7f72\u6d4b\u8bd5\uff0c\u5e0c\u671b\u8fd9\u7bc7\u8bb0\u5f55\u80fd\u7ed9\u4f60\u4e00\u70b9\u53c2\u8003\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6211\u5728\u672c\u5730\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668\u4e0a\u505a\u4e86\u4e00\u6b21vLLM \u90e8\u7f72\u6d4b\u8bd5\uff1a\u54ea\u4e9b\u53c2\u6570\u771f\u7684\u6709\u7528\uff1f \u6700\u8fd1\u6211\u5728\u672c\u5730\u4e00\u53f0\u5355\u8282\u70b9\u53cc\u5361\u670d\u52a1\u5668\u4e0a\u505a [&hellip;]<\/p>\n","protected":false},"author":2,"featured_media":117,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[37,36],"class_list":["post-112","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-mtsb","tag-llm","tag-vllm"],"_links":{"self":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/112","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/comments?post=112"}],"version-history":[{"count":4,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/112\/revisions"}],"predecessor-version":[{"id":116,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/posts\/112\/revisions\/116"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media\/117"}],"wp:attachment":[{"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/media?parent=112"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/categories?post=112"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bojinhu.xyz\/index.php\/wp-json\/wp\/v2\/tags?post=112"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}