{"id":104710,"date":"2026-09-13T02:57:09","date_gmt":"2026-09-12T18:57:09","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/104710.html"},"modified":"2026-09-13T02:57:09","modified_gmt":"2026-09-12T18:57:09","slug":"%e3%80%90%e6%8e%a8%e7%90%86%e5%bc%95%e6%93%8e%e3%80%91%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8e%a8%e7%90%86%e6%9c%8d%e5%8a%a1%e9%83%a8%e7%bd%b2%e3%80%81%e6%80%a7%e8%83%bd%e8%b0%83%e4%bc%98%e4%b8%8e%e7%94%9f","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/104710.html","title":{"rendered":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM"},"content":{"rendered":"<h2>\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM<\/h2>\n<\/p>\n<h4>\u6587\u7ae0\u76ee\u5f55<\/h4>\n<ul>\n<li>\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM<\/li>\n<li>\n<ul>\n<li>\u4e00\u3001\u63a8\u7406\u5f15\u64ce&#xff1a;\u9009\u578b\u4e0e\u8fd0\u884c\u539f\u7406&#xff08;SGLang\u3001vLLM\u3001TensorRT-LLM&#xff09;<\/li>\n<li>\n<ul>\n<li>1. \u63a8\u7406\u5f15\u64ce\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/li>\n<li>2. \u4e09\u79cd\u5f15\u64ce\u7684\u5173\u952e\u6267\u884c\u8def\u5f84<\/li>\n<li>3. \u524d\u7f00\u7f13\u5b58\u4e0e KV Cache<\/li>\n<\/ul>\n<\/li>\n<li>\u4e8c\u3001\u6a21\u578b\u90e8\u7f72&#xff1a;\u73af\u5883\u51c6\u5907\u3001\u670d\u52a1\u90e8\u7f72\u4e0e API \u8c03\u7528<\/li>\n<li>\n<ul>\n<li>1. \u73af\u5883\u4e0e\u7248\u672c\u68c0\u67e5<\/li>\n<li>2. SGLang Docker \u5355\u673a\u90e8\u7f72<\/li>\n<li>3. vLLM Docker \u90e8\u7f72<\/li>\n<li>4. TensorRT-LLM Engine \u6784\u5efa\u4e0e\u670d\u52a1<\/li>\n<li>5. Kubernetes \u90e8\u7f72\u9aa8\u67b6&#xff08;\u4ee5 SGLang \u4e3a\u4f8b&#xff09;<\/li>\n<li>6. \u4ece\u7f51\u5173\u8bbf\u95ee\u63a8\u7406\u670d\u52a1<\/li>\n<li>7. OpenAI \u517c\u5bb9 API \u8c03\u7528<\/li>\n<li>8. \u91c7\u6837\u53c2\u6570\u5982\u4f55\u5f71\u54cd\u7ed3\u679c<\/li>\n<li>9. \u4f7f\u7528\u5f15\u64ce\u539f\u751f\u524d\u7aef<\/li>\n<\/ul>\n<\/li>\n<li>\u4e09\u3001\u6027\u80fd\u8c03\u4f18&#xff1a;\u4ece\u663e\u5b58\u5230\u541e\u5410&#xff08;\u6307\u6807\u3001\u91cf\u5316\u4e0e PD \u5206\u79bb&#xff09;<\/li>\n<li>\n<ul>\n<li>1. \u5148\u5efa\u7acb\u6b63\u786e\u7684\u6027\u80fd\u6307\u6807<\/li>\n<li>2. \u663e\u5b58\u9884\u7b97\u4e0e\u4e0a\u4e0b\u6587\u957f\u5ea6<\/li>\n<li>3. Tensor Parallel\u3001Data Parallel \u4e0e\u526f\u672c<\/li>\n<li>4. \u8c03\u5ea6\u548c\u524d\u7f00\u7f13\u5b58\u8c03\u4f18<\/li>\n<li>5. PD \u5206\u79bb&#xff1a;\u628a Prefill \u4e0e Decode \u62c6\u6210\u72ec\u7acb\u670d\u52a1\u6c60<\/li>\n<li>6. \u4e0e PD \u5206\u79bb\u914d\u5957\u7684\u8fdb\u9636\u4f18\u5316<\/li>\n<li>7. \u91cf\u5316\u3001FlashInfer \u4e0e Kernel \u9009\u62e9<\/li>\n<\/ul>\n<\/li>\n<li>\u56db\u3001\u751f\u4ea7\u5b9e\u8df5&#xff1a;\u751f\u4ea7\u6cbb\u7406\u3001\u76d1\u63a7\u4e0e\u6392\u969c<\/li>\n<li>\n<ul>\n<li>1. \u76d1\u63a7\u548c\u65e5\u5fd7<\/li>\n<li>2. \u544a\u8b66\u548c\u5f39\u6027\u7b56\u7565<\/li>\n<li>3. \u53d1\u5e03\u3001\u5347\u7ea7\u4e0e\u5b89\u5168\u8fb9\u754c<\/li>\n<li>4. \u5e38\u89c1\u6545\u969c\u6392\u67e5<\/li>\n<li>5. \u5e38\u89c1\u95ee\u7b54<\/li>\n<li>6. \u4e00\u4efd\u4e0a\u7ebf\u68c0\u67e5\u6e05\u5355<\/li>\n<\/ul>\n<\/li>\n<li>\u4e94\u3001\u5e73\u53f0\u5316\u6f14\u8fdb&#xff1a;\u591a\u6a21\u578b\u3001\u591a\u79df\u6237\u3001\u5bb9\u91cf\u4e0e\u6210\u672c\u6cbb\u7406<\/li>\n<li>\n<ul>\n<li>1. \u591a\u6a21\u578b\u6ce8\u518c\u4e0e\u667a\u80fd\u8def\u7531<\/li>\n<li>2. \u591a\u79df\u6237\u9694\u79bb\u3001\u914d\u989d\u4e0e\u516c\u5e73\u8c03\u5ea6<\/li>\n<li>3. \u5bb9\u91cf\u89c4\u5212\u4e0e\u5355\u4f4d Token \u6210\u672c<\/li>\n<li>4. \u5f02\u6784 GPU\u3001Engine \u517c\u5bb9\u4e0e\u591a\u5730\u57df\u5bb9\u707e<\/li>\n<li>5. \u4ece\u5355\u670d\u52a1\u5230\u63a8\u7406\u5e73\u53f0\u7684\u6f14\u8fdb\u8def\u7ebf<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<p>\u628a\u4e00\u4e2a\u5927\u6a21\u578b\u8dd1\u8d77\u6765&#xff0c;\u53ea\u80fd\u8bc1\u660e\u6a21\u578b\u80fd\u63a8\u7406&#xff1b;\u628a\u5b83\u4f5c\u4e3a\u7ebf\u4e0a\u670d\u52a1\u7a33\u5b9a\u5730\u63d0\u4f9b\u7ed9\u591a\u4e2a\u8c03\u7528\u65b9&#xff0c;\u8fd8\u8981\u89e3\u51b3\u663e\u5b58\u3001\u5e76\u53d1\u3001\u6d41\u5f0f\u8f93\u51fa\u3001\u6a21\u578b\u52a0\u8f7d\u3001\u76d1\u63a7\u548c\u6545\u969c\u6062\u590d\u7b49\u95ee\u9898\u3002<\/p>\n<p>SGLang\u3001vLLM \u548c TensorRT-LLM \u90fd\u662f\u9762\u5411\u5927\u8bed\u8a00\u6a21\u578b\u548c\u591a\u6a21\u6001\u6a21\u578b\u7684\u63a8\u7406\u5f15\u64ce\u3002\u5b83\u4eec\u901a\u5e38\u63d0\u4f9b OpenAI \u517c\u5bb9\u63a5\u53e3&#xff0c;\u4e5f\u4f1a\u5728\u8bf7\u6c42\u8c03\u5ea6\u3001KV Cache\u3001\u6279\u5904\u7406\u3001\u91cf\u5316\u548c GPU Kernel \u7b49\u5c42\u9762\u505a\u4f18\u5316\u3002\u63a8\u7406\u5f15\u64ce\u7684\u6838\u5fc3\u4ef7\u503c\u4e0d\u662f\u201c\u628a\u6a21\u578b\u5305\u4e00\u5c42 HTTP\u201d&#xff0c;\u800c\u662f\u628a\u6a21\u578b\u6743\u91cd\u3001\u8bf7\u6c42\u961f\u5217\u548c GPU \u6267\u884c\u7ec4\u7ec7\u6210\u53ef\u6301\u7eed\u4f18\u5316\u7684\u8fd0\u884c\u65f6\u3002<\/p>\n<p>\u672c\u6587\u4ee5 Qwen\u3001Llama \u7b49 Hugging Face \u683c\u5f0f\u6a21\u578b\u4e3a\u4f8b&#xff0c;\u8bf4\u660e SGLang\u3001vLLM \u548c TensorRT-LLM \u7684\u5b9a\u4f4d\u5dee\u5f02\u3001\u90e8\u7f72\u65b9\u5f0f\u3001\u5173\u952e\u53c2\u6570\u548c\u751f\u4ea7\u6392\u969c\u8def\u5f84\u3002\u547d\u4ee4\u4e2d\u7684\u7248\u672c\u3001\u6a21\u578b\u540d\u548c GPU \u6570\u91cf\u9700\u8981\u6309\u5b9e\u9645\u73af\u5883\u66ff\u6362&#xff0c;\u751f\u4ea7\u73af\u5883\u5e94\u56fa\u5b9a\u63a8\u7406\u5f15\u64ce\u3001CUDA\u3001\u9a71\u52a8\u548c\u6a21\u578b revision&#xff0c;\u800c\u4e0d\u662f\u76f4\u63a5\u4f7f\u7528 latest\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260912185707-6aa5a0836e6a6.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<h3>\u4e00\u3001\u63a8\u7406\u5f15\u64ce&#xff1a;\u9009\u578b\u4e0e\u8fd0\u884c\u539f\u7406&#xff08;SGLang\u3001vLLM\u3001TensorRT-LLM&#xff09;<\/h3>\n<h4>1. \u63a8\u7406\u5f15\u64ce\u89e3\u51b3\u4ec0\u4e48\u95ee\u9898<\/h4>\n<p>\u4e00\u4e2a\u63a8\u7406\u8bf7\u6c42\u901a\u5e38\u7ecf\u5386\u4ee5\u4e0b\u9636\u6bb5&#xff1a;\u8bfb\u53d6\u804a\u5929\u6d88\u606f\u3001\u5206\u8bcd\u3001\u6784\u9020 KV Cache\u3001\u8c03\u5ea6\u6279\u6b21\u3001\u6267\u884c Prefill\u3001\u9010 Token Decode\u3001\u91c7\u6837\u5e76\u8fd4\u56de\u7ed3\u679c\u3002\u5e76\u53d1\u4e00\u9ad8&#xff0c;\u771f\u6b63\u7684\u74f6\u9888\u5f80\u5f80\u4e0d\u662f\u6a21\u578b\u524d\u5411\u672c\u8eab&#xff0c;\u800c\u662f\u8bf7\u6c42\u6392\u961f\u3001\u663e\u5b58\u788e\u7247\u3001\u91cd\u590d\u8ba1\u7b97\u548c\u4e0d\u540c\u957f\u5ea6\u8bf7\u6c42\u4e4b\u95f4\u7684\u76f8\u4e92\u963b\u585e\u3002<\/p>\n<p>\u4e09\u79cd\u5f15\u64ce\u90fd\u5c06\u8fd9\u4e9b\u901a\u7528\u80fd\u529b\u653e\u5165 runtime&#xff0c;\u5e76\u5728\u5176\u4e0a\u63d0\u4f9b\u5728\u7ebf API \u6216\u7f16\u7a0b\u5165\u53e3&#xff1a;<\/p>\n<table>\n<tr>\u5165\u53e3\u4f5c\u7528\u9002\u7528\u573a\u666f<\/tr>\n<tbody>\n<tr>\n<td>OpenAI \u517c\u5bb9 API<\/td>\n<td>\u901a\u8fc7 \/v1\/chat\/completions\u3001\/v1\/completions \u63d0\u4f9b\u5728\u7ebf\u63a8\u7406<\/td>\n<td>\u63a5\u5165\u73b0\u6709 SDK\u3001\u7f51\u5173\u548c\u4e1a\u52a1\u670d\u52a1<\/td>\n<\/tr>\n<tr>\n<td>\u5f15\u64ce\u539f\u751f API<\/td>\n<td>\u7528 SGLang Python \u524d\u7aef\u3001vLLM Python API \u6216 TensorRT-LLM Runtime \u7f16\u6392\u63a8\u7406<\/td>\n<td>Agent\u3001\u79bb\u7ebf\u6279\u5904\u7406\u3001\u590d\u6742\u751f\u6210\u903b\u8f91<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e09\u79cd\u5f15\u64ce\u7684\u4f18\u5316\u4fa7\u91cd\u70b9\u5e76\u4e0d\u5b8c\u5168\u76f8\u540c&#xff1a;<\/p>\n<table>\n<tr>\u5bf9\u6bd4\u9879SGLangvLLMTensorRT-LLM<\/tr>\n<tbody>\n<tr>\n<td>\u670d\u52a1\u63a5\u53e3<\/td>\n<td>OpenAI \u517c\u5bb9 API &#043; Python \u524d\u7aef<\/td>\n<td>OpenAI \u517c\u5bb9 API &#043; Python API<\/td>\n<td>TensorRT-LLM Runtime &#043; trtllm-serve \u7b49\u670d\u52a1\u5165\u53e3<\/td>\n<\/tr>\n<tr>\n<td>\u6838\u5fc3\u4f18\u5316<\/td>\n<td>RadixAttention \u524d\u7f00\u7f13\u5b58\u3001\u8fde\u7eed\u6279\u5904\u7406\u3001\u7ed3\u6784\u5316\u751f\u6210<\/td>\n<td>PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u3001\u5e7f\u6cdb\u6a21\u578b\u9002\u914d<\/td>\n<td>TensorRT-LLM Engine\u3001\u878d\u5408 Kernel\u3001\u91cf\u5316\u548c CUDA Graph<\/td>\n<\/tr>\n<tr>\n<td>\u5178\u578b\u4f18\u52bf<\/td>\n<td>\u591a\u8f6e\u5bf9\u8bdd\u3001\u91cd\u590d\u524d\u7f00\u548c\u7ed3\u6784\u5316\u751f\u6210<\/td>\n<td>\u901a\u7528\u5728\u7ebf\u670d\u52a1\u3001\u751f\u6001\u6210\u719f\u3001\u4e0a\u624b\u5feb<\/td>\n<td>NVIDIA GPU \u4e0a\u8ffd\u6c42\u6781\u81f4\u541e\u5410\u548c\u4f4e\u5ef6\u8fdf<\/td>\n<\/tr>\n<tr>\n<td>\u4e3b\u8981\u4ee3\u4ef7<\/td>\n<td>\u7248\u672c\u53d8\u5316\u5feb&#xff0c;\u9700\u8981\u9a8c\u8bc1\u6a21\u578b\u9002\u914d<\/td>\n<td>\u7279\u5b9a\u6a21\u578b\u548c\u573a\u666f\u672a\u5fc5\u8fbe\u5230\u6700\u4f18<\/td>\n<td>Engine \u6784\u5efa\u3001\u7248\u672c\u7ed1\u5b9a\u548c\u90e8\u7f72\u590d\u6742\u5ea6\u66f4\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>\u9009\u578b\u5efa\u8bae<\/td>\n<td>\u7528\u771f\u5b9e\u4e1a\u52a1\u538b\u6d4b TTFT\u3001TPOT\u3001\u7f13\u5b58\u547d\u4e2d\u7387\u548c\u6210\u672c<\/td>\n<td>\u7528\u771f\u5b9e\u4e1a\u52a1\u538b\u6d4b\u541e\u5410\u3001P99 \u548c\u8fd0\u7ef4\u6210\u672c<\/td>\n<td>\u7528\u76ee\u6807 GPU\u3001\u76ee\u6807\u7cbe\u5ea6\u548c\u56fa\u5b9a\u6a21\u578b\u6784\u5efa\u540e\u518d\u6bd4\u8f83<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8fd9\u91cc\u7684\u6bd4\u8f83\u4e0d\u662f\u4e3a\u4e86\u8bc1\u660e\u54ea\u4e2a\u6846\u67b6\u201c\u7edd\u5bf9\u66f4\u5feb\u201d\u3002\u6a21\u578b\u67b6\u6784\u3001\u8f93\u5165\u957f\u5ea6\u3001\u8f93\u51fa\u957f\u5ea6\u3001\u5e76\u53d1\u5ea6\u3001GPU \u578b\u53f7\u548c\u91cf\u5316\u65b9\u5f0f\u90fd\u4f1a\u6539\u53d8\u7ed3\u679c\u3002<\/p>\n<h4>2. \u4e09\u79cd\u5f15\u64ce\u7684\u5173\u952e\u6267\u884c\u8def\u5f84<\/h4>\n<p>\u5ba2\u6237\u7aef \/ AI Gateway<br \/>\n       \u2502 OpenAI API \u6216\u5f15\u64ce\u539f\u751f\u8bf7\u6c42<br \/>\n       \u25bc<br \/>\nSGLang \/ vLLM \/ TensorRT-LLM Server<br \/>\n       \u251c\u2500 \u8bf7\u6c42\u89e3\u6790\u3001\u9274\u6743\u8fb9\u754c\u548c\u53c2\u6570\u6821\u9a8c<br \/>\n       \u251c\u2500 Tokenizer \u4e0e Prefill \u8c03\u5ea6<br \/>\n       \u251c\u2500 Prefix Cache \u6216 Paged KV Cache \u67e5\u8be2<br \/>\n       \u251c\u2500 Continuous Batching \u52a8\u6001\u7ec4\u6279<br \/>\n       \u251c\u2500 GPU Worker \u6267\u884c Prefill \u4e0e Decode<br \/>\n       \u2514\u2500 SSE \u6d41\u5f0f\u8fd4\u56de Token<br \/>\n       \u2502<br \/>\n       \u251c\u2500 \/health\u3001\/metrics\u3001\u8fd0\u884c\u65e5\u5fd7<br \/>\n       \u2514\u2500 Prometheus\u3001Grafana\u3001\u65e5\u5fd7\u5e73\u53f0<\/p>\n<p>Prefill \u8d1f\u8d23\u4e00\u6b21\u6027\u5904\u7406\u8f93\u5165\u4e0a\u4e0b\u6587&#xff0c;\u8ba1\u7b97\u91cf\u5927&#xff0c;\u4e3b\u8981\u5f71\u54cd\u9996 Token \u5ef6\u8fdf&#xff08;TTFT&#xff09;&#xff1b;Decode \u6309 Token \u8fed\u4ee3\u751f\u6210&#xff0c;\u4e3b\u8981\u5f71\u54cd\u6bcf Token \u5ef6\u8fdf&#xff08;TPOT&#xff09;\u548c\u8f93\u51fa\u541e\u5410\u3002SGLang \u66f4\u5f3a\u8c03 RadixAttention \u524d\u7f00\u590d\u7528&#xff0c;vLLM \u4ee5 PagedAttention \u7ba1\u7406 KV Cache&#xff0c;TensorRT-LLM \u5219\u4f9d\u8d56\u6784\u5efa\u597d\u7684 Engine\u3001\u878d\u5408 Kernel \u548c CUDA Graph\u3002\u628a\u957f\u4e0a\u4e0b\u6587\u8bf7\u6c42\u548c\u77ed\u8bf7\u6c42\u6df7\u5728\u4e00\u8d77\u65f6&#xff0c;\u4e09\u79cd\u5f15\u64ce\u90fd\u9700\u8981\u5728\u541e\u5410\u4e0e\u4ea4\u4e92\u5ef6\u8fdf\u4e4b\u95f4\u505a\u53d6\u820d\u3002<\/p>\n<h4>3. \u524d\u7f00\u7f13\u5b58\u4e0e KV Cache<\/h4>\n<p>\u591a\u8f6e\u5bf9\u8bdd\u3001\u7cfb\u7edf\u63d0\u793a\u8bcd\u548c\u6279\u91cf\u6587\u6863\u95ee\u7b54\u7ecf\u5e38\u5171\u4eab\u4e00\u6bb5\u524d\u7f00\u3002\u5982\u679c\u6bcf\u6b21\u8bf7\u6c42\u90fd\u4ece\u5934\u8ba1\u7b97&#xff0c;GPU \u4f1a\u91cd\u590d\u6267\u884c\u76f8\u540c\u7684 Prefill\u3002SGLang \u7684 RadixAttention \u4f1a\u5c06\u5df2\u8ba1\u7b97\u7684 Token \u524d\u7f00\u7ec4\u7ec7\u6210\u53ef\u590d\u7528\u7684\u6811\u7ed3\u6784&#xff1b;vLLM \u4e5f\u652f\u6301 Prefix Caching&#xff0c;\u4f46\u5b9e\u73b0\u548c\u53c2\u6570\u540d\u79f0\u968f\u7248\u672c\u53d8\u5316&#xff1b;TensorRT-LLM \u5219\u901a\u5e38\u7ed3\u5408 KV Cache \u7ba1\u7406\u3001Paged Context FMHA \u548c CUDA Graph \u7b49\u80fd\u529b\u4f18\u5316\u6267\u884c\u3002<\/p>\n<p>\u524d\u7f00\u7f13\u5b58\u80fd\u5426\u751f\u6548&#xff0c;\u53d6\u51b3\u4e8e\u8f93\u5165 Token \u662f\u5426\u771f\u7684\u76f8\u540c&#xff1a;\u7cfb\u7edf\u63d0\u793a\u8bcd\u7684\u7a7a\u683c\u3001\u5de5\u5177\u63cf\u8ff0\u7684\u6392\u5e8f\u3001\u6d88\u606f\u5e8f\u5217\u5316\u65b9\u5f0f\u548c\u6a21\u677f\u7248\u672c\u53d1\u751f\u53d8\u5316&#xff0c;\u90fd\u4f1a\u5bfc\u81f4\u7f13\u5b58\u672a\u547d\u4e2d\u3002\u56e0\u6b64\u5e94\u56fa\u5b9a Chat Template&#xff0c;\u628a\u7a33\u5b9a\u5185\u5bb9\u653e\u5728\u524d\u9762&#xff0c;\u628a\u7528\u6237\u52a8\u6001\u5185\u5bb9\u653e\u5728\u540e\u9762&#xff0c;\u5e76\u5728\u6307\u6807\u4e2d\u89c2\u5bdf\u547d\u4e2d\u7387&#xff0c;\u800c\u4e0d\u662f\u53ea\u51ed\u611f\u89c9\u5224\u65ad\u3002<\/p>\n<p>\u7f13\u5b58\u4e0d\u662f\u65e0\u9650\u7684\u3002\u63a8\u7406\u5f15\u64ce\u4f1a\u5728\u663e\u5b58\u9884\u7b97\u5185\u7ba1\u7406 KV Cache&#xff0c;\u7f13\u5b58\u8fc7\u591a\u4f1a\u6324\u538b\u6a21\u578b\u6743\u91cd\u548c\u65b0\u8bf7\u6c42\u7a7a\u95f4\u3002\u751f\u4ea7\u73af\u5883\u9700\u8981\u540c\u65f6\u5173\u6ce8\u7f13\u5b58\u547d\u4e2d\u7387\u3001KV Cache \u4f7f\u7528\u7387\u3001\u8bf7\u6c42\u6392\u961f\u65f6\u95f4\u548c OOM \u6b21\u6570\u3002\u4e0d\u8981\u628a SGLang \u7684\u7f13\u5b58\u53c2\u6570\u76f4\u63a5\u7167\u642c\u5230 vLLM \u6216 TensorRT-LLM&#xff0c;\u5e94\u4ee5\u5b9e\u9645\u7248\u672c\u7684\u542f\u52a8\u5e2e\u52a9\u548c\u6307\u6807\u4e3a\u51c6\u3002<\/p>\n<h3>\u4e8c\u3001\u6a21\u578b\u90e8\u7f72&#xff1a;\u73af\u5883\u51c6\u5907\u3001\u670d\u52a1\u90e8\u7f72\u4e0e API \u8c03\u7528<\/h3>\n<p>\u672c\u7ae0\u6309\u7167\u6a21\u578b\u670d\u52a1\u4ece\u51c6\u5907\u5230\u53ef\u8c03\u7528\u7684\u987a\u5e8f\u5c55\u5f00&#xff1a;\u5148\u6821\u9a8c\u8fd0\u884c\u73af\u5883&#xff0c;\u518d\u5206\u522b\u542f\u52a8\u4e09\u79cd\u63a8\u7406\u5f15\u64ce&#xff0c;\u6700\u540e\u901a\u8fc7\u7edf\u4e00 API \u548c\u7f51\u5173\u5b8c\u6210\u9a8c\u8bc1\u3002<\/p>\n<h4>1. \u73af\u5883\u4e0e\u7248\u672c\u68c0\u67e5<\/h4>\n<p>\u4e0a\u7ebf\u524d\u5148\u8bb0\u5f55\u4ee5\u4e0b\u4fe1\u606f&#xff0c;\u540e\u7eed\u6392\u969c\u65f6\u4e0d\u8981\u53ea\u770b\u201cCUDA \u80fd\u4e0d\u80fd\u7528\u201d&#xff1a;<\/p>\n<p>nvidia-smi<br \/>\npython3 <span class=\"token parameter variable\">&#8211;version<\/span><br \/>\npython3 <span class=\"token parameter variable\">-c<\/span> <span class=\"token string\">&#039;import torch; print(torch.__version__, torch.version.cuda, torch.cuda.device_count())&#039;<\/span><\/p>\n<p>\u9700\u8981\u786e\u8ba4\u7684\u517c\u5bb9\u5173\u7cfb\u5305\u62ec&#xff1a;NVIDIA \u9a71\u52a8\u662f\u5426\u652f\u6301\u76ee\u6807 CUDA\u3001PyTorch \u4e0e\u63a8\u7406\u5f15\u64ce\u7684 CUDA \u6784\u5efa\u662f\u5426\u5339\u914d\u3001GPU \u663e\u5b58\u662f\u5426\u80fd\u5bb9\u7eb3\u6a21\u578b\u6743\u91cd\u548c KV Cache\u3001\u8de8\u5361\u8fde\u63a5\u662f\u5426\u5177\u5907 NVLink \u6216\u9ad8\u901f PCIe&#xff0c;\u4ee5\u53ca\u5bb9\u5668\u7684 \/dev\/shm \u662f\u5426\u8db3\u591f\u3002TensorRT-LLM \u8fd8\u8981\u989d\u5916\u6821\u9a8c TensorRT\u3001CUDA\u3001\u9a71\u52a8\u548c Engine \u6784\u5efa\u73af\u5883\u7684\u4e00\u81f4\u6027\u3002<\/p>\n<p>\u6a21\u578b\u4e0b\u8f7d\u76ee\u5f55\u6700\u597d\u4f7f\u7528\u672c\u5730 NVMe \u6216\u9ad8\u6027\u80fd\u5171\u4eab\u5b58\u50a8&#xff0c;\u5e76\u63d0\u524d\u5b8c\u6210\u6743\u9650\u3001\u6821\u9a8c\u548c\u7f13\u5b58\u3002\u7b2c\u4e00\u6b21\u542f\u52a8\u4f1a\u52a0\u8f7d\u6743\u91cd\u3001\u521d\u59cb\u5316 CUDA Kernel \u548c\u5efa\u7acb\u901a\u4fe1\u7ec4&#xff0c;\u4e0d\u80fd\u628a\u201c\u9996\u6b21\u542f\u52a8\u65f6\u95f4\u201d\u5f53\u4f5c\u6b63\u5e38\u8bf7\u6c42\u5ef6\u8fdf\u3002<\/p>\n<h4>2. SGLang Docker \u5355\u673a\u90e8\u7f72<\/h4>\n<p>\u5f00\u53d1\u548c\u538b\u6d4b\u53ef\u4ee5\u4f7f\u7528\u5b98\u65b9\u955c\u50cf\u3002\u751f\u4ea7\u73af\u5883\u628a &lt;version&gt; \u66ff\u6362\u4e3a\u7ecf\u8fc7\u9a8c\u8bc1\u7684\u56fa\u5b9a\u6807\u7b7e&#xff0c;\u5e76\u5c06\u6a21\u578b\u901a\u8fc7\u53ea\u8bfb\u5377\u6302\u8f7d\u5230\u5bb9\u5668&#xff1a;<\/p>\n<p><span class=\"token function\">docker<\/span> run <span class=\"token parameter variable\">&#8211;rm<\/span> <span class=\"token parameter variable\">&#8211;gpus<\/span> all <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;ipc<\/span><span class=\"token operator\">&#061;<\/span>host <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;shm-size<span class=\"token operator\">&#061;<\/span>32g <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;network<\/span> <span class=\"token function\">host<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-v<\/span> \/data\/models\/Qwen2.5-32B-Instruct:\/models\/Qwen2.5-32B-Instruct:ro <span class=\"token punctuation\">\\\\<\/span><br \/>\n  lmsysorg\/sglang:<span class=\"token operator\">&lt;<\/span>version<span class=\"token operator\">&gt;<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  python3 <span class=\"token parameter variable\">-m<\/span> sglang.launch_server <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;model-path \/models\/Qwen2.5-32B-Instruct <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;served-model-name qwen2.5-32b <span class=\"token punctuation\">\\\\<\/span><br \/>\n    <span class=\"token parameter variable\">&#8211;host<\/span> <span class=\"token number\">0.0<\/span>.0.0 <span class=\"token punctuation\">\\\\<\/span><br \/>\n    <span class=\"token parameter variable\">&#8211;port<\/span> <span class=\"token number\">30000<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;tp-size <span class=\"token number\">4<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;context-length <span class=\"token number\">32768<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;mem-fraction-static <span class=\"token number\">0.85<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n    &#8211;enable-metrics<\/p>\n<p>\u51e0\u4e2a\u53c2\u6570\u7684\u542b\u4e49&#xff1a;<\/p>\n<table>\n<tr>\u53c2\u6570\u4f5c\u7528\u4f7f\u7528\u5efa\u8bae<\/tr>\n<tbody>\n<tr>\n<td>&#8211;model-path<\/td>\n<td>\u672c\u5730\u76ee\u5f55\u6216 Hugging Face \u6a21\u578b\u8def\u5f84<\/td>\n<td>\u751f\u4ea7\u4f7f\u7528\u56fa\u5b9a revision \u7684\u672c\u5730\u526f\u672c<\/td>\n<\/tr>\n<tr>\n<td>&#8211;served-model-name<\/td>\n<td>API \u4e2d\u66b4\u9732\u7684\u6a21\u578b\u540d<\/td>\n<td>\u4e0e\u7f51\u5173\u8def\u7531\u548c\u8ba1\u8d39\u53e3\u5f84\u4fdd\u6301\u4e00\u81f4<\/td>\n<\/tr>\n<tr>\n<td>&#8211;tp-size<\/td>\n<td>Tensor Parallel \u5e76\u884c\u5ea6<\/td>\n<td>\u901a\u5e38\u4e0e\u53c2\u4e0e\u63a8\u7406\u7684 GPU \u6570\u91cf\u4e00\u81f4<\/td>\n<\/tr>\n<tr>\n<td>&#8211;context-length<\/td>\n<td>\u5355\u8bf7\u6c42\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6<\/td>\n<td>\u6309\u4e1a\u52a1\u9700\u8981\u8bbe\u7f6e&#xff0c;\u8d8a\u5927\u8d8a\u5360 KV Cache<\/td>\n<\/tr>\n<tr>\n<td>&#8211;mem-fraction-static<\/td>\n<td>\u9759\u6001\u663e\u5b58\u9884\u7b97\u6bd4\u4f8b<\/td>\n<td>\u7559\u51fa\u901a\u4fe1\u3001\u8fd0\u884c\u65f6\u548c\u788e\u7247\u7a7a\u95f4\u540e\u518d\u63d0\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>&#8211;enable-metrics<\/td>\n<td>\u66b4\u9732 Prometheus \u6307\u6807<\/td>\n<td>\u751f\u4ea7\u73af\u5883\u5e94\u9ed8\u8ba4\u5f00\u542f\u5e76\u7eb3\u5165\u544a\u8b66<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>&#8211;tp-size 4 \u5e76\u4e0d\u610f\u5473\u7740\u56db\u5f20\u5361\u4e00\u5b9a\u6bd4\u4e00\u5f20\u5361\u5feb\u3002Tensor Parallel \u4f1a\u5f15\u5165\u8de8\u5361\u901a\u4fe1&#xff0c;GPU \u62d3\u6251\u3001\u6a21\u578b\u5927\u5c0f\u548c\u8bf7\u6c42\u957f\u5ea6\u90fd\u4f1a\u5f71\u54cd\u6536\u76ca\u3002\u5c0f\u6a21\u578b\u5728 PCIe \u673a\u5668\u4e0a\u5f3a\u884c\u6269\u5927 TP&#xff0c;\u53ef\u80fd\u53cd\u800c\u964d\u4f4e\u541e\u5410\u3002<\/p>\n<h4>3. vLLM Docker \u90e8\u7f72<\/h4>\n<p>vLLM \u53ef\u4ee5\u901a\u8fc7 OpenAI-compatible server \u76f4\u63a5\u63d0\u4f9b\u7edf\u4e00\u7684 \/v1 \u63a5\u53e3\u3002\u4e0b\u9762\u7684\u547d\u4ee4\u9002\u5408\u5355\u673a\u538b\u6d4b\u6216\u4f5c\u4e3a Kubernetes \u5bb9\u5668\u7684\u542f\u52a8\u53c2\u6570&#xff1b;vllm\/vllm-openai \u7684\u955c\u50cf\u6807\u7b7e\u548c\u53c2\u6570\u4f1a\u968f\u7248\u672c\u53d8\u5316&#xff0c;\u6b63\u5f0f\u90e8\u7f72\u524d\u5e94\u4ee5\u5bf9\u5e94\u7248\u672c\u7684 vllm serve &#8211;help \u4e3a\u51c6\u3002<\/p>\n<p><span class=\"token function\">docker<\/span> run <span class=\"token parameter variable\">&#8211;rm<\/span> <span class=\"token parameter variable\">&#8211;gpus<\/span> all <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;ipc<\/span><span class=\"token operator\">&#061;<\/span>host <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;shm-size<span class=\"token operator\">&#061;<\/span>32g <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;network<\/span> <span class=\"token function\">host<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-v<\/span> \/data\/models\/Qwen2.5-32B-Instruct:\/models\/Qwen2.5-32B-Instruct:ro <span class=\"token punctuation\">\\\\<\/span><br \/>\n  vllm\/vllm-openai:<span class=\"token operator\">&lt;<\/span>version<span class=\"token operator\">&gt;<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;model<\/span> \/models\/Qwen2.5-32B-Instruct <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;served-model-name qwen2.5-32b <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;host<\/span> <span class=\"token number\">0.0<\/span>.0.0 <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;port<\/span> <span class=\"token number\">30000<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;tensor-parallel-size <span class=\"token number\">4<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;max-model-len <span class=\"token number\">32768<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;gpu-memory-utilization <span class=\"token number\">0.85<\/span><\/p>\n<p>vLLM \u7684 &#8211;tensor-parallel-size \u4e0e SGLang \u7684 &#8211;tp-size \u8bed\u4e49\u76f8\u8fd1&#xff0c;\u4f46\u4e0d\u80fd\u6df7\u7528\u53c2\u6570\u540d\u3002Prefix Caching\u3001\u91cf\u5316\u683c\u5f0f\u3001\u6700\u5927\u5e76\u53d1\u548c\u8c03\u5ea6\u53c2\u6570\u4e5f\u5e94\u6309 vLLM \u5f53\u524d\u7248\u672c\u7684\u5e2e\u52a9\u4fe1\u606f\u9010\u9879\u786e\u8ba4\u3002\u751f\u4ea7\u73af\u5883\u5efa\u8bae\u4e3a\u955c\u50cf\u3001\u6a21\u578b revision \u548c\u542f\u52a8\u53c2\u6570\u5efa\u7acb\u53ef\u56de\u6eda\u7684\u914d\u7f6e\u6e05\u5355\u3002<\/p>\n<h4>4. TensorRT-LLM Engine \u6784\u5efa\u4e0e\u670d\u52a1<\/h4>\n<p>TensorRT-LLM \u901a\u5e38\u4e0d\u662f\u542f\u52a8\u5bb9\u5668\u540e\u76f4\u63a5\u8bfb\u53d6 Hugging Face \u6743\u91cd&#xff0c;\u800c\u662f\u5148\u628a\u6a21\u578b\u8f6c\u6362\u4e3a checkpoint&#xff0c;\u518d\u9488\u5bf9\u76ee\u6807 GPU\u3001\u7cbe\u5ea6\u548c\u5e76\u884c\u7b56\u7565\u6784\u5efa TensorRT Engine\u3002\u5178\u578b\u6d41\u7a0b\u5982\u4e0b&#xff1a;<\/p>\n<p>Hugging Face \u6a21\u578b<br \/>\n    \u2192 TensorRT-LLM checkpoint \u8f6c\u6362<br \/>\n    \u2192 trtllm-build \u6784\u5efa Engine<br \/>\n    \u2192 trtllm-serve \u6216 Triton Inference Server<br \/>\n    \u2192 OpenAI \u517c\u5bb9 API \/ \u5185\u90e8\u670d\u52a1<\/p>\n<p>\u4ee5\u4e0b\u547d\u4ee4\u662f\u7248\u672c\u76f8\u5173\u793a\u4f8b&#xff0c;\u5177\u4f53\u7684\u8f6c\u6362\u811a\u672c\u8def\u5f84\u3001\u53c2\u6570\u540d\u548c\u670d\u52a1\u5165\u53e3\u4ee5\u6240\u7528 TensorRT-LLM \u7248\u672c\u6587\u6863\u53ca &#8211;help \u8f93\u51fa\u4e3a\u51c6&#xff1a;<\/p>\n<p><span class=\"token comment\"># 1. \u5c06 Hugging Face \u6743\u91cd\u8f6c\u6362\u4e3a TensorRT-LLM checkpoint<\/span><br \/>\npython examples\/convert_checkpoint.py <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;model_dir<\/span> \/models\/Qwen2.5-32B-Instruct <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;output_dir<\/span> \/tmp\/qwen2.5-32b-ckpt <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;dtype<\/span> float16 <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;tp_size<\/span> <span class=\"token number\">4<\/span><\/p>\n<p><span class=\"token comment\"># 2. \u9488\u5bf9\u76ee\u6807 GPU \u6784\u5efa Engine&#xff1b;FP16 \u4e5f\u53ef\u4ee5\u66ff\u6362\u4e3a BF16\u3001FP8 \u6216 INT8<\/span><br \/>\ntrtllm-build <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;checkpoint_dir<\/span> \/tmp\/qwen2.5-32b-ckpt <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;output_dir<\/span> \/engines\/qwen2.5-32b-tp4 <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;gemm_plugin<\/span> float16<\/p>\n<p><span class=\"token comment\"># 3. \u542f\u52a8\u670d\u52a1&#xff08;\u4e0d\u540c\u7248\u672c\u53ef\u80fd\u4f7f\u7528 Triton \u6216\u5176\u4ed6 serving \u5165\u53e3&#xff09;<\/span><br \/>\ntrtllm-serve \/engines\/qwen2.5-32b-tp4 <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;host<\/span> <span class=\"token number\">0.0<\/span>.0.0 <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">&#8211;port<\/span> <span class=\"token number\">30000<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  &#8211;served-model-name qwen2.5-32b<\/p>\n<p>Engine \u4e0e GPU \u67b6\u6784\u3001TensorRT-LLM \u7248\u672c\u3001CUDA \u7248\u672c\u548c\u5e76\u884c\u914d\u7f6e\u5b58\u5728\u7ed1\u5b9a\u5173\u7cfb\u3002\u5347\u7ea7\u5176\u4e2d\u4efb\u4e00\u9879\u90fd\u5e94\u91cd\u65b0\u6784\u5efa\u5e76\u56de\u5f52\u6d4b\u8bd5&#xff0c;\u4e0d\u80fd\u628a\u65e7 Engine \u76f4\u63a5\u590d\u5236\u5230\u65b0\u8fd0\u884c\u65f6\u3002TensorRT-LLM \u9002\u5408\u6a21\u578b\u3001GPU \u548c\u7cbe\u5ea6\u76f8\u5bf9\u56fa\u5b9a\u4e14\u8ffd\u6c42\u6781\u81f4\u541e\u5410\u7684\u573a\u666f&#xff1b;\u6a21\u578b\u9891\u7e41\u53d8\u66f4\u6216\u9700\u8981\u5feb\u901f\u9002\u914d\u65b0\u67b6\u6784\u65f6&#xff0c;SGLang\u3001vLLM \u901a\u5e38\u66f4\u6613\u8fed\u4ee3\u3002<\/p>\n<h4>5. Kubernetes \u90e8\u7f72\u9aa8\u67b6&#xff08;\u4ee5 SGLang \u4e3a\u4f8b&#xff09;<\/h4>\n<p>Kubernetes \u4e2d\u5efa\u8bae\u4e00\u4e2a Pod \u7ed1\u5b9a\u4e00\u7ec4\u56fa\u5b9a GPU&#xff0c;\u5e76\u901a\u8fc7\u8282\u70b9\u6807\u7b7e\u3001RuntimeClass \u548c GPU \u8d44\u6e90\u9650\u5236\u4fdd\u8bc1\u8c03\u5ea6\u7a33\u5b9a\u3002\u4e0b\u9762\u662f\u53ef\u4f5c\u4e3a\u8d77\u70b9\u7684 Deployment&#xff0c;\u6a21\u578b\u76ee\u5f55\u5047\u8bbe\u7531 PVC \u63d0\u4f9b&#xff1a;<\/p>\n<p><span class=\"token key atrule\">apiVersion<\/span><span class=\"token punctuation\">:<\/span> apps\/v1<br \/>\n<span class=\"token key atrule\">kind<\/span><span class=\"token punctuation\">:<\/span> Deployment<br \/>\n<span class=\"token key atrule\">metadata<\/span><span class=\"token punctuation\">:<\/span><br \/>\n  <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n  <span class=\"token key atrule\">labels<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token key atrule\">app<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n<span class=\"token key atrule\">spec<\/span><span class=\"token punctuation\">:<\/span><br \/>\n  <span class=\"token key atrule\">replicas<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">1<\/span><br \/>\n  <span class=\"token key atrule\">strategy<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token key atrule\">type<\/span><span class=\"token punctuation\">:<\/span> Recreate<br \/>\n  <span class=\"token key atrule\">selector<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token key atrule\">matchLabels<\/span><span class=\"token punctuation\">:<\/span><br \/>\n      <span class=\"token key atrule\">app<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n  <span class=\"token key atrule\">template<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token key atrule\">metadata<\/span><span class=\"token punctuation\">:<\/span><br \/>\n      <span class=\"token key atrule\">labels<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token key atrule\">app<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n    <span class=\"token key atrule\">spec<\/span><span class=\"token punctuation\">:<\/span><br \/>\n      <span class=\"token key atrule\">nodeSelector<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token key atrule\">accelerator<\/span><span class=\"token punctuation\">:<\/span> nvidia<span class=\"token punctuation\">&#8211;<\/span>a100<br \/>\n      <span class=\"token key atrule\">terminationGracePeriodSeconds<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">120<\/span><br \/>\n      <span class=\"token key atrule\">containers<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> server<br \/>\n          <span class=\"token key atrule\">image<\/span><span class=\"token punctuation\">:<\/span> lmsysorg\/sglang<span class=\"token punctuation\">:<\/span>&lt;version<span class=\"token punctuation\">&gt;<\/span><br \/>\n          <span class=\"token key atrule\">command<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;python3&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;-m&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;sglang.launch_server&#034;<\/span><span class=\"token punctuation\">]<\/span><br \/>\n          <span class=\"token key atrule\">args<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>model<span class=\"token punctuation\">&#8211;<\/span>path&#061;\/models\/Qwen2.5<span class=\"token punctuation\">&#8211;<\/span>32B<span class=\"token punctuation\">&#8211;<\/span>Instruct<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>served<span class=\"token punctuation\">&#8211;<\/span>model<span class=\"token punctuation\">&#8211;<\/span>name&#061;qwen2.5<span class=\"token punctuation\">&#8211;<\/span>32b<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>host&#061;0.0.0.0<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>port&#061;30000<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>tp<span class=\"token punctuation\">&#8211;<\/span>size&#061;4<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>context<span class=\"token punctuation\">&#8211;<\/span>length&#061;32768<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>mem<span class=\"token punctuation\">&#8211;<\/span>fraction<span class=\"token punctuation\">&#8211;<\/span>static&#061;0.85<br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token punctuation\">&#8211;<\/span><span class=\"token punctuation\">&#8211;<\/span>enable<span class=\"token punctuation\">&#8211;<\/span>metrics<br \/>\n          <span class=\"token key atrule\">ports<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> http<br \/>\n              <span class=\"token key atrule\">containerPort<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">30000<\/span><br \/>\n          <span class=\"token key atrule\">resources<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token key atrule\">limits<\/span><span class=\"token punctuation\">:<\/span><br \/>\n              <span class=\"token key atrule\">nvidia.com\/gpu<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;4&#034;<\/span><br \/>\n              <span class=\"token key atrule\">memory<\/span><span class=\"token punctuation\">:<\/span> 64Gi<br \/>\n          <span class=\"token key atrule\">readinessProbe<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token key atrule\">httpGet<\/span><span class=\"token punctuation\">:<\/span><br \/>\n              <span class=\"token key atrule\">path<\/span><span class=\"token punctuation\">:<\/span> \/health<br \/>\n              <span class=\"token key atrule\">port<\/span><span class=\"token punctuation\">:<\/span> http<br \/>\n            <span class=\"token key atrule\">periodSeconds<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">10<\/span><br \/>\n            <span class=\"token key atrule\">failureThreshold<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">30<\/span><br \/>\n          <span class=\"token key atrule\">livenessProbe<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token key atrule\">httpGet<\/span><span class=\"token punctuation\">:<\/span><br \/>\n              <span class=\"token key atrule\">path<\/span><span class=\"token punctuation\">:<\/span> \/health<br \/>\n              <span class=\"token key atrule\">port<\/span><span class=\"token punctuation\">:<\/span> http<br \/>\n            <span class=\"token key atrule\">periodSeconds<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">20<\/span><br \/>\n            <span class=\"token key atrule\">failureThreshold<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">6<\/span><br \/>\n          <span class=\"token key atrule\">volumeMounts<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> model<br \/>\n              <span class=\"token key atrule\">mountPath<\/span><span class=\"token punctuation\">:<\/span> \/models<br \/>\n              <span class=\"token key atrule\">readOnly<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token boolean important\">true<\/span><br \/>\n      <span class=\"token key atrule\">volumes<\/span><span class=\"token punctuation\">:<\/span><br \/>\n        <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> model<br \/>\n          <span class=\"token key atrule\">persistentVolumeClaim<\/span><span class=\"token punctuation\">:<\/span><br \/>\n            <span class=\"token key atrule\">claimName<\/span><span class=\"token punctuation\">:<\/span> llm<span class=\"token punctuation\">&#8211;<\/span>models<br \/>\n<span class=\"token punctuation\">&#8212;<\/span><br \/>\n<span class=\"token key atrule\">apiVersion<\/span><span class=\"token punctuation\">:<\/span> v1<br \/>\n<span class=\"token key atrule\">kind<\/span><span class=\"token punctuation\">:<\/span> Service<br \/>\n<span class=\"token key atrule\">metadata<\/span><span class=\"token punctuation\">:<\/span><br \/>\n  <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n<span class=\"token key atrule\">spec<\/span><span class=\"token punctuation\">:<\/span><br \/>\n  <span class=\"token key atrule\">selector<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token key atrule\">app<\/span><span class=\"token punctuation\">:<\/span> sglang<span class=\"token punctuation\">&#8211;<\/span>qwen<br \/>\n  <span class=\"token key atrule\">ports<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    <span class=\"token punctuation\">&#8211;<\/span> <span class=\"token key atrule\">name<\/span><span class=\"token punctuation\">:<\/span> http<br \/>\n      <span class=\"token key atrule\">port<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token number\">30000<\/span><br \/>\n      <span class=\"token key atrule\">targetPort<\/span><span class=\"token punctuation\">:<\/span> http<\/p>\n<p>\u6a21\u578b\u52a0\u8f7d\u901a\u5e38\u9700\u8981\u51e0\u5341\u79d2\u5230\u6570\u5206\u949f&#xff0c;readiness \u5fc5\u987b\u5728\u670d\u52a1\u771f\u6b63\u53ef\u4ee5\u63a5\u6536\u8bf7\u6c42\u540e\u624d\u901a\u8fc7\u3002\u751f\u4ea7\u53d1\u5e03\u5efa\u8bae\u4f7f\u7528\u65b0\u65e7 Deployment \u5e76\u884c\u9884\u70ed&#xff0c;\u518d\u7531\u7f51\u5173\u5207\u6d41&#xff1b;\u76f4\u63a5\u6eda\u52a8\u91cd\u542f\u552f\u4e00\u526f\u672c\u4f1a\u9020\u6210\u8f83\u957f\u4e0d\u53ef\u7528\u7a97\u53e3\u3002\u6a21\u578b\u6587\u4ef6\u5982\u679c\u901a\u8fc7 Init Container \u4e0b\u8f7d&#xff0c;\u8981\u8bbe\u7f6e\u4e0b\u8f7d\u8d85\u65f6\u3001\u6821\u9a8c\u548c\u4ee5\u53ca\u65ad\u70b9\u7eed\u4f20&#xff0c;\u907f\u514d\u6bcf\u6b21\u91cd\u542f\u90fd\u4ece\u516c\u7f51\u91cd\u65b0\u62c9\u53d6\u3002<\/p>\n<h4>6. \u4ece\u7f51\u5173\u8bbf\u95ee\u63a8\u7406\u670d\u52a1<\/h4>\n<p>\u4e09\u79cd\u63a8\u7406\u670d\u52a1\u90fd\u9002\u5408\u653e\u5728\u5185\u7f51&#xff0c;\u5916\u90e8\u9274\u6743\u3001\u79df\u6237\u9650\u6d41\u3001\u914d\u989d\u548c\u5ba1\u8ba1\u4ea4\u7ed9 AI Gateway\u3002\u6700\u5c11\u5e94\u5728\u7f51\u5173\u5c42\u8865\u9f50\u4ee5\u4e0b\u5b57\u6bb5&#xff1a;request_id\u3001trace_id\u3001tenant\u3001model\u3001stream \u548c max_tokens\u3002\u8fd9\u6837\u624d\u80fd\u628a\u4e00\u6b21\u8c03\u7528\u548c\u6307\u6807\u3001\u65e5\u5fd7\u3001Trace \u5173\u8054\u8d77\u6765\u3002<\/p>\n<h4>7. OpenAI \u517c\u5bb9 API \u8c03\u7528<\/h4>\n<p>\u542f\u52a8\u670d\u52a1\u540e&#xff0c;\u53ef\u4ee5\u4f7f\u7528 curl \u9a8c\u8bc1\u975e\u6d41\u5f0f\u63a5\u53e3&#xff1a;<\/p>\n<p><span class=\"token function\">curl<\/span> http:\/\/127.0.0.1:30000\/v1\/chat\/completions <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-H<\/span> <span class=\"token string\">&#039;Content-Type: application\/json&#039;<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-d<\/span> <span class=\"token string\">&#039;{<br \/>\n    &#034;model&#034;: &#034;qwen2.5-32b&#034;,<br \/>\n    &#034;messages&#034;: [<br \/>\n      {&#034;role&#034;: &#034;system&#034;, &#034;content&#034;: &#034;\u4f60\u662f\u4e00\u4e2a\u4e25\u8c28\u7684\u8fd0\u7ef4\u52a9\u624b\u3002&#034;},<br \/>\n      {&#034;role&#034;: &#034;user&#034;, &#034;content&#034;: &#034;\u89e3\u91ca\u4e00\u6b21 API \u5ef6\u8fdf\u5347\u9ad8\u7684\u6392\u67e5\u987a\u5e8f\u3002&#034;}<br \/>\n    ],<br \/>\n    &#034;temperature&#034;: 0.2,<br \/>\n    &#034;top_p&#034;: 0.9,<br \/>\n    &#034;max_tokens&#034;: 512,<br \/>\n    &#034;stream&#034;: false<br \/>\n  }&#039;<\/span><\/p>\n<p>\u6d41\u5f0f\u54cd\u5e94\u901a\u8fc7 SSE \u8fd4\u56de&#xff0c;\u9002\u5408\u804a\u5929\u754c\u9762\u548c\u957f\u6587\u672c\u751f\u6210&#xff1a;<\/p>\n<p><span class=\"token function\">curl<\/span> <span class=\"token parameter variable\">-N<\/span> http:\/\/127.0.0.1:30000\/v1\/chat\/completions <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-H<\/span> <span class=\"token string\">&#039;Content-Type: application\/json&#039;<\/span> <span class=\"token punctuation\">\\\\<\/span><br \/>\n  <span class=\"token parameter variable\">-d<\/span> <span class=\"token string\">&#039;{<br \/>\n    &#034;model&#034;: &#034;qwen2.5-32b&#034;,<br \/>\n    &#034;messages&#034;: [{&#034;role&#034;: &#034;user&#034;, &#034;content&#034;: &#034;\u7528\u4e09\u70b9\u8bf4\u660e SGLang \u7684\u524d\u7f00\u7f13\u5b58\u3002&#034;}],<br \/>\n    &#034;max_tokens&#034;: 256,<br \/>\n    &#034;stream&#034;: true<br \/>\n  }&#039;<\/span><\/p>\n<p>\u4e1a\u52a1\u670d\u52a1\u4e5f\u53ef\u4ee5\u590d\u7528 OpenAI SDK&#xff0c;\u53ea\u9700\u628a base_url \u6307\u5411\u5bf9\u5e94\u63a8\u7406\u670d\u52a1&#xff1a;<\/p>\n<p><span class=\"token keyword\">from<\/span> openai <span class=\"token keyword\">import<\/span> OpenAI<\/p>\n<p>client <span class=\"token operator\">&#061;<\/span> OpenAI<span class=\"token punctuation\">(<\/span><br \/>\n    base_url<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;http:\/\/inference-qwen:30000\/v1&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    api_key<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;internal-token&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><\/p>\n<p>response <span class=\"token operator\">&#061;<\/span> client<span class=\"token punctuation\">.<\/span>chat<span class=\"token punctuation\">.<\/span>completions<span class=\"token punctuation\">.<\/span>create<span class=\"token punctuation\">(<\/span><br \/>\n    model<span class=\"token operator\">&#061;<\/span><span class=\"token string\">&#034;qwen2.5-32b&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    messages<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token punctuation\">{<\/span><span class=\"token string\">&#034;role&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;user&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;content&#034;<\/span><span class=\"token punctuation\">:<\/span> <span class=\"token string\">&#034;\u5199\u4e00\u4e2a PostgreSQL \u7d22\u5f15\u68c0\u67e5\u6e05\u5355\u3002&#034;<\/span><span class=\"token punctuation\">}<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    temperature<span class=\"token operator\">&#061;<\/span><span class=\"token number\">0.1<\/span><span class=\"token punctuation\">,<\/span><br \/>\n    max_tokens<span class=\"token operator\">&#061;<\/span><span class=\"token number\">512<\/span><span class=\"token punctuation\">,<\/span><br \/>\n<span class=\"token punctuation\">)<\/span><br \/>\n<span class=\"token keyword\">print<\/span><span class=\"token punctuation\">(<\/span>response<span class=\"token punctuation\">.<\/span>choices<span class=\"token punctuation\">[<\/span><span class=\"token number\">0<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">.<\/span>message<span class=\"token punctuation\">.<\/span>content<span class=\"token punctuation\">)<\/span><\/p>\n<p>\u4e09\u79cd\u5f15\u64ce\u90fd\u63d0\u4f9b\u4e0d\u540c\u7a0b\u5ea6\u7684 OpenAI \u517c\u5bb9\u80fd\u529b&#xff0c;\u4f46\u4e0d\u7b49\u4e8e\u6240\u6709 OpenAI \u53c2\u6570\u5728\u6240\u6709\u6a21\u578b\u548c\u7248\u672c\u4e0a\u90fd\u5b8c\u5168\u4e00\u81f4\u3002\u63a5\u5165\u524d\u5e94\u5efa\u7acb\u53c2\u6570\u517c\u5bb9\u6027\u6d4b\u8bd5&#xff0c;\u91cd\u70b9\u9a8c\u8bc1 stop\u3001\u5de5\u5177\u8c03\u7528\u3001JSON \u8f93\u51fa\u3001\u89c6\u89c9\u8f93\u5165\u3001\u6d41\u5f0f\u4e2d\u65ad\u548c\u9519\u8bef\u7801\u3002<\/p>\n<h4>8. \u91c7\u6837\u53c2\u6570\u5982\u4f55\u5f71\u54cd\u7ed3\u679c<\/h4>\n<table>\n<tr>\u53c2\u6570\u542b\u4e49\u5de5\u7a0b\u5efa\u8bae<\/tr>\n<tbody>\n<tr>\n<td>temperature<\/td>\n<td>\u8c03\u6574\u91c7\u6837\u968f\u673a\u6027<\/td>\n<td>\u5206\u7c7b\u3001\u62bd\u53d6\u7b49\u4efb\u52a1\u4f7f\u7528\u4f4e\u503c&#xff1b;\u521b\u4f5c\u4efb\u52a1\u518d\u63d0\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>top_p<\/td>\n<td>nucleus sampling \u622a\u65ad\u8303\u56f4<\/td>\n<td>\u901a\u5e38\u53ea\u8c03\u4e00\u4e2a\u968f\u673a\u6027\u53c2\u6570&#xff0c;\u907f\u514d\u7ec4\u5408\u5931\u63a7<\/td>\n<\/tr>\n<tr>\n<td>max_tokens<\/td>\n<td>\u6700\u5927\u8f93\u51fa Token \u6570<\/td>\n<td>\u7531\u4e1a\u52a1\u4e0a\u9650\u63a7\u5236&#xff0c;\u9632\u6b62\u957f\u8f93\u51fa\u5360\u6ee1\u5e76\u53d1<\/td>\n<\/tr>\n<tr>\n<td>stop<\/td>\n<td>\u63d0\u524d\u7ed3\u675f\u751f\u6210\u7684\u5b57\u7b26\u4e32<\/td>\n<td>\u4e0e\u6a21\u677f\u548c\u6d41\u5f0f\u534f\u8bae\u4e00\u8d77\u6d4b\u8bd5<\/td>\n<\/tr>\n<tr>\n<td>stream<\/td>\n<td>\u662f\u5426\u5206\u7247\u8fd4\u56de<\/td>\n<td>\u957f\u54cd\u5e94\u4f7f\u7528\u6d41\u5f0f&#xff0c;\u4f46\u7f51\u5173\u8981\u6b63\u786e\u5904\u7406\u65ad\u8fde<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5bf9\u7ed3\u6784\u5316\u8f93\u51fa&#xff0c;\u4f18\u5148\u4f7f\u7528\u6a21\u578b\u548c\u6846\u67b6\u652f\u6301\u7684 JSON Schema\u3001\u6b63\u5219\u6216 choices \u7ea6\u675f&#xff0c;\u5e76\u5728\u670d\u52a1\u7aef\u505a\u6700\u7ec8 JSON \u6821\u9a8c\u3002\u4e0d\u8981\u53ea\u5728 Prompt \u4e2d\u5199\u201c\u5fc5\u987b\u8f93\u51fa JSON\u201d&#xff0c;\u7136\u540e\u628a\u672a\u7ecf\u6821\u9a8c\u7684\u6587\u672c\u76f4\u63a5\u5199\u5165\u6570\u636e\u5e93\u3002<\/p>\n<h4>9. \u4f7f\u7528\u5f15\u64ce\u539f\u751f\u524d\u7aef<\/h4>\n<p>\u5f53\u6d41\u7a0b\u5305\u542b\u591a\u4e2a\u751f\u6210\u6b65\u9aa4\u3001\u5019\u9009\u9009\u62e9\u6216\u5171\u4eab\u524d\u7f00\u65f6&#xff0c;\u5f15\u64ce\u539f\u751f\u524d\u7aef\u53ef\u4ee5\u628a\u7f16\u6392\u903b\u8f91\u653e\u5230\u670d\u52a1\u7aef\u3002\u4e0b\u9762\u4ee5 SGLang \u4e3a\u4f8b&#xff0c;vLLM \u548c TensorRT-LLM \u53ef\u4f7f\u7528\u5404\u81ea\u7684 Python API \u6216 Runtime&#xff1a;<\/p>\n<p><span class=\"token keyword\">import<\/span> sglang <span class=\"token keyword\">as<\/span> sgl<\/p>\n<p><span class=\"token decorator annotation punctuation\">&#064;sgl<span class=\"token punctuation\">.<\/span>function<\/span><br \/>\n<span class=\"token keyword\">def<\/span> <span class=\"token function\">classify_and_explain<\/span><span class=\"token punctuation\">(<\/span>s<span class=\"token punctuation\">,<\/span> question<span class=\"token punctuation\">)<\/span><span class=\"token punctuation\">:<\/span><br \/>\n    s <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token string\">&#034;\u95ee\u9898&#xff1a;&#034;<\/span> <span class=\"token operator\">&#043;<\/span> question <span class=\"token operator\">&#043;<\/span> <span class=\"token string\">&#034;\\\\n&#034;<\/span><br \/>\n    s <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token string\">&#034;\u7c7b\u522b&#xff1a;&#034;<\/span> <span class=\"token operator\">&#043;<\/span> sgl<span class=\"token punctuation\">.<\/span>gen<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;category&#034;<\/span><span class=\"token punctuation\">,<\/span> choices<span class=\"token operator\">&#061;<\/span><span class=\"token punctuation\">[<\/span><span class=\"token string\">&#034;\u6545\u969c&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;\u914d\u7f6e&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;\u6027\u80fd&#034;<\/span><span class=\"token punctuation\">,<\/span> <span class=\"token string\">&#034;\u5176\u4ed6&#034;<\/span><span class=\"token punctuation\">]<\/span><span class=\"token punctuation\">)<\/span> <span class=\"token operator\">&#043;<\/span> <span class=\"token string\">&#034;\\\\n&#034;<\/span><br \/>\n    s <span class=\"token operator\">&#043;&#061;<\/span> <span class=\"token string\">&#034;\u89e3\u91ca&#xff1a;&#034;<\/span> <span class=\"token operator\">&#043;<\/span> sgl<span class=\"token punctuation\">.<\/span>gen<span class=\"token punctuation\">(<\/span><span class=\"token string\">&#034;explanation&#034;<\/span><span class=\"token punctuation\">,<\/span> max_tokens<span class=\"token operator\">&#061;<\/span><span class=\"token number\">256<\/span><span class=\"token punctuation\">)<\/span><\/p>\n<p>\u8fd9\u7c7b\u7f16\u6392\u9002\u5408\u56fa\u5b9a\u6d41\u7a0b\u548c\u53ef\u590d\u7528\u524d\u7f00\u3002\u52a8\u6001 Agent \u4ecd\u5e94\u8bbe\u7f6e\u6b65\u9aa4\u6570\u3001\u603b Token\u3001\u5de5\u5177\u8c03\u7528\u6b21\u6570\u548c\u8d85\u65f6&#xff0c;\u907f\u514d\u6a21\u578b\u5728\u670d\u52a1\u7aef\u65e0\u9650\u5faa\u73af\u3002<\/p>\n<h3>\u4e09\u3001\u6027\u80fd\u8c03\u4f18&#xff1a;\u4ece\u663e\u5b58\u5230\u541e\u5410&#xff08;\u6307\u6807\u3001\u91cf\u5316\u4e0e PD \u5206\u79bb&#xff09;<\/h3>\n<h4>1. \u5148\u5efa\u7acb\u6b63\u786e\u7684\u6027\u80fd\u6307\u6807<\/h4>\n<p>\u53ea\u770b\u5e73\u5747\u54cd\u5e94\u65f6\u95f4\u4f1a\u63a9\u76d6\u6392\u961f\u548c\u957f\u5c3e\u95ee\u9898\u3002\u5728\u7ebf\u63a8\u7406\u81f3\u5c11\u8bb0\u5f55&#xff1a;<\/p>\n<table>\n<tr>\u6307\u6807\u5b9a\u4e49\u4e3b\u8981\u53d7\u4ec0\u4e48\u5f71\u54cd<\/tr>\n<tbody>\n<tr>\n<td>TTFT<\/td>\n<td>\u8bf7\u6c42\u5230\u9996\u4e2a\u8f93\u51fa Token \u7684\u65f6\u95f4<\/td>\n<td>\u6392\u961f\u3001Prefill\u3001\u8f93\u5165\u957f\u5ea6\u3001\u524d\u7f00\u7f13\u5b58<\/td>\n<\/tr>\n<tr>\n<td>TPOT<\/td>\n<td>\u76f8\u90bb\u8f93\u51fa Token \u7684\u5e73\u5747\u95f4\u9694<\/td>\n<td>Decode\u3001GPU \u5229\u7528\u7387\u3001\u5e76\u53d1\u548c\u91c7\u6837<\/td>\n<\/tr>\n<tr>\n<td>E2E Latency<\/td>\n<td>\u8bf7\u6c42\u5f00\u59cb\u5230\u5b8c\u6210\u7684\u603b\u65f6\u95f4<\/td>\n<td>TTFT\u3001\u8f93\u51fa\u957f\u5ea6\u3001\u7f51\u7edc\u548c\u5ba2\u6237\u7aef\u8bfb\u53d6<\/td>\n<\/tr>\n<tr>\n<td>Throughput<\/td>\n<td>\u6bcf\u79d2\u751f\u6210 Token \u6216\u5b8c\u6210\u8bf7\u6c42\u6570<\/td>\n<td>\u6279\u5904\u7406\u3001\u6a21\u578b\u5927\u5c0f\u3001\u5e76\u53d1\u548c GPU<\/td>\n<\/tr>\n<tr>\n<td>Queue Time<\/td>\n<td>\u8fdb\u5165\u961f\u5217\u5230\u5f00\u59cb\u6267\u884c\u7684\u65f6\u95f4<\/td>\n<td>\u5e76\u53d1\u4e0a\u9650\u3001\u8c03\u5ea6\u7b56\u7565\u548c\u7a81\u53d1\u6d41\u91cf<\/td>\n<\/tr>\n<tr>\n<td>Cache Hit Rate<\/td>\n<td>\u524d\u7f00\u7f13\u5b58\u547d\u4e2d\u6bd4\u4f8b<\/td>\n<td>Prompt \u7a33\u5b9a\u6027\u548c\u7f13\u5b58\u5bb9\u91cf<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u538b\u6d4b\u65f6\u56fa\u5b9a\u6a21\u578b revision\u3001\u8f93\u5165\u8f93\u51fa\u957f\u5ea6\u548c\u5e76\u53d1\u66f2\u7ebf&#xff0c;\u5206\u522b\u6d4b\u51b7\u542f\u52a8\u3001\u7f13\u5b58\u547d\u4e2d\u3001\u5355\u8bf7\u6c42\u3001\u7a33\u5b9a\u5e76\u53d1\u548c\u7a81\u53d1\u6d41\u91cf\u3002\u628a P50\u3001P95\u3001P99 \u4e0e GPU \u663e\u5b58\u3001\u529f\u8017\u3001\u6e29\u5ea6\u3001PCIe\/NVLink \u901a\u4fe1\u4e00\u8d77\u8bb0\u5f55&#xff0c;\u624d\u80fd\u5224\u65ad\u662f\u6a21\u578b\u8ba1\u7b97\u6162\u8fd8\u662f\u6392\u961f\u6162\u3002<\/p>\n<h4>2. \u663e\u5b58\u9884\u7b97\u4e0e\u4e0a\u4e0b\u6587\u957f\u5ea6<\/h4>\n<p>\u663e\u5b58\u5927\u81f4\u7531\u56db\u90e8\u5206\u6784\u6210&#xff1a;\u6a21\u578b\u6743\u91cd\u3001\u8fd0\u884c\u65f6\u4e34\u65f6\u5f20\u91cf\u3001KV Cache \u548c\u901a\u4fe1\/\u6846\u67b6\u5f00\u9500\u3002\u91cf\u5316\u53ea\u4e3b\u8981\u964d\u4f4e\u6743\u91cd\u5360\u7528&#xff0c;\u4e0d\u4f1a\u6309\u76f8\u540c\u6bd4\u4f8b\u6d88\u9664 KV Cache&#xff1b;\u5f53\u4e0a\u4e0b\u6587\u5f88\u957f\u3001\u5e76\u53d1\u5f88\u9ad8\u65f6&#xff0c;KV Cache \u4ecd\u7136\u53ef\u80fd\u6210\u4e3a\u74f6\u9888\u3002<\/p>\n<p>\u5e38\u7528\u8c03\u4f18\u987a\u5e8f\u5982\u4e0b&#xff1a;<\/p>\n<li>\u5148\u786e\u8ba4\u6a21\u578b\u6743\u91cd\u548c Tokenizer \u80fd\u5728\u76ee\u6807 GPU \u4e0a\u7a33\u5b9a\u52a0\u8f7d\u3002<\/li>\n<li>\u5c06\u4e0a\u4e0b\u6587\u4e0a\u9650\u8bbe\u4e3a\u4e1a\u52a1\u771f\u5b9e\u4e0a\u9650&#xff1a;SGLang \u4f7f\u7528 &#8211;context-length&#xff0c;vLLM \u4f7f\u7528 &#8211;max-model-len&#xff0c;TensorRT-LLM \u901a\u5e38\u5728 Engine\/profile \u914d\u7f6e\u4e2d\u56fa\u5b9a&#xff0c;\u907f\u514d\u4e3a\u6781\u5c11\u6570\u8bf7\u6c42\u9884\u7559\u8fc7\u5927\u7684 KV Cache\u3002<\/li>\n<li>\u4e3a\u8fd0\u884c\u65f6\u548c\u901a\u4fe1\u9884\u7559\u663e\u5b58\u4f59\u91cf&#xff1a;SGLang \u53ef\u8c03\u6574 &#8211;mem-fraction-static&#xff0c;vLLM \u53ef\u8c03\u6574 &#8211;gpu-memory-utilization&#xff0c;TensorRT-LLM \u5219\u9700\u8981\u5728 Engine \u6784\u5efa\u548c\u8fd0\u884c\u65f6\u914d\u7f6e workspace\u3001KV Cache \u9884\u7b97&#xff1b;\u786e\u8ba4\u7a33\u5b9a\u540e\u518d\u9010\u6b65\u63d0\u9ad8\u5229\u7528\u7387\u3002<\/li>\n<li>\u8bbe\u7f6e\u5f15\u64ce\u5bf9\u5e94\u7684\u5728\u7ebf\u5e76\u53d1\u4e0a\u9650&#xff1a;SGLang \u5e38\u7528 &#8211;max-running-requests&#xff0c;vLLM \u53ef\u4f7f\u7528 &#8211;max-num-seqs \u7b49\u8c03\u5ea6\u53c2\u6570&#xff0c;TensorRT-LLM \u901a\u8fc7 batching\/scheduler \u914d\u7f6e\u63a7\u5236&#xff1b;\u4f7f\u7528\u7f51\u5173\u6392\u961f\u800c\u4e0d\u662f\u8ba9\u8fdb\u7a0b\u65e0\u754c\u5806\u79ef\u3002<\/li>\n<li>\u6839\u636e\u6a21\u578b\u652f\u6301\u60c5\u51b5\u9009\u62e9 FP8\u3001AWQ\u3001GPTQ \u7b49\u91cf\u5316\u65b9\u6848&#xff0c;\u5e76\u91cd\u65b0\u8bc4\u4f30\u51c6\u786e\u7387\u3001\u9996 Token \u5ef6\u8fdf\u548c\u541e\u5410\u3002<\/li>\n<p>\u53d1\u751f OOM \u65f6&#xff0c;\u4e0d\u8981\u53ea\u91cd\u542f Pod\u3002\u5148\u786e\u8ba4\u662f\u542f\u52a8\u9636\u6bb5\u6743\u91cd OOM\u3001\u8fd0\u884c\u9636\u6bb5 KV Cache OOM&#xff0c;\u8fd8\u662f\u957f\u5c3e\u8bf7\u6c42\u89e6\u53d1\u7684\u788e\u7247\u95ee\u9898&#xff0c;\u518d\u5206\u522b\u8c03\u6574 TP\u3001\u4e0a\u4e0b\u6587\u3001\u5e76\u53d1\u548c\u91cf\u5316\u914d\u7f6e\u3002<\/p>\n<h4>3. Tensor Parallel\u3001Data Parallel \u4e0e\u526f\u672c<\/h4>\n<p>Tensor Parallel&#xff08;TP&#xff09;\u628a\u540c\u4e00\u4e2a\u6a21\u578b\u62c6\u5230\u591a\u5f20 GPU \u4e0a&#xff0c;\u9002\u5408\u5355\u5361\u653e\u4e0d\u4e0b\u6216\u9700\u8981\u66f4\u9ad8\u5355\u5b9e\u4f8b\u541e\u5410\u7684\u6a21\u578b&#xff1b;\u4ee3\u4ef7\u662f\u6bcf\u4e00\u6b65\u90fd\u53ef\u80fd\u53d1\u751f\u8de8\u5361\u901a\u4fe1\u3002Data Parallel&#xff08;DP&#xff09;\u662f\u591a\u4e2a\u63a8\u7406\u526f\u672c\u5171\u540c\u627f\u63a5\u8bf7\u6c42&#xff0c;\u9002\u5408\u6a21\u578b\u80fd\u591f\u653e\u5165\u6bcf\u7ec4 GPU\u3001\u5e76\u4e14\u5e0c\u671b\u63d0\u5347\u6574\u4f53\u5e76\u53d1\u7684\u573a\u666f\u3002<\/p>\n<p>\u53ef\u4ee5\u6309\u4e0b\u9762\u7684\u65b9\u5f0f\u505a\u5bb9\u91cf\u8bbe\u8ba1&#xff1a;<\/p>\n<p>\u5355 Pod \u4f7f\u7528 GPU \u6570 &#061; TP \u00d7 DP<br \/>\n\u96c6\u7fa4\u603b\u5bb9\u91cf &#061; \u5355 Pod \u6bcf\u79d2\u751f\u6210 Token \u00d7 Pod \u6570\u91cf<br \/>\n\u7a33\u5b9a\u5e76\u53d1\u4e0a\u9650 \u2248 \u53ef\u63a5\u53d7\u6392\u961f\u65f6\u95f4\u5185\u7684\u5904\u7406\u80fd\u529b<\/p>\n<p>\u5728\u6709 NVLink \u7684\u673a\u5668\u4e0a&#xff0c;\u4f18\u5148\u8ba9 TP \u8de8\u540c\u4e00\u4e92\u8054\u57df&#xff1b;\u8de8 NUMA\u3001\u8de8\u4e3b\u673a\u7684 TP \u9700\u8981\u9ad8\u901f\u7f51\u7edc\u548c NCCL \u914d\u7f6e&#xff0c;\u4e0d\u80fd\u7167\u642c\u5355\u673a\u53c2\u6570\u3002\u82e5\u6a21\u578b\u80fd\u5728\u5355\u5361\u6216\u5c0f TP \u4e0a\u8fd0\u884c&#xff0c;\u901a\u5e38\u589e\u52a0\u591a\u4e2a\u526f\u672c\u6bd4\u7ee7\u7eed\u6269\u5927 TP \u66f4\u5bb9\u6613\u83b7\u5f97\u7ebf\u6027\u541e\u5410\u3002<\/p>\n<h4>4. \u8c03\u5ea6\u548c\u524d\u7f00\u7f13\u5b58\u8c03\u4f18<\/h4>\n<p>\u8fde\u7eed\u6279\u5904\u7406\u4f1a\u5728\u6bcf\u8f6e Decode \u4e2d\u52a8\u6001\u52a0\u5165\u65b0\u8bf7\u6c42\u3001\u79fb\u9664\u5df2\u5b8c\u6210\u8bf7\u6c42&#xff0c;\u63d0\u9ad8 GPU \u7a7a\u95f2\u65f6\u95f4\u7684\u5229\u7528\u7387\u3002\u4f46\u5e76\u53d1\u4e0d\u662f\u8d8a\u5927\u8d8a\u597d&#xff1a;\u8fc7\u5927\u7684\u6279\u6b21\u4f1a\u589e\u52a0 TTFT \u548c\u957f\u5c3e&#xff0c;\u8fc7\u5c0f\u7684\u6279\u6b21\u53c8\u65e0\u6cd5\u53d1\u6325 GPU \u541e\u5410\u3002<\/p>\n<p>\u5efa\u8bae\u5efa\u7acb\u4e24\u7c7b\u670d\u52a1\u6c60&#xff1a;\u4ea4\u4e92\u6c60\u4f18\u5148\u4fdd\u8bc1 TTFT \u548c P99&#xff0c;\u79bb\u7ebf\u6c60\u4f18\u5148\u8ffd\u6c42\u6bcf\u79d2 Token&#xff1b;\u4e0d\u8981\u8ba9\u6279\u91cf\u8bc4\u6d4b\u6216\u957f\u6587\u672c\u603b\u7ed3\u5360\u6ee1\u4ea4\u4e92\u6c60\u3002\u5bf9\u5171\u4eab\u7cfb\u7edf\u63d0\u793a\u8bcd\u548c\u5de5\u5177\u5b9a\u4e49&#xff0c;\u56fa\u5b9a\u6a21\u677f\u548c\u5b57\u6bb5\u987a\u5e8f&#xff0c;\u89c2\u5bdf\u7f13\u5b58\u547d\u4e2d\u7387&#xff1b;\u547d\u4e2d\u7387\u4f4e\u65f6\u5148\u68c0\u67e5 Prompt \u62fc\u63a5\u903b\u8f91&#xff0c;\u518d\u8c03\u6574\u7f13\u5b58\u9884\u7b97\u3002<\/p>\n<h4>5. PD \u5206\u79bb&#xff1a;\u628a Prefill \u4e0e Decode \u62c6\u6210\u72ec\u7acb\u670d\u52a1\u6c60<\/h4>\n<p>PD \u5206\u79bb&#xff08;Prefill\/Decode Disaggregation&#xff0c;\u9884\u586b\u5145\u4e0e\u89e3\u7801\u5206\u79bb&#xff09;\u786e\u5b9e\u5c5e\u4e8e\u6027\u80fd\u8c03\u4f18\u5b9e\u8df5&#xff0c;\u4f46\u5b83\u66f4\u63a5\u8fd1\u201c\u90e8\u7f72\u67b6\u6784\u7ea7\u8c03\u4f18\u201d&#xff0c;\u800c\u4e0d\u662f\u5355\u4e2a\u542f\u52a8\u53c2\u6570\u3002Prefill \u9700\u8981\u4e00\u6b21\u6027\u5904\u7406\u5927\u91cf\u8f93\u5165 Token&#xff0c;\u901a\u5e38\u66f4\u504f\u8ba1\u7b97\u5bc6\u96c6&#xff1b;Decode \u6bcf\u8f6e\u53ea\u751f\u6210\u5c11\u91cf Token&#xff0c;\u66f4\u5bb9\u6613\u53d7\u663e\u5b58\u5e26\u5bbd\u3001KV Cache \u8bbf\u95ee\u548c\u8c03\u5ea6\u6296\u52a8\u5f71\u54cd\u3002\u4e24\u8005\u6df7\u5728\u540c\u4e00\u4e2a\u5b9e\u4f8b\u65f6&#xff0c;\u957f\u4e0a\u4e0b\u6587 Prefill \u53ef\u80fd\u6324\u5360 Decode \u7684\u6267\u884c\u673a\u4f1a&#xff0c;\u5bfc\u81f4 TTFT\u3001TPOT \u4e92\u76f8\u5f71\u54cd\u3002<\/p>\n<p>\u5178\u578b\u7684 PD \u67b6\u6784\u5982\u4e0b&#xff1a;<\/p>\n<p>\u5ba2\u6237\u7aef<br \/>\n  \u2502<br \/>\n  \u25bc<br \/>\nAI Gateway \/ PD Router<br \/>\n  \u251c\u2500 Prefill Pool&#xff1a;\u5904\u7406\u8f93\u5165 Token&#xff0c;\u751f\u6210 KV Cache<br \/>\n  \u2502                  \u2502 KV Transfer&#xff08;RDMA \/ NVLink \/ \u9ad8\u901f\u7f51\u7edc&#xff09;<br \/>\n  \u2502                  \u25bc<br \/>\n  \u2514\u2500 Decode Pool&#xff1a;\u52a0\u8f7d\u6216\u63a5\u6536 KV Cache&#xff0c;\u6301\u7eed\u751f\u6210 Token \u5e76\u8fd4\u56de SSE<\/p>\n<p>PD \u5206\u79bb\u7684\u4ef7\u503c\u5728\u4e8e\u53ef\u4ee5\u5206\u522b\u6269\u7f29\u5bb9\u548c\u8bbe\u7f6e\u8d44\u6e90\u914d\u989d\u3002\u4f8b\u5982&#xff0c;\u8f93\u5165\u4e0a\u4e0b\u6587\u5f88\u957f\u4f46\u8f93\u51fa\u8f83\u77ed\u7684\u77e5\u8bc6\u5e93\u95ee\u7b54&#xff0c;\u53ef\u4ee5\u589e\u52a0 Prefill \u526f\u672c&#xff1b;\u804a\u5929\u548c\u4ee3\u7801\u8865\u5168\u8f93\u51fa\u8f83\u957f&#xff0c;\u5219\u53ef\u4ee5\u589e\u52a0 Decode \u526f\u672c\u3002\u8def\u7531\u5668\u8fd8\u53ef\u4ee5\u6309\u79df\u6237\u3001\u6a21\u578b\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u548c\u670d\u52a1\u7b49\u7ea7\u628a\u8bf7\u6c42\u9001\u5165\u4e0d\u540c\u8d44\u6e90\u6c60\u3002<\/p>\n<p>SGLang\u3001vLLM \u548c TensorRT-LLM \u5bf9 PD \u5206\u79bb\u3001KV Transfer Connector\u3001\u8fdc\u7a0b KV Cache \u548c Triton \u7f16\u6392\u7684\u652f\u6301\u65b9\u5f0f\u5e76\u4e0d\u76f8\u540c&#xff0c;\u90e8\u5206\u80fd\u529b\u4ecd\u5904\u4e8e\u5b9e\u9a8c\u6216\u5feb\u901f\u6f14\u8fdb\u9636\u6bb5\u3002\u4e0d\u8981\u53ea\u770b\u5230\u6587\u6863\u4e2d\u7684\u529f\u80fd\u540d\u5c31\u76f4\u63a5\u4e0a\u7ebf&#xff0c;\u5e94\u786e\u8ba4\u76ee\u6807\u7248\u672c\u662f\u5426\u652f\u6301\u5f53\u524d\u6a21\u578b\u3001\u5e76\u884c\u65b9\u5f0f\u548c\u7f51\u7edc\u62d3\u6251&#xff0c;\u5e76\u5b8c\u6210\u7aef\u5230\u7aef\u538b\u6d4b\u3002<\/p>\n<table>\n<tr>\u5224\u65ad\u6761\u4ef6\u66f4\u9002\u5408\u91c7\u7528 PD \u5206\u79bb\u6682\u4e0d\u5efa\u8bae\u91c7\u7528<\/tr>\n<tbody>\n<tr>\n<td>\u8d1f\u8f7d\u7279\u5f81<\/td>\n<td>\u957f\u4e0a\u4e0b\u6587\u3001Prefill \u548c Decode \u8d44\u6e90\u9700\u6c42\u5dee\u5f02\u660e\u663e<\/td>\n<td>\u5355\u8bf7\u6c42\u3001\u77ed\u4e0a\u4e0b\u6587\u3001\u5e76\u53d1\u5f88\u4f4e<\/td>\n<\/tr>\n<tr>\n<td>\u76ee\u6807<\/td>\n<td>\u540c\u65f6\u538b\u4f4e TTFT \u548c TPOT&#xff0c;\u6216\u9700\u8981\u72ec\u7acb\u6269\u7f29\u5bb9<\/td>\n<td>\u53ea\u8ffd\u6c42\u5355\u8bf7\u6c42\u6700\u4f4e\u5ef6\u8fdf<\/td>\n<\/tr>\n<tr>\n<td>\u7f51\u7edc<\/td>\n<td>\u5177\u5907\u4f4e\u5ef6\u8fdf\u3001\u9ad8\u5e26\u5bbd\u7684 NVLink\u3001RoCE \u6216 InfiniBand<\/td>\n<td>\u4ec5\u6709\u666e\u901a\u8de8\u673a\u7f51\u7edc&#xff0c;KV \u4f20\u8f93\u5360\u6bd4\u9ad8<\/td>\n<\/tr>\n<tr>\n<td>\u8fd0\u7ef4<\/td>\n<td>\u6709\u6210\u719f\u7684\u8def\u7531\u3001\u9884\u70ed\u3001\u6545\u969c\u8f6c\u79fb\u548c\u5bb9\u91cf\u6cbb\u7406<\/td>\n<td>\u56e2\u961f\u5c1a\u672a\u5efa\u7acb\u591a\u6c60\u53d1\u5e03\u548c\u6392\u969c\u80fd\u529b<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u843d\u5730\u65f6\u81f3\u5c11\u89c2\u6d4b\u56db\u7c7b\u6570\u636e&#xff1a;Prefill \u6392\u961f\u65f6\u95f4\u3001Decode \u6392\u961f\u65f6\u95f4\u3001KV \u4f20\u8f93\u5b57\u8282\u6570\u4e0e\u8017\u65f6\u3001\u7aef\u5230\u7aef TTFT\/TPOT\u3002\u53ea\u6709\u5f53\u201c\u51cf\u5c11\u4e92\u76f8\u5e72\u6270\u201d\u5e26\u6765\u7684\u6536\u76ca\u5927\u4e8e KV \u4f20\u8f93\u548c\u989d\u5916\u526f\u672c\u6210\u672c\u65f6&#xff0c;PD \u5206\u79bb\u624d\u662f\u6b63\u6536\u76ca\u3002KV \u4f20\u8f93\u5931\u8d25\u65f6\u5e94\u6709\u91cd\u8bd5\u6216\u56de\u9000\u5230\u5408\u5e76\u5b9e\u4f8b\u7684\u7b56\u7565&#xff0c;\u5426\u5219\u4e00\u4e2a\u7f51\u7edc\u6296\u52a8\u5c31\u53ef\u80fd\u653e\u5927\u4e3a\u6574\u6761\u8bf7\u6c42\u94fe\u8def\u5931\u8d25\u3002<\/p>\n<h4>6. \u4e0e PD \u5206\u79bb\u914d\u5957\u7684\u8fdb\u9636\u4f18\u5316<\/h4>\n<p>PD \u5206\u79bb\u4e4b\u5916&#xff0c;\u8fd8\u6709\u51e0\u7c7b\u7ecf\u5e38\u7ec4\u5408\u4f7f\u7528\u7684\u4f18\u5316\u3002\u5b83\u4eec\u89e3\u51b3\u7684\u95ee\u9898\u4e0d\u540c&#xff0c;\u4e0d\u80fd\u7b80\u5355\u53e0\u52a0\u540e\u671f\u5f85\u7ebf\u6027\u6536\u76ca&#xff1a;<\/p>\n<table>\n<tr>\u4f18\u5316\u5b9e\u8df5\u4e3b\u8981\u89e3\u51b3\u7684\u95ee\u9898\u9002\u7528\u573a\u666f\u4e3b\u8981\u4ee3\u4ef7\u6216\u98ce\u9669<\/tr>\n<tbody>\n<tr>\n<td>Chunked Prefill<\/td>\n<td>\u957f\u8f93\u5165\u4e00\u6b21 Prefill \u963b\u585e Decode<\/td>\n<td>\u957f\u4e0a\u4e0b\u6587\u548c\u4ea4\u4e92\u8bf7\u6c42\u6df7\u90e8<\/td>\n<td>\u5207\u5757\u8fc7\u5c0f\u4f1a\u589e\u52a0\u8c03\u5ea6\u5f00\u9500&#xff0c;TTFT \u53ef\u80fd\u53d8\u957f<\/td>\n<\/tr>\n<tr>\n<td>Speculative Decoding<\/td>\n<td>Decode \u6bcf\u6b65\u751f\u6210\u901f\u5ea6\u4e0d\u8db3<\/td>\n<td>\u8f93\u51fa\u8f83\u957f\u3001\u8349\u7a3f\u6a21\u578b\u4e0e\u76ee\u6807\u6a21\u578b\u76f8\u8fd1<\/td>\n<td>\u9700\u8981\u989d\u5916\u8349\u7a3f\u6a21\u578b&#xff0c;\u63a5\u53d7\u7387\u4f4e\u65f6\u6536\u76ca\u6709\u9650<\/td>\n<\/tr>\n<tr>\n<td>KV Cache \u5206\u5c42\/\u5378\u8f7d<\/td>\n<td>GPU \u663e\u5b58\u4e0d\u8db3\u3001\u5e76\u53d1\u53d7\u9650<\/td>\n<td>\u957f\u4f1a\u8bdd\u3001\u4f4e\u9891\u7f13\u5b58\u548c\u8d85\u957f\u4e0a\u4e0b\u6587<\/td>\n<td>CPU\/NVMe \u4f20\u8f93\u5ef6\u8fdf\u9ad8&#xff0c;\u9700\u63a7\u5236\u51b7\u70ed\u6570\u636e<\/td>\n<\/tr>\n<tr>\n<td>Prefix Cache<\/td>\n<td>\u91cd\u590d\u7cfb\u7edf\u63d0\u793a\u8bcd\u5bfc\u81f4 Prefill \u91cd\u7b97<\/td>\n<td>\u591a\u8f6e\u5bf9\u8bdd\u3001RAG\u3001\u5de5\u5177\u8c03\u7528<\/td>\n<td>Prompt \u89c4\u8303\u7a0d\u6709\u53d8\u5316\u5c31\u4f1a\u5931\u6548\u5e76\u5360\u7528\u663e\u5b58<\/td>\n<\/tr>\n<tr>\n<td>MoE Expert Parallel<\/td>\n<td>MoE \u6a21\u578b\u4e13\u5bb6\u8ba1\u7b97\u548c\u663e\u5b58\u5206\u5e03\u4e0d\u5747<\/td>\n<td>Mixtral\u3001Qwen-MoE \u7b49\u7a00\u758f\u6a21\u578b<\/td>\n<td>All-to-All \u901a\u4fe1\u548c\u4e13\u5bb6\u8d1f\u8f7d\u503e\u659c<\/td>\n<\/tr>\n<tr>\n<td>\u8bf7\u6c42\u5206\u7ea7\u4e0e\u8def\u7531<\/td>\n<td>\u957f\u77ed\u8bf7\u6c42\u4e92\u76f8\u5f71\u54cd\u3001\u7a81\u53d1\u6d41\u91cf\u6392\u961f<\/td>\n<td>\u4ea4\u4e92\u3001\u6279\u5904\u7406\u3001\u79bb\u7ebf\u4efb\u52a1\u5e76\u5b58<\/td>\n<td>\u9700\u8981\u7f51\u5173\u8bc6\u522b Token \u9884\u7b97\u5e76\u7ef4\u62a4\u591a\u4e2a\u670d\u52a1\u6c60<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>Chunked Prefill\u3002<\/p>\n<p>Chunked Prefill \u628a\u957f\u8f93\u5165\u62c6\u6210\u591a\u4e2a\u5c0f\u5757&#xff0c;\u5728\u5757\u4e0e\u5757\u4e4b\u95f4\u8ba9 Decode \u8bf7\u6c42\u83b7\u5f97\u8c03\u5ea6\u673a\u4f1a\u3002\u5b83\u53ef\u4ee5\u5728\u4e0d\u5b8c\u5168\u62c6\u5206\u670d\u52a1\u7684\u60c5\u51b5\u4e0b\u7f13\u89e3\u201c\u957f Prefill \u963b\u585e\u77ed\u8bf7\u6c42\u201d&#xff0c;\u9002\u5408\u5148\u505a\u5355\u5b9e\u4f8b\u4f18\u5316\u3002\u5757\u5927\u5c0f\u5e94\u901a\u8fc7 TTFT\u3001TPOT \u548c GPU \u5229\u7528\u7387\u538b\u6d4b\u786e\u5b9a&#xff1b;\u5e76\u4e0d\u662f\u8d8a\u5c0f\u8d8a\u597d\u3002<\/p>\n<p>Speculative Decoding\u3002<\/p>\n<p>Speculative Decoding \u5148\u7531\u8f83\u5c0f\u7684\u8349\u7a3f\u6a21\u578b\u63d0\u51fa\u591a\u4e2a\u5019\u9009 Token&#xff0c;\u518d\u7531\u76ee\u6807\u6a21\u578b\u6279\u91cf\u6821\u9a8c&#xff0c;\u6821\u9a8c\u901a\u8fc7\u7684 Token \u53ef\u4ee5\u4e00\u6b21\u63d0\u4ea4&#xff0c;\u4ece\u800c\u51cf\u5c11\u76ee\u6807\u6a21\u578b\u9010 Token Decode \u7684\u8f6e\u6570\u3002\u5b83\u901a\u5e38\u5bf9\u8f93\u51fa\u8f83\u957f\u3001\u8349\u7a3f\u6a21\u578b\u4e0e\u76ee\u6807\u6a21\u578b\u5206\u5e03\u63a5\u8fd1\u7684\u4efb\u52a1\u66f4\u6709\u6548&#xff1b;\u6458\u8981\u3001\u4ee3\u7801\u548c\u4e25\u683c\u683c\u5f0f\u8f93\u51fa\u5e94\u5206\u522b\u6d4b\u91cf\u63a5\u53d7\u7387\u548c\u8d28\u91cf&#xff0c;\u4e0d\u80fd\u53ea\u770b\u7406\u8bba\u52a0\u901f\u6bd4\u3002<\/p>\n<p>KV Cache \u5206\u5c42\u4e0e\u5378\u8f7d\u3002<\/p>\n<p>\u53ef\u4ee5\u628a KV Cache \u6309\u51b7\u70ed\u7a0b\u5ea6\u653e\u5728 GPU HBM\u3001CPU \u5185\u5b58\u751a\u81f3\u672c\u5730 NVMe \u4e2d&#xff0c;\u6216\u8005\u4f7f\u7528\u72ec\u7acb\u7684\u8fdc\u7a0b KV Cache \u670d\u52a1\u3002\u70ed\u6570\u636e\u7559\u5728 GPU&#xff0c;\u51b7\u6570\u636e\u5728\u8bf7\u6c42\u518d\u6b21\u547d\u4e2d\u65f6\u56de\u8fc1\u3002\u8be5\u65b9\u6848\u80fd\u63d0\u9ad8\u53ef\u5bb9\u7eb3\u7684\u4f1a\u8bdd\u6570&#xff0c;\u4f46\u4f1a\u5f15\u5165\u4f20\u8f93\u5ef6\u8fdf\u548c\u4e00\u81f4\u6027\u95ee\u9898&#xff1b;\u5e94\u8bbe\u7f6e TTL\u3001\u5bb9\u91cf\u4e0a\u9650\u548c\u79df\u6237\u9694\u79bb&#xff0c;\u5e76\u628a\u547d\u4e2d\u7387\u3001\u56de\u8fc1\u8017\u65f6\u548c\u6dd8\u6c70\u6b21\u6570\u7eb3\u5165\u76d1\u63a7\u3002<\/p>\n<p>MoE \u7684 Expert Parallel\u3002<\/p>\n<p>\u5bf9 Mixture-of-Experts&#xff08;MoE&#xff09;\u6a21\u578b&#xff0c;Expert Parallel&#xff08;EP&#xff09;\u53ef\u4ee5\u628a\u4e0d\u540c\u4e13\u5bb6\u5206\u5e03\u5230\u4e0d\u540c GPU&#xff0c;\u518d\u901a\u8fc7\u8def\u7531\u53ea\u6fc0\u6d3b\u90e8\u5206\u4e13\u5bb6\u3002\u5b83\u4e0e TP\u3001DP \u53ef\u4ee5\u7ec4\u5408&#xff0c;\u4f46\u4f1a\u589e\u52a0 All-to-All \u901a\u4fe1\u3002\u4f18\u5316\u91cd\u70b9\u662f\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861\u3001\u901a\u4fe1\u62d3\u6251\u548c batch \u7ec4\u6210&#xff1b;\u5f53\u5c11\u6570\u4e13\u5bb6\u957f\u671f\u8fc7\u70ed\u65f6&#xff0c;\u5355\u7eaf\u589e\u52a0 GPU \u6570\u91cf\u4e0d\u4e00\u5b9a\u6709\u6548&#xff0c;\u9700\u8981\u8c03\u6574\u8def\u7531\u3001\u5bb9\u91cf\u56e0\u5b50\u6216\u526f\u672c\u7b56\u7565\u3002<\/p>\n<p>\u8fd9\u4e9b\u4f18\u5316\u5efa\u8bae\u6309\u201c\u5148\u6d4b\u91cf\u3001\u540e\u6539\u52a8\u201d\u7684\u987a\u5e8f\u63a8\u8fdb&#xff1a;\u5148\u7528\u7edf\u4e00\u8bf7\u6c42\u96c6\u5efa\u7acb\u57fa\u7ebf&#xff0c;\u518d\u4e00\u6b21\u53ea\u6539\u53d8\u4e00\u4e2a\u53d8\u91cf&#xff0c;\u8bb0\u5f55 TTFT\u3001TPOT\u3001P99\u3001\u541e\u5410\u3001GPU \u663e\u5b58\u3001KV Cache \u547d\u4e2d\u7387\u548c\u5355\u4f4d Token \u6210\u672c\u3002\u4efb\u4f55\u4f18\u5316\u90fd\u5fc5\u987b\u6709\u56de\u6eda\u5f00\u5173&#xff0c;\u907f\u514d\u6027\u80fd\u63d0\u5347\u5374\u5e26\u6765\u51c6\u786e\u7387\u3001\u7a33\u5b9a\u6027\u6216\u8fd0\u7ef4\u590d\u6742\u5ea6\u7684\u4e0d\u53ef\u63a7\u53d8\u5316\u3002<\/p>\n<h4>7. \u91cf\u5316\u3001FlashInfer \u4e0e Kernel \u9009\u62e9<\/h4>\n<p>\u91cf\u5316\u53ef\u4ee5\u964d\u4f4e\u663e\u5b58\u548c\u5e26\u5bbd\u538b\u529b&#xff0c;\u4f46\u4e0d\u540c\u6a21\u578b\u3001\u5c42\u548c\u4efb\u52a1\u5bf9\u7cbe\u5ea6\u654f\u611f\u5ea6\u4e0d\u540c\u3002\u4e0a\u7ebf\u524d\u81f3\u5c11\u6bd4\u8f83\u56db\u9879&#xff1a;\u4efb\u52a1\u51c6\u786e\u7387\u3001TTFT\u3001TPOT \u548c\u5cf0\u503c\u663e\u5b58\u3002\u4e0d\u8981\u4ec5\u51ed\u6a21\u578b\u6587\u4ef6\u5927\u5c0f\u9009\u62e9\u91cf\u5316\u683c\u5f0f\u3002<\/p>\n<p>SGLang\u3001vLLM \u548c TensorRT-LLM \u90fd\u4f1a\u6839\u636e\u7248\u672c\u548c\u786c\u4ef6\u9009\u62e9 Attention\u3001\u91c7\u6837\u548c\u901a\u4fe1\u5b9e\u73b0\u3002SGLang\u3001vLLM \u5e38\u89c1 FlashInfer\u3001PagedAttention \u6216 CUDA Graph \u540e\u7aef&#xff0c;TensorRT-LLM \u5219\u66f4\u591a\u4f9d\u8d56\u6784\u5efa Engine \u65f6\u786e\u5b9a\u7684\u878d\u5408 Kernel \u548c CUDA Graph \u914d\u7f6e\u3002\u9047\u5230\u5347\u7ea7\u540e\u6027\u80fd\u56de\u9000\u65f6&#xff0c;\u5e94\u4fdd\u5b58\u542f\u52a8\u65e5\u5fd7\u3001\u5b9e\u9645\u751f\u6548\u7684 backend \u548c Engine \u5143\u6570\u636e&#xff0c;\u4f7f\u7528\u540c\u4e00\u7ec4\u8bf7\u6c42\u505a A\/B \u6d4b\u8bd5&#xff0c;\u907f\u514d\u540c\u65f6\u66f4\u6362\u6a21\u578b\u3001\u9a71\u52a8\u548c\u6846\u67b6\u7248\u672c\u3002<\/p>\n<h3>\u56db\u3001\u751f\u4ea7\u5b9e\u8df5&#xff1a;\u751f\u4ea7\u6cbb\u7406\u3001\u76d1\u63a7\u4e0e\u6392\u969c<\/h3>\n<h4>1. \u76d1\u63a7\u548c\u65e5\u5fd7<\/h4>\n<p>\u5f00\u542f\u5f15\u64ce\u5bf9\u5e94\u7684 metrics \u9009\u9879\u540e&#xff0c;\u5c06 SGLang\u3001vLLM \u6216 Triton\/TensorRT-LLM \u7684\u6307\u6807\u7aef\u70b9\u7eb3\u5165 Prometheus \u6216 vmagent\u3002Grafana \u770b\u677f\u81f3\u5c11\u5e94\u5305\u62ec&#xff1a;\u8bf7\u6c42\u6570\u3001\u6210\u529f\u7387\u30015xx\u3001TTFT\u3001TPOT\u3001E2E P99\u3001\u961f\u5217\u957f\u5ea6\u3001\u8fd0\u884c\u4e2d\u8bf7\u6c42\u6570\u3001\u8f93\u5165\/\u8f93\u51fa Token\u3001KV Cache \u4f7f\u7528\u7387\u3001\u7f13\u5b58\u547d\u4e2d\u7387\u548c\u6a21\u578b\u52a0\u8f7d\u72b6\u6001\u3002\u91c7\u7528 PD \u5206\u79bb\u65f6&#xff0c;\u8fd8\u8981\u6309 Prefill Pool\u3001Decode Pool \u5206\u522b\u5c55\u793a\u961f\u5217\u957f\u5ea6\u3001GPU \u5229\u7528\u7387\u3001\u5904\u7406 Token \u6570&#xff0c;\u5e76\u8bb0\u5f55 KV Transfer \u7684\u5b57\u8282\u6570\u3001\u8017\u65f6\u3001\u5931\u8d25\u7387\u548c\u56de\u9000\u6b21\u6570\u3002TensorRT-LLM \u7ecf Triton \u90e8\u7f72\u65f6&#xff0c;\u8fd8\u5e94\u91c7\u96c6 Triton \u7684\u8bf7\u6c42\u961f\u5217\u3001\u6279\u5904\u7406\u548c\u5b9e\u4f8b\u7ea7\u6307\u6807\u3002<\/p>\n<p>\u4e0d\u540c\u7248\u672c\u7684\u6307\u6807\u540d\u79f0\u53ef\u80fd\u53d8\u5316&#xff0c;\u63a5\u5165\u65f6\u5e94\u5148\u6293\u53d6 \/metrics \u786e\u8ba4\u5b9e\u9645\u540d\u79f0&#xff0c;\u518d\u7f16\u5199\u5f55\u5236\u89c4\u5219\u548c\u544a\u8b66\u3002\u4e0d\u8981\u5728\u770b\u677f\u4e2d\u786c\u7f16\u7801\u67d0\u4e2a\u7248\u672c\u7684\u6307\u6807\u540d\u800c\u4e0d\u505a\u5347\u7ea7\u9a8c\u8bc1\u3002<\/p>\n<p>\u65e5\u5fd7\u5efa\u8bae\u5305\u542b\u4ee5\u4e0b\u5b57\u6bb5&#xff0c;\u5e76\u4f20\u9012\u5230 ELK&#xff1a;<\/p>\n<p><span class=\"token punctuation\">{<\/span><br \/>\n  <span class=\"token string-property property\">&#034;timestamp&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;2026-09-12T08:30:00Z&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;service&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;inference-qwen&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;engine&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;sglang&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;model&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;qwen2.5-32b&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;request_id&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;req-abc123&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;trace_id&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;4bf92f3577b34da6a3ce929d0e0e4736&#034;<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;input_tokens&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">2048<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;output_tokens&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">256<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;ttft_ms&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token number\">420<\/span><span class=\"token punctuation\">,<\/span><br \/>\n  <span class=\"token string-property property\">&#034;status&#034;<\/span><span class=\"token operator\">:<\/span> <span class=\"token string\">&#034;ok&#034;<\/span><br \/>\n<span class=\"token punctuation\">}<\/span><\/p>\n<p>\u6a21\u578b Prompt\u3001\u7528\u6237\u8f93\u5165\u548c\u8f93\u51fa\u5185\u5bb9\u53ef\u80fd\u5305\u542b\u9690\u79c1\u6216\u5bc6\u94a5\u3002\u9ed8\u8ba4\u53ea\u8bb0\u5f55\u957f\u5ea6\u3001\u54c8\u5e0c\u3001\u9519\u8bef\u7c7b\u578b\u548c\u91c7\u6837\u540e\u7684\u7247\u6bb5&#xff1b;\u9700\u8981\u5ba1\u8ba1\u5168\u6587\u65f6\u5355\u72ec\u8d70\u6743\u9650\u3001\u8131\u654f\u548c\u4fdd\u7559\u5468\u671f&#xff0c;\u4e0d\u8981\u628a\u539f\u59cb\u5185\u5bb9\u65e0\u6761\u4ef6\u5199\u5165\u666e\u901a\u65e5\u5fd7\u3002<\/p>\n<h4>2. \u544a\u8b66\u548c\u5f39\u6027\u7b56\u7565<\/h4>\n<p>\u6307\u6807\u544a\u8b66\u53ef\u4ee5\u7531\u591c\u83ba\u6216\u5176\u4ed6\u544a\u8b66\u5e73\u53f0\u7edf\u4e00\u7ba1\u7406&#xff0c;\u5178\u578b\u89c4\u5219\u5305\u62ec&#xff1a;<\/p>\n<p>\u63a8\u7406\u670d\u52a1\u8fdb\u7a0b\u4e0d\u53ef\u7528\u6301\u7eed 2 \u5206\u949f       \u2192 Critical<br \/>\n5 \u5206\u949f P99 TTFT \u8d85\u8fc7\u76ee\u6807\u503c          \u2192 Warning \/ Critical<br \/>\nKV Cache \u4f7f\u7528\u7387\u6301\u7eed\u8d85\u8fc7 90%         \u2192 Warning<br \/>\nGPU \u6e29\u5ea6\u3001XID\u3001ECC \u9519\u8bef\u5f02\u5e38          \u2192 Critical<br \/>\n5xx \u6216 OOM \u6b21\u6570\u5728\u7a97\u53e3\u5185\u5feb\u901f\u589e\u957f      \u2192 Critical<br \/>\n\u8bf7\u6c42\u961f\u5217\u6301\u7eed\u589e\u957f\u4e14\u65e0\u4e0b\u964d\u8d8b\u52bf         \u2192 Warning<br \/>\nKV Transfer \u5931\u8d25\u7387\u6216\u56de\u9000\u6b21\u6570\u6301\u7eed\u5347\u9ad8  \u2192 Critical<\/p>\n<p>\u5f39\u6027\u4f38\u7f29\u4e0d\u8981\u53ea\u4f9d\u636e CPU\u3002\u63a8\u7406\u670d\u52a1\u7684\u4e3b\u8981\u74f6\u9888\u5f80\u5f80\u662f GPU \u663e\u5b58\u3001KV Cache \u548c\u961f\u5217\u957f\u5ea6&#xff1b;\u800c\u65b0 Pod \u7684\u6a21\u578b\u52a0\u8f7d\u65f6\u95f4\u53c8\u53ef\u80fd\u5f88\u957f\u3002\u66f4\u53ef\u9760\u7684\u65b9\u5f0f\u662f&#xff1a;\u7f51\u5173\u6309\u79df\u6237\u548c\u6a21\u578b\u9650\u6d41&#xff0c;\u4f7f\u7528\u961f\u5217\u957f\u5ea6\u6216\u7b49\u5f85\u65f6\u95f4\u89e6\u53d1\u6269\u5bb9&#xff0c;\u63d0\u524d\u9884\u70ed\u5907\u7528\u526f\u672c&#xff0c;\u7f29\u5bb9\u65f6\u5148\u505c\u6b62\u63a5\u6536\u65b0\u8bf7\u6c42\u5e76\u7b49\u5f85\u6d41\u5f0f\u8bf7\u6c42\u5b8c\u6210\u3002<\/p>\n<h4>3. \u53d1\u5e03\u3001\u5347\u7ea7\u4e0e\u5b89\u5168\u8fb9\u754c<\/h4>\n<p>\u751f\u4ea7\u53d1\u5e03\u5efa\u8bae\u91c7\u7528\u4ee5\u4e0b\u6d41\u7a0b&#xff1a;<\/p>\n<li>\u56fa\u5b9a\u955c\u50cf\u3001\u63a8\u7406\u5f15\u64ce\u3001PyTorch\/TensorRT\u3001CUDA\u3001\u9a71\u52a8\u548c\u6a21\u578b revision&#xff0c;\u8bb0\u5f55 SBOM \u4e0e\u6821\u9a8c\u503c\u3002<\/li>\n<li>\u4f7f\u7528\u79bb\u7ebf\u6216\u53d7\u63a7\u7f51\u7edc\u4e0b\u8f7d\u6a21\u578b&#xff0c;\u6821\u9a8c\u6587\u4ef6\u5b8c\u6574\u6027&#xff0c;\u7981\u6b62\u8fd0\u884c\u65f6\u4efb\u610f\u62c9\u53d6\u8fdc\u7a0b\u4ee3\u7801\u3002<\/li>\n<li>\u65b0\u7248\u672c\u5148\u542f\u52a8\u5f71\u5b50\u5b9e\u4f8b&#xff0c;\u56de\u653e\u8131\u654f\u8bf7\u6c42&#xff0c;\u6bd4\u8f83\u51c6\u786e\u7387\u3001TTFT\u3001TPOT\u3001\u663e\u5b58\u548c\u9519\u8bef\u7387\u3002<\/li>\n<li>\u901a\u8fc7\u7f51\u5173\u6309\u79df\u6237\u6216\u6bd4\u4f8b\u7070\u5ea6&#xff0c;\u4fdd\u7559\u65e7\u7248\u672c\u76f4\u5230\u9519\u8bef\u9884\u7b97\u548c\u957f\u5c3e\u6307\u6807\u7a33\u5b9a\u3002<\/li>\n<li>\u56de\u6eda\u65f6\u5207\u56de\u65e7\u670d\u52a1&#xff0c;\u4e0d\u8981\u5728\u539f Pod \u4e0a\u76f4\u63a5\u66ff\u6362\u6a21\u578b\u76ee\u5f55\u3002<\/li>\n<p>\u63a8\u7406\u670d\u52a1\u4e0d\u5e94\u76f4\u63a5\u66b4\u9732\u516c\u7f51\u3002API Key\u3001OAuth\u3001\u79df\u6237\u914d\u989d\u3001\u8bf7\u6c42\u4f53\u5927\u5c0f\u3001\u6a21\u578b\u767d\u540d\u5355\u548c\u5ba1\u8ba1\u5e94\u653e\u5728\u7f51\u5173&#xff1b;\u670d\u52a1\u5230\u670d\u52a1\u901a\u4fe1\u4f7f\u7528 mTLS \u6216\u5185\u7f51\u7b56\u7565\u3002&#8211;trust-remote-code \u4f1a\u5141\u8bb8\u6267\u884c\u6a21\u578b\u4ed3\u5e93\u4e2d\u7684\u81ea\u5b9a\u4e49\u4ee3\u7801&#xff0c;\u53ea\u6709\u5728\u5ba1\u6838\u6765\u6e90\u548c\u7248\u672c\u540e\u624d\u5f00\u542f&#xff0c;\u5e76\u5c06\u6a21\u578b\u76ee\u5f55\u8bbe\u7f6e\u4e3a\u53ea\u8bfb\u3002TensorRT-LLM \u7684 Engine \u6587\u4ef6\u4e5f\u5e94\u89c6\u4e3a\u53d7\u4fdd\u62a4\u7684\u6784\u5efa\u4ea7\u7269&#xff0c;\u53ea\u5141\u8bb8\u6765\u81ea\u5df2\u5ba1\u6838\u6d41\u6c34\u7ebf\u3002<\/p>\n<h4>4. \u5e38\u89c1\u6545\u969c\u6392\u67e5<\/h4>\n<p>\u95ee\u9898\u4e00&#xff1a;\u670d\u52a1\u542f\u52a8\u65f6\u62a5 CUDA out of memory\u3002<\/p>\n<p>\u5148\u533a\u5206\u6a21\u578b\u6743\u91cd\u65e0\u6cd5\u52a0\u8f7d&#xff0c;\u8fd8\u662f\u8fd0\u884c\u65f6\u521d\u59cb\u5316\u5931\u8d25\u3002\u68c0\u67e5\u6a21\u578b\u7cbe\u5ea6\u3001GPU \u663e\u5b58\u3001TP \u6570\u91cf\u548c\u662f\u5426\u6709\u5176\u4ed6\u8fdb\u7a0b\u5360\u5361&#xff1b;\u518d\u964d\u4f4e\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u5f15\u64ce\u663e\u5b58\u9884\u7b97\u6216\u9009\u62e9\u91cf\u5316\u6a21\u578b\u3002\u4e0d\u8981\u53ea\u589e\u52a0 TP&#xff0c;\u8de8\u5361\u901a\u4fe1\u548c\u6a21\u578b\u5207\u5206\u7ea6\u675f\u53ef\u80fd\u8ba9\u95ee\u9898\u66f4\u590d\u6742\u3002<\/p>\n<p>\u95ee\u9898\u4e8c&#xff1a;GPU \u5229\u7528\u7387\u4e0d\u9ad8&#xff0c;\u4f46\u8bf7\u6c42\u5f88\u6162\u3002<\/p>\n<p>\u4f18\u5148\u67e5\u770b Queue Time\u3001TTFT \u548c\u8f93\u5165\u957f\u5ea6\u3002\u5982\u679c\u961f\u5217\u65f6\u95f4\u9ad8&#xff0c;\u8bf4\u660e\u5e76\u53d1\u9650\u5236\u6216\u8c03\u5ea6\u7b56\u7565\u5728\u7b49\u5f85&#xff1b;\u5982\u679c Prefill \u65f6\u95f4\u9ad8&#xff0c;\u68c0\u67e5\u957f\u4e0a\u4e0b\u6587\u548c\u524d\u7f00\u7f13\u5b58&#xff1b;\u5982\u679c Decode \u95f4\u9694\u9ad8&#xff0c;\u68c0\u67e5\u8f93\u51fa\u957f\u5ea6\u3001\u6279\u6b21\u3001\u91c7\u6837\u548c GPU \u65f6\u949f\u3002CPU Tokenizer\u3001\u7f51\u7edc\u53d1\u9001\u963b\u585e\u548c\u5ba2\u6237\u7aef\u6ca1\u6709\u53ca\u65f6\u8bfb\u53d6 SSE&#xff0c;\u4e5f\u53ef\u80fd\u8ba9 GPU \u770b\u8d77\u6765\u4e0d\u5fd9\u3002<\/p>\n<p>\u95ee\u9898\u4e09&#xff1a;\u591a\u5361\u542f\u52a8\u540e\u5361\u4f4f\u6216 NCCL \u62a5\u9519\u3002<\/p>\n<p>\u68c0\u67e5\u5bb9\u5668\u662f\u5426\u4f7f\u7528 &#8211;ipc&#061;host\u3001\u5171\u4eab\u5185\u5b58\u662f\u5426\u8db3\u591f\u3001\u5404\u5361\u9a71\u52a8\u548c CUDA \u662f\u5426\u4e00\u81f4\u3001NCCL \u7f51\u5361\u9009\u62e9\u662f\u5426\u6b63\u786e&#xff0c;\u4ee5\u53ca Pod \u662f\u5426\u88ab\u8c03\u5ea6\u5230\u540c\u4e00\u53f0\u673a\u5668\u3002\u8de8\u4e3b\u673a TP \u8fd8\u8981\u9a8c\u8bc1 RDMA\u3001\u7aef\u53e3\u3001\u5b89\u5168\u7ec4\u548c NCCL_SOCKET_IFNAME \u7b49\u914d\u7f6e\u3002<\/p>\n<p>\u95ee\u9898\u56db&#xff1a;\u9996 Token \u5ef6\u8fdf\u7a81\u7136\u5347\u9ad8\u3002<\/p>\n<p>\u68c0\u67e5\u662f\u5426\u53d1\u751f\u524d\u7f00\u7f13\u5b58\u672a\u547d\u4e2d\u3001\u6a21\u578b\u521a\u91cd\u542f\u3001\u957f\u8bf7\u6c42\u7a81\u53d1\u6216 GPU \u964d\u9891\u3002\u5c06\u7cfb\u7edf Prompt\u3001\u5de5\u5177\u5b9a\u4e49\u548c\u6d88\u606f\u6a21\u677f\u56fa\u5b9a\u4e0b\u6765&#xff0c;\u533a\u5206\u51b7\u7f13\u5b58\u4e0e\u70ed\u7f13\u5b58\u538b\u6d4b&#xff1b;\u4e0d\u8981\u53ea\u770b\u5e73\u5747\u503c&#xff0c;\u8981\u89c2\u5bdf P95\/P99 \u548c\u6392\u961f\u65f6\u95f4\u3002<\/p>\n<p>\u95ee\u9898\u4e94&#xff1a;\u6d41\u5f0f\u8f93\u51fa\u4e2d\u9014\u65ad\u5f00\u3002<\/p>\n<p>\u68c0\u67e5\u7f51\u5173\u548c\u8d1f\u8f7d\u5747\u8861\u5668\u7684\u54cd\u5e94\u8d85\u65f6\u3001SSE \u7f13\u51b2\u3001\u5ba2\u6237\u7aef\u53d6\u6d88\u4f20\u64ad\u548c\u670d\u52a1\u7aef\u4f18\u96c5\u9000\u51fa\u3002\u7f51\u5173\u8981\u900f\u4f20 text\/event-stream&#xff0c;\u5b9a\u671f\u53d1\u9001\u5fc3\u8df3\u6216\u4fdd\u6301\u6570\u636e\u5237\u65b0&#xff0c;\u5e76\u5728\u53d6\u6d88\u65f6\u53ca\u65f6\u91ca\u653e\u751f\u6210\u8bf7\u6c42\u548c KV Cache\u3002<\/p>\n<p>\u95ee\u9898\u516d&#xff1a;\u7ed3\u6784\u5316\u8f93\u51fa\u5076\u5c14\u4e0d\u662f\u5408\u6cd5 JSON\u3002<\/p>\n<p>\u786e\u8ba4\u4f7f\u7528\u7684\u6a21\u578b\u3001\u63a8\u7406\u5f15\u64ce\u7248\u672c\u548c API \u53c2\u6570\u786e\u5b9e\u652f\u6301\u76ee\u6807\u7ea6\u675f&#xff1b;\u4e3a JSON Schema\u3001\u6b63\u5219\u548c stop \u6761\u4ef6\u5efa\u7acb\u56de\u5f52\u6d4b\u8bd5&#xff0c;\u670d\u52a1\u7aef\u5bf9\u7ed3\u679c\u505a\u89e3\u6790\u6821\u9a8c&#xff0c;\u5931\u8d25\u65f6\u8fdb\u884c\u6709\u9650\u6b21\u6570\u91cd\u8bd5\u6216\u8fd4\u56de\u660e\u786e\u9519\u8bef&#xff0c;\u4e0d\u80fd\u628a\u6821\u9a8c\u8d23\u4efb\u5168\u90e8\u4ea4\u7ed9\u524d\u7aef\u3002<\/p>\n<p>\u95ee\u9898\u4e03&#xff1a;\u542f\u7528 PD \u5206\u79bb\u540e&#xff0c;TTFT \u53cd\u800c\u5347\u9ad8\u6216\u8bf7\u6c42\u5076\u53d1\u5931\u8d25\u3002<\/p>\n<p>\u5148\u5206\u522b\u67e5\u770b Prefill\u3001KV Transfer \u548c Decode \u4e09\u6bb5\u8017\u65f6&#xff0c;\u786e\u8ba4\u662f\u5426\u662f\u7f51\u7edc\u4f20\u8f93\u3001KV Cache \u683c\u5f0f\u4e0d\u517c\u5bb9\u3001\u4e24\u4fa7 TP\/PP \u914d\u7f6e\u4e0d\u4e00\u81f4\u6216\u8def\u7531\u5668\u91cd\u590d\u8f6c\u53d1\u3002\u666e\u901a\u8de8\u673a\u7f51\u7edc\u4e0b&#xff0c;KV Transfer \u53ef\u80fd\u62b5\u6d88 Prefill \u5e76\u884c\u6536\u76ca&#xff1b;\u5e94\u964d\u4f4e\u4f20\u8f93\u8ddd\u79bb\u3001\u4f7f\u7528 RDMA\/NVLink&#xff0c;\u6216\u6682\u65f6\u56de\u9000\u5230\u5408\u5e76\u5b9e\u4f8b\u3002\u751f\u4ea7\u4e0a\u5fc5\u987b\u4fdd\u7559\u56de\u9000\u8def\u5f84&#xff0c;\u5e76\u4e3a\u8f6c\u79fb\u8d85\u65f6\u3001\u7248\u672c\u4e0d\u5339\u914d\u548c\u76ee\u6807 Decode \u5b9e\u4f8b\u4e0d\u53ef\u7528\u8bbe\u7f6e\u660e\u786e\u9519\u8bef\u7801\u3002<\/p>\n<h4>5. \u5e38\u89c1\u95ee\u7b54<\/h4>\n<p>Q&#xff1a;SGLang\u3001vLLM \u548c TensorRT-LLM&#xff0c;\u54ea\u4e2a\u4e00\u5b9a\u66f4\u5feb&#xff1f;<\/p>\n<p>A&#xff1a;\u6ca1\u6709\u8131\u79bb\u8d1f\u8f7d\u7684\u56fa\u5b9a\u7b54\u6848\u3002SGLang \u5728\u5171\u4eab\u524d\u7f00\u3001\u591a\u8f6e\u5bf9\u8bdd\u548c\u7ed3\u6784\u5316\u751f\u6210\u573a\u666f\u53ef\u80fd\u4f18\u52bf\u660e\u663e&#xff0c;vLLM \u901a\u7528\u670d\u52a1\u548c\u6a21\u578b\u9002\u914d\u7ecf\u9a8c\u4e30\u5bcc&#xff0c;TensorRT-LLM \u5728\u56fa\u5b9a NVIDIA GPU \u548c Engine \u7684\u573a\u666f\u53ef\u80fd\u53d6\u5f97\u66f4\u9ad8\u541e\u5410\u3002\u6700\u7ec8\u5e94\u4f7f\u7528\u751f\u4ea7\u8bf7\u6c42\u5206\u5e03\u505a\u538b\u6d4b&#xff0c;\u81f3\u5c11\u6bd4\u8f83 TTFT\u3001TPOT\u3001P99\u3001\u541e\u5410\u3001\u663e\u5b58\u548c\u8fd0\u7ef4\u6210\u672c\u3002<\/p>\n<p>Q&#xff1a;TP \u8d8a\u5927\u8d8a\u597d\u5417&#xff1f;<\/p>\n<p>A&#xff1a;\u4e0d\u662f\u3002SGLang \u4f7f\u7528 &#8211;tp-size&#xff0c;vLLM \u4f7f\u7528 &#8211;tensor-parallel-size&#xff0c;TensorRT-LLM \u901a\u5e38\u5728 Engine \u6784\u5efa\u9636\u6bb5\u914d\u7f6e TP\/PP\u3002TP \u89e3\u51b3\u5355\u5361\u653e\u4e0d\u4e0b\u6216\u5355\u5b9e\u4f8b\u541e\u5410\u4e0d\u8db3\u7684\u95ee\u9898&#xff0c;\u4f46\u4f1a\u589e\u52a0\u8de8\u5361\u901a\u4fe1\u3002\u6a21\u578b\u80fd\u5728\u5355\u5361\u6216\u5c0f TP \u4e0a\u7a33\u5b9a\u8fd0\u884c\u65f6&#xff0c;\u589e\u52a0\u591a\u4e2a\u526f\u672c\u5f80\u5f80\u6bd4\u6269\u5927 TP \u66f4\u5bb9\u6613\u63d0\u5347\u6574\u4f53\u541e\u5410\u3002<\/p>\n<p>Q&#xff1a;\u4e3a\u4ec0\u4e48\u5df2\u7ecf\u5f00\u542f\u524d\u7f00\u7f13\u5b58&#xff0c;\u547d\u4e2d\u7387\u8fd8\u662f\u5f88\u4f4e&#xff1f;<\/p>\n<p>A&#xff1a;\u7f13\u5b58\u6309 Token \u524d\u7f00\u5339\u914d\u3002\u7cfb\u7edf\u63d0\u793a\u8bcd\u7684\u7a7a\u683c\u3001\u5de5\u5177\u5217\u8868\u987a\u5e8f\u3001Chat Template\u3001\u7248\u672c\u5b57\u6bb5\u6216\u52a8\u6001\u5185\u5bb9\u653e\u7f6e\u4f4d\u7f6e\u53d8\u5316&#xff0c;\u90fd\u4f1a\u7834\u574f\u590d\u7528\u3002\u5e94\u56fa\u5b9a\u5e8f\u5217\u5316\u6a21\u677f&#xff0c;\u628a\u7a33\u5b9a\u5185\u5bb9\u653e\u5728\u524d\u9762&#xff0c;\u5e76\u76f4\u63a5\u67e5\u770b\u8fd0\u884c\u6307\u6807\u786e\u8ba4\u547d\u4e2d\u7387\u3002<\/p>\n<p>Q&#xff1a;\u63a8\u7406\u670d\u52a1\u9700\u8981\u81ea\u5df1\u5b9e\u73b0\u9274\u6743\u5417&#xff1f;<\/p>\n<p>A&#xff1a;\u670d\u52a1\u53ef\u4ee5\u653e\u7f6e\u6700\u57fa\u7840\u7684\u8bbf\u95ee\u63a7\u5236&#xff0c;\u4f46\u751f\u4ea7\u4e0a\u5e94\u7531 AI Gateway \u7edf\u4e00\u8d1f\u8d23 API Key\u3001OAuth\u3001\u79df\u6237\u9650\u6d41\u3001\u914d\u989d\u3001\u5ba1\u8ba1\u548c\u6a21\u578b\u767d\u540d\u5355\u3002\u4e09\u79cd\u5f15\u64ce\u53ea\u66b4\u9732\u5728\u53d7\u63a7\u7f51\u7edc\u4e2d&#xff0c;\u4fbf\u4e8e\u5347\u7ea7\u548c\u9694\u79bb\u3002<\/p>\n<h4>6. \u4e00\u4efd\u4e0a\u7ebf\u68c0\u67e5\u6e05\u5355<\/h4>\n<p>[ ] \u56fa\u5b9a\u63a8\u7406\u5f15\u64ce\u3001PyTorch\/TensorRT\u3001CUDA\u3001\u9a71\u52a8\u548c\u6a21\u578b revision<br \/>\n[ ] \u5b8c\u6210\u5355\u8bf7\u6c42\u3001\u7f13\u5b58\u547d\u4e2d\u3001\u7a33\u5b9a\u5e76\u53d1\u548c\u7a81\u53d1\u6d41\u91cf\u538b\u6d4b<br \/>\n[ ] \u660e\u786e TP\/DP\u3001GPU \u62d3\u6251\u3001\/dev\/shm \u548c NCCL \u914d\u7f6e<br \/>\n[ ] \u8bbe\u7f6e\u4e0a\u4e0b\u6587\u3001\u8f93\u51fa Token\u3001\u5e76\u53d1\u548c\u8bf7\u6c42\u4f53\u5927\u5c0f\u4e0a\u9650<br \/>\n[ ] readiness \u7b49\u5f85\u6a21\u578b\u771f\u6b63\u52a0\u8f7d\u5b8c\u6210&#xff0c;\u53d1\u5e03\u652f\u6301\u56de\u6eda<br \/>\n[ ] \/metrics \u5df2\u63a5\u5165 Prometheus\/vmagent&#xff0c;Grafana \u770b\u677f\u53ef\u7528<br \/>\n[ ] TTFT\u3001TPOT\u3001P99\u3001\u961f\u5217\u3001KV Cache\u3001GPU \u548c OOM \u5df2\u544a\u8b66<br \/>\n[ ] \u82e5\u91c7\u7528 PD \u5206\u79bb&#xff0c;\u5df2\u9a8c\u8bc1 KV Transfer\u3001\u7f51\u7edc\u5e26\u5bbd\u3001\u56de\u9000\u8def\u5f84\u548c\u4e24\u4fa7\u7248\u672c\u517c\u5bb9\u6027<br \/>\n[ ] \u65e5\u5fd7\u5305\u542b request_id\u3001trace_id\u3001model \u548c Token \u7edf\u8ba1\u5e76\u5b8c\u6210\u8131\u654f<br \/>\n[ ] \u7f51\u5173\u5df2\u914d\u7f6e\u9274\u6743\u3001\u9650\u6d41\u3001SSE \u8d85\u65f6\u548c\u5ba2\u6237\u7aef\u53d6\u6d88\u4f20\u64ad<br \/>\n[ ] \u5df2\u6f14\u7ec3 OOM\u3001NCCL \u6545\u969c\u3001\u6a21\u578b\u52a0\u8f7d\u5931\u8d25\u548c\u901a\u77e5\u5931\u8d25<\/p>\n<h3>\u4e94\u3001\u5e73\u53f0\u5316\u6f14\u8fdb&#xff1a;\u591a\u6a21\u578b\u3001\u591a\u79df\u6237\u3001\u5bb9\u91cf\u4e0e\u6210\u672c\u6cbb\u7406<\/h3>\n<p>\u5355\u4e2a\u6a21\u578b\u80fd\u591f\u7a33\u5b9a\u63d0\u4f9b\u670d\u52a1&#xff0c;\u53ea\u662f\u63a8\u7406\u5e73\u53f0\u7684\u8d77\u70b9\u3002\u5f53\u6a21\u578b\u6570\u91cf\u3001\u4e1a\u52a1\u56e2\u961f\u548c\u8c03\u7528\u91cf\u589e\u52a0\u540e&#xff0c;\u65b0\u7684\u74f6\u9888\u901a\u5e38\u51fa\u73b0\u5728\u6a21\u578b\u53d1\u5e03\u3001\u8d44\u6e90\u5206\u914d\u3001\u79df\u6237\u9694\u79bb\u548c\u6210\u672c\u6838\u7b97&#xff0c;\u800c\u4e0d\u518d\u53ea\u662f\u67d0\u4e00\u5f20 GPU \u7684\u5229\u7528\u7387\u3002\u5e73\u53f0\u5316\u7684\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u4ece\u201c\u624b\u5de5\u542f\u52a8\u7684\u8fdb\u7a0b\u201d\u53d8\u6210\u53ef\u6ce8\u518c\u3001\u53ef\u8def\u7531\u3001\u53ef\u8ba1\u91cf\u3001\u53ef\u56de\u6eda\u7684\u6807\u51c6\u670d\u52a1\u3002<\/p>\n<h4>1. \u591a\u6a21\u578b\u6ce8\u518c\u4e0e\u667a\u80fd\u8def\u7531<\/h4>\n<p>\u5efa\u8bae\u5efa\u7acb\u7edf\u4e00\u7684\u6a21\u578b\u6ce8\u518c\u8868&#xff0c;\u81f3\u5c11\u8bb0\u5f55\u6a21\u578b\u540d\u79f0\u3001\u7248\u672c\u6216 revision\u3001\u91cf\u5316\u683c\u5f0f\u3001\u4e0a\u4e0b\u6587\u4e0a\u9650\u3001\u6240\u9700 GPU\u3001\u9002\u914d\u7684\u5f15\u64ce\u3001SLO \u548c\u6570\u636e\u5408\u89c4\u7b49\u7ea7\u3002\u6a21\u578b\u6743\u91cd\u3001\u542f\u52a8\u53c2\u6570\u548c Engine \u5e94\u4f5c\u4e3a\u4e0d\u53ef\u53d8\u5236\u54c1\u7ba1\u7406&#xff0c;\u907f\u514d\u540c\u4e00\u4e2a\u6a21\u578b\u540d\u5728\u4e0d\u540c\u8282\u70b9\u5b9e\u9645\u8fd0\u884c\u4e0d\u540c\u7248\u672c\u3002<\/p>\n<p>\u7f51\u5173\u6216\u6a21\u578b\u8def\u7531\u5c42\u53ef\u4ee5\u6839\u636e\u4ee5\u4e0b\u7ef4\u5ea6\u9009\u62e9\u540e\u7aef&#xff1a;<\/p>\n<table>\n<tr>\u8def\u7531\u7ef4\u5ea6\u793a\u4f8b\u7b56\u7565\u76ee\u7684<\/tr>\n<tbody>\n<tr>\n<td>\u4efb\u52a1\u7c7b\u578b<\/td>\n<td>\u5bf9\u8bdd\u8d70\u901a\u7528\u6a21\u578b&#xff0c;\u4ee3\u7801\u8d70\u4ee3\u7801\u6a21\u578b&#xff0c;Embedding \u8d70\u4e13\u7528\u670d\u52a1<\/td>\n<td>\u51cf\u5c11\u9519\u8bef\u9009\u578b\u548c\u65e0\u6548\u6d88\u8017<\/td>\n<\/tr>\n<tr>\n<td>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/td>\n<td>\u77ed\u8bf7\u6c42\u8d70\u4f4e\u5ef6\u8fdf\u6c60&#xff0c;\u8d85\u957f\u8bf7\u6c42\u8d70\u957f\u4e0a\u4e0b\u6587\u6c60<\/td>\n<td>\u907f\u514d\u957f\u8bf7\u6c42\u62d6\u6162\u4ea4\u4e92\u6d41\u91cf<\/td>\n<\/tr>\n<tr>\n<td>\u670d\u52a1\u7b49\u7ea7<\/td>\n<td>\u9ad8\u4f18\u5148\u7ea7\u79df\u6237\u8d70\u9884\u70ed\u526f\u672c&#xff0c;\u666e\u901a\u4efb\u52a1\u5141\u8bb8\u6392\u961f<\/td>\n<td>\u4fdd\u8bc1\u4e0d\u540c SLA<\/td>\n<\/tr>\n<tr>\n<td>\u5f15\u64ce\u80fd\u529b<\/td>\n<td>\u9700\u8981\u7ed3\u6784\u5316\u751f\u6210\u65f6\u8def\u7531\u5230\u652f\u6301\u7ea6\u675f\u7684\u5f15\u64ce<\/td>\n<td>\u907f\u514d API \u80fd\u529b\u4e0d\u5339\u914d<\/td>\n<\/tr>\n<tr>\n<td>\u53d1\u5e03\u72b6\u6001<\/td>\n<td>\u6309\u79df\u6237\u6216\u6bd4\u4f8b\u5207\u5230 canary&#xff0c;\u5f02\u5e38\u65f6\u56de\u9000 stable<\/td>\n<td>\u652f\u6301\u7070\u5ea6\u4e0e\u5feb\u901f\u56de\u6eda<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8def\u7531\u7b56\u7565\u5e94\u907f\u514d\u201c\u53ea\u770b GPU \u7a7a\u95f2\u7387\u201d\u3002\u66f4\u6709\u4ef7\u503c\u7684\u4fe1\u53f7\u5305\u62ec\u9884\u8ba1\u8f93\u5165 Token\u3001\u9884\u8ba1\u8f93\u51fa Token\u3001\u961f\u5217\u7b49\u5f85\u65f6\u95f4\u3001KV Cache \u547d\u4e2d\u7387\u548c\u6a21\u578b\u52a0\u8f7d\u72b6\u6001\u3002\u5bf9\u540c\u4e00\u6a21\u578b\u7684\u591a\u4e2a\u526f\u672c&#xff0c;\u53ef\u4ee5\u4f7f\u7528\u6700\u77ed\u961f\u5217\u3001\u52a0\u6743\u8f6e\u8be2\u6216\u57fa\u4e8e Token \u9884\u7b97\u7684\u8c03\u5ea6&#xff1b;\u5bf9\u4e0d\u540c\u6a21\u578b&#xff0c;\u5219\u8981\u628a\u8d28\u91cf\u3001\u5ef6\u8fdf\u548c\u6210\u672c\u4e00\u8d77\u7eb3\u5165\u9009\u62e9\u3002<\/p>\n<h4>2. \u591a\u79df\u6237\u9694\u79bb\u3001\u914d\u989d\u4e0e\u516c\u5e73\u8c03\u5ea6<\/h4>\n<p>\u591a\u79df\u6237\u73af\u5883\u81f3\u5c11\u9700\u8981\u56db\u5c42\u9694\u79bb&#xff1a;\u7f51\u7edc\u8bbf\u95ee\u9694\u79bb\u3001\u6a21\u578b\u767d\u540d\u5355\u9694\u79bb\u3001\u5e76\u53d1\u4e0e Token \u914d\u989d\u9694\u79bb\u3001\u65e5\u5fd7\u548c\u6570\u636e\u6743\u9650\u9694\u79bb\u3002\u4e0d\u8981\u53ea\u9650\u5236 QPS&#xff0c;\u56e0\u4e3a\u4e00\u4e2a\u8bf7\u6c42\u53ef\u80fd\u5305\u542b\u5f88\u957f\u7684\u8f93\u5165\u548c\u8f93\u51fa&#xff1b;\u66f4\u5408\u7406\u7684\u914d\u989d\u5355\u4f4d\u662f\u8f93\u5165 Token\u3001\u8f93\u51fa Token\u3001\u5e76\u53d1\u8bf7\u6c42\u6570\u548c GPU \u65f6\u95f4\u7684\u7ec4\u5408\u3002<\/p>\n<p>\u53ef\u4ee5\u6309\u79df\u6237\u7ef4\u62a4\u4ee5\u4e0b\u7b56\u7565&#xff1a;<\/p>\n<p>\u79df\u6237\u914d\u989d &#061; \u6bcf\u5206\u949f\u8bf7\u6c42\u6570 &#043; \u6bcf\u5206\u949f\u8f93\u5165 Token &#043; \u6bcf\u5206\u949f\u8f93\u51fa Token<br \/>\n\u5355\u8bf7\u6c42\u4e0a\u9650 &#061; \u6700\u5927\u4e0a\u4e0b\u6587 Token &#043; \u6700\u5927\u8f93\u51fa Token &#043; \u8bf7\u6c42\u4f53\u5927\u5c0f<br \/>\n\u5e76\u53d1\u4e0a\u9650 &#061; \u4ea4\u4e92\u8bf7\u6c42\u5e76\u53d1 &#043; \u6d41\u5f0f\u8bf7\u6c42\u5e76\u53d1 &#043; \u79bb\u7ebf\u4efb\u52a1\u5e76\u53d1<\/p>\n<p>\u8c03\u5ea6\u5668\u5e94\u652f\u6301\u8d85\u989d\u62d2\u7edd\u3001\u6392\u961f\u3001\u964d\u7ea7\u6a21\u578b\u548c\u4f18\u5148\u7ea7\u7ee7\u627f\u3002\u9ad8\u4f18\u5148\u7ea7\u4e0d\u80fd\u65e0\u9650\u6324\u5360\u7cfb\u7edf\u8d44\u6e90&#xff0c;\u5426\u5219\u4f4e\u4f18\u5148\u7ea7\u4efb\u52a1\u4f1a\u5f62\u6210\u9965\u997f&#xff1b;\u5efa\u8bae\u4e3a\u6bcf\u4e2a\u79df\u6237\u8bbe\u7f6e\u6700\u5927\u6392\u961f\u65f6\u95f4&#xff0c;\u5e76\u5728\u7f51\u5173\u8fd4\u56de\u53ef\u91cd\u8bd5\u7684\u9519\u8bef\u7801\u3002\u6307\u6807\u6807\u7b7e\u4e2d\u4e0d\u8981\u76f4\u63a5\u4f7f\u7528\u4efb\u610f\u7528\u6237 ID&#xff0c;\u5426\u5219 Prometheus \u4f1a\u4ea7\u751f\u9ad8\u57fa\u6570&#xff1b;\u79df\u6237\u660e\u7ec6\u53ef\u4ee5\u653e\u5728\u65e5\u5fd7\u6216\u8ba1\u8d39\u7cfb\u7edf\u4e2d\u3002<\/p>\n<h4>3. \u5bb9\u91cf\u89c4\u5212\u4e0e\u5355\u4f4d Token \u6210\u672c<\/h4>\n<p>\u5bb9\u91cf\u89c4\u5212\u4e0d\u80fd\u53ea\u7528\u201cGPU \u5229\u7528\u7387\u8fbe\u5230 80%\u201d\u4f5c\u4e3a\u6269\u5bb9\u4f9d\u636e\u3002\u5e94\u5148\u6839\u636e\u751f\u4ea7\u8bf7\u6c42\u5206\u5e03\u6d4b\u51fa\u6709\u6548\u541e\u5410&#xff0c;\u518d\u7ed3\u5408 SLO \u53cd\u63a8\u526f\u672c\u6570\u91cf\u3002\u4e00\u4e2a\u7b80\u5355\u7684\u4f30\u7b97\u516c\u5f0f\u662f&#xff1a;<\/p>\n<p>\u6240\u9700\u526f\u672c\u6570 \u2248 \u5cf0\u503c Token \u901f\u7387 \u00f7 \u5355\u526f\u672c\u5728\u76ee\u6807 P99 \u4e0b\u7684\u6709\u6548 Token \u901f\u7387<br \/>\n\u5355\u4f4d Token \u6210\u672c \u2248 &#xff08;GPU \u5c0f\u65f6\u6210\u672c \u00d7 GPU \u6570 \u00d7 \u8fd0\u884c\u5c0f\u65f6 &#043; \u5b58\u50a8\u4e0e\u7f51\u7edc\u6210\u672c&#xff09;\u00f7 \u5b9e\u9645\u5904\u7406 Token \u6570<\/p>\n<p>\u201c\u6709\u6548 Token \u901f\u7387\u201d\u5fc5\u987b\u5305\u542b\u6392\u961f\u3001\u5931\u8d25\u91cd\u8bd5\u3001\u9884\u70ed\u548c\u4f4e\u8d1f\u8f7d\u65f6\u6bb5&#xff0c;\u4e0d\u80fd\u76f4\u63a5\u4f7f\u7528\u538b\u6d4b\u62a5\u544a\u4e2d\u7684\u5cf0\u503c\u3002\u5bf9\u4e8e PD \u5206\u79bb\u67b6\u6784&#xff0c;\u8fd8\u8981\u5206\u522b\u8ba1\u7b97 Prefill \u548c Decode \u6c60\u7684\u5bb9\u91cf&#xff0c;\u907f\u514d\u4e00\u4fa7\u6269\u5bb9\u540e\u53e6\u4e00\u4fa7\u6210\u4e3a\u74f6\u9888\u3002\u6269\u5bb9\u89e6\u53d1\u53ef\u4ee5\u7ec4\u5408\u961f\u5217\u7b49\u5f85\u65f6\u95f4\u3001\u9884\u8ba1 Token \u8d1f\u8f7d\u3001GPU \u663e\u5b58\u6c34\u4f4d\u548c\u9519\u8bef\u7387&#xff1b;\u7f29\u5bb9\u524d\u8981\u7b49\u5f85\u6d41\u5f0f\u8bf7\u6c42\u5b8c\u6210&#xff0c;\u5e76\u4fdd\u7559\u8db3\u591f\u7684\u70ed\u526f\u672c\u3002<\/p>\n<p>\u6210\u672c\u4f18\u5316\u901a\u5e38\u6709\u4ee5\u4e0b\u987a\u5e8f&#xff1a;\u5148\u6e05\u7406\u65e0\u6548\u91cd\u8bd5\u548c\u8fc7\u5927\u7684 max_tokens&#xff0c;\u518d\u4f18\u5316 Prefix Cache \u547d\u4e2d\u7387\u548c\u6279\u5904\u7406&#xff0c;\u4e4b\u540e\u624d\u8003\u8651\u66f4\u6fc0\u8fdb\u7684\u91cf\u5316\u3001\u4f4e\u4ef7 GPU \u6216\u8de8\u5730\u57df\u8c03\u5ea6\u3002\u4efb\u4f55\u964d\u672c\u65b9\u6848\u90fd\u5fc5\u987b\u540c\u65f6\u56de\u5f52\u51c6\u786e\u7387\u3001P99 \u548c\u6545\u969c\u6062\u590d\u65f6\u95f4&#xff0c;\u4e0d\u80fd\u53ea\u770b\u4e91\u8d26\u5355\u4e0b\u964d\u3002<\/p>\n<h4>4. \u5f02\u6784 GPU\u3001Engine \u517c\u5bb9\u4e0e\u591a\u5730\u57df\u5bb9\u707e<\/h4>\n<p>\u4e0d\u540c GPU \u67b6\u6784\u5bf9 FP8\u3001BF16\u3001FlashInfer\u3001TensorRT Engine \u548c\u663e\u5b58\u5bb9\u91cf\u7684\u652f\u6301\u4e0d\u540c\u3002\u5e73\u53f0\u5e94\u7ef4\u62a4\u201c\u6a21\u578b\u2014\u5f15\u64ce\u2014GPU\u2014\u7cbe\u5ea6\u201d\u7684\u517c\u5bb9\u77e9\u9635&#xff0c;\u8c03\u5ea6\u65f6\u628a\u8282\u70b9\u6807\u7b7e\u3001\u9a71\u52a8\u7248\u672c\u3001CUDA \u7248\u672c\u548c Engine \u5143\u6570\u636e\u4f5c\u4e3a\u786c\u7ea6\u675f\u3002TensorRT-LLM Engine \u5c24\u5176\u4e0d\u80fd\u8de8\u4e0d\u517c\u5bb9\u7684 GPU \u67b6\u6784\u76f4\u63a5\u590d\u7528&#xff1b;SGLang \u548c vLLM \u4e5f\u9700\u8981\u9a8c\u8bc1\u5bf9\u5e94\u7248\u672c\u7684 Kernel \u548c\u91cf\u5316\u540e\u7aef\u3002<\/p>\n<p>\u591a\u5730\u57df\u90e8\u7f72\u65f6&#xff0c;\u53ef\u4ee5\u6309\u6570\u636e\u5408\u89c4\u3001\u7f51\u7edc\u8ddd\u79bb\u548c\u5bb9\u91cf\u6c34\u4f4d\u9009\u62e9\u4e3b\u533a\u57df\u4e0e\u5907\u7528\u533a\u57df\u3002\u8de8\u5730\u57df\u53ea\u540c\u6b65\u6a21\u578b\u5236\u54c1\u548c\u914d\u7f6e&#xff0c;\u4e0d\u5e94\u9ed8\u8ba4\u8f6c\u53d1\u5305\u542b\u7528\u6237\u539f\u6587\u7684\u8bf7\u6c42&#xff1b;\u6545\u969c\u5207\u6362\u524d\u8981\u660e\u786e\u4f1a\u8bdd\u72b6\u6001\u3001Prefix Cache \u548c\u6d41\u5f0f\u8bf7\u6c42\u5982\u4f55\u5904\u7406\u3002\u63a8\u8350\u91c7\u7528\u201c\u533a\u57df\u5185\u591a\u526f\u672c &#043; \u533a\u57df\u95f4\u51b7\u5907\u6216\u70ed\u5907\u201d\u7684\u5206\u5c42\u7b56\u7565&#xff0c;\u5e76\u5b9a\u671f\u6f14\u7ec3 DNS\u3001\u7f51\u5173\u3001\u6a21\u578b\u4ed3\u5e93\u548c\u5bc6\u94a5\u7cfb\u7edf\u540c\u65f6\u5f02\u5e38\u7684\u60c5\u51b5\u3002<\/p>\n<h4>5. \u4ece\u5355\u670d\u52a1\u5230\u63a8\u7406\u5e73\u53f0\u7684\u6f14\u8fdb\u8def\u7ebf<\/h4>\n<p>\u53ef\u4ee5\u6309\u4ee5\u4e0b\u9636\u6bb5\u63a8\u8fdb&#xff0c;\u907f\u514d\u4e00\u5f00\u59cb\u5c31\u5f15\u5165\u8fc7\u591a\u590d\u6742\u7ec4\u4ef6&#xff1a;<\/p>\n<table>\n<tr>\u9636\u6bb5\u91cd\u70b9\u80fd\u529b\u9a8c\u6536\u6807\u51c6<\/tr>\n<tbody>\n<tr>\n<td>P0 \u5355\u6a21\u578b\u670d\u52a1<\/td>\n<td>\u56fa\u5b9a\u7248\u672c\u3001\u5065\u5eb7\u68c0\u67e5\u3001OpenAI API\u3001\u57fa\u7840\u6307\u6807<\/td>\n<td>\u80fd\u7a33\u5b9a\u53d1\u5e03\u548c\u56de\u6eda<\/td>\n<\/tr>\n<tr>\n<td>P1 \u6807\u51c6\u5316\u670d\u52a1<\/td>\n<td>\u6a21\u578b\u6ce8\u518c\u8868\u3001\u7edf\u4e00\u7f51\u5173\u3001\u914d\u989d\u3001\u65e5\u5fd7\u548c\u6210\u672c\u6807\u7b7e<\/td>\n<td>\u591a\u4e2a\u6a21\u578b\u4f7f\u7528\u540c\u4e00\u5957\u6d41\u7a0b<\/td>\n<\/tr>\n<tr>\n<td>P2 \u6027\u80fd\u6c60\u5316<\/td>\n<td>\u4ea4\u4e92\/\u79bb\u7ebf\u5206\u6c60\u3001Prefix Cache\u3001Chunked Prefill\u3001PD \u5206\u79bb<\/td>\n<td>TTFT\u3001TPOT \u548c\u541e\u5410\u8fbe\u5230\u76ee\u6807<\/td>\n<\/tr>\n<tr>\n<td>P3 \u5e73\u53f0\u6cbb\u7406<\/td>\n<td>\u591a\u5730\u57df\u3001\u707e\u5907\u3001\u8ba1\u8d39\u3001\u8d28\u91cf\u8bc4\u4f30\u548c\u81ea\u52a8\u5316\u6269\u7f29\u5bb9<\/td>\n<td>\u6545\u969c\u53ef\u5207\u6362&#xff0c;\u6210\u672c\u53ef\u89e3\u91ca<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u6bcf\u4e2a\u9636\u6bb5\u90fd\u5e94\u4fdd\u7559\u6e05\u6670\u7684\u56de\u9000\u8def\u5f84\u3002\u53ea\u6709\u5f53\u8bf7\u6c42\u5206\u5e03\u3001\u7f51\u7edc\u6761\u4ef6\u548c\u56e2\u961f\u8fd0\u7ef4\u80fd\u529b\u90fd\u8fbe\u5230\u8981\u6c42\u65f6&#xff0c;\u624d\u628a\u5355\u4f53\u670d\u52a1\u5347\u7ea7\u4e3a PD \u5206\u79bb\u3001\u8fdc\u7a0b KV Cache \u6216\u8de8\u5730\u57df\u8c03\u5ea6&#xff1b;\u590d\u6742\u5ea6\u672c\u8eab\u4e0d\u662f\u6027\u80fd\u6307\u6807\u3002<\/p>\n<p>\u7ed3\u8bed<\/p>\n<p>\u63a8\u7406\u5f15\u64ce\u7684\u5de5\u7a0b\u4ef7\u503c\u5728\u4e8e\u628a\u6a21\u578b\u63a8\u7406\u4e2d\u7684\u91cd\u590d\u8ba1\u7b97\u3001\u8bf7\u6c42\u8c03\u5ea6\u548c\u751f\u6210\u63a7\u5236\u4ea4\u7ed9\u4e13\u95e8\u7684 runtime \u7ba1\u7406\u3002\u771f\u6b63\u4e0a\u7ebf\u65f6&#xff0c;\u91cd\u70b9\u4e0d\u5e94\u53ea\u662f\u542f\u52a8\u547d\u4ee4&#xff0c;\u800c\u662f\u5efa\u7acb\u4e00\u5957\u5b8c\u6574\u8fb9\u754c&#xff1a;SGLang\u3001vLLM \u6216 TensorRT-LLM \u8d1f\u8d23\u9ad8\u6548\u6267\u884c&#xff0c;GPU \u548c Kubernetes \u8d1f\u8d23\u8d44\u6e90\u9694\u79bb&#xff0c;AI Gateway \u8d1f\u8d23\u8eab\u4efd\u4e0e\u6d41\u91cf\u6cbb\u7406&#xff0c;Prometheus\/vmagent\u3001Grafana\u3001ELK \u548c\u591c\u83ba\u8d1f\u8d23\u53ef\u89c2\u6d4b\u4e0e\u544a\u8b66&#xff1b;\u5e73\u53f0\u5c42\u518d\u8d1f\u8d23\u6a21\u578b\u6ce8\u518c\u3001\u8def\u7531\u3001\u914d\u989d\u3001\u5bb9\u91cf\u548c\u6210\u672c\u6cbb\u7406\u3002<\/p>\n<p>\u5f53\u56e2\u961f\u80fd\u591f\u540c\u65f6\u56de\u7b54\u201c\u5f53\u524d\u6392\u961f\u591a\u957f\u3001\u9996 Token \u4e3a\u4ec0\u4e48\u6162\u3001\u524d\u7f00\u7f13\u5b58\u662f\u5426\u547d\u4e2d\u3001\u663e\u5b58\u8fd8\u5269\u591a\u5c11\u3001\u54ea\u4e2a\u79df\u6237\u53d7\u5230\u5f71\u54cd\u3001\u5982\u4f55\u5feb\u901f\u56de\u6eda\u201d\u65f6&#xff0c;\u6a21\u578b\u624d\u7b97\u4ece\u5b9e\u9a8c\u73af\u5883\u8fdb\u5165\u4e86\u53ef\u8fd0\u8425\u7684\u63a8\u7406\u5e73\u53f0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM \u6587\u7ae0\u76ee\u5f55\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM\u4e00\u3001\u63a8\u7406\u5f15\u64ce&#xff1a;\u9009\u578b\u4e0e\u8fd0\u884c\u539f\u7406&#xff08;SGLang\u3001vLLM\u3001TensorRT-LLM&#xff09;1. \u63a8\u7406\u5f15\u64ce\u89e3\u51b3\u4ec0\u4e48\u95ee\u98982. \u4e09\u79cd\u5f15\u64ce\u7684\u5173\u952e\u6267\u884c\u8def\u5f843. \u524d\u7f00\u7f13\u5b58\u4e0e KV Cache\u4e8c\u3001\u6a21\u578b\u90e8\u7f72&#xff1a;\u73af\u5883<\/p>\n","protected":false},"author":2,"featured_media":104709,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[11908,6994,11568,2681,2094],"topic":[],"class_list":["post-104710","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-pd-","tag-sglang","tag-11568","tag-vllm","tag-2094"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/104710.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM \u6587\u7ae0\u76ee\u5f55\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM\u4e00\u3001\u63a8\u7406\u5f15\u64ce&#xff1a;\u9009\u578b\u4e0e\u8fd0\u884c\u539f\u7406&#xff08;SGLang\u3001vLLM\u3001TensorRT-LLM&#xff09;1. \u63a8\u7406\u5f15\u64ce\u89e3\u51b3\u4ec0\u4e48\u95ee\u98982. \u4e09\u79cd\u5f15\u64ce\u7684\u5173\u952e\u6267\u884c\u8def\u5f843. \u524d\u7f00\u7f13\u5b58\u4e0e KV Cache\u4e8c\u3001\u6a21\u578b\u90e8\u7f72&#xff1a;\u73af\u5883\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/104710.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-12T18:57:09+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260912185707-6aa5a0836e6a6.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"11 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/104710.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/104710.html\",\"name\":\"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-12T18:57:09+00:00\",\"dateModified\":\"2026-09-12T18:57:09+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/104710.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/104710.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/104710.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/104710.html","og_locale":"zh_CN","og_type":"article","og_title":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM \u6587\u7ae0\u76ee\u5f55\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5&#xff1a;SGLang\u3001vLLM\u3001TensorRT-LLM\u4e00\u3001\u63a8\u7406\u5f15\u64ce&#xff1a;\u9009\u578b\u4e0e\u8fd0\u884c\u539f\u7406&#xff08;SGLang\u3001vLLM\u3001TensorRT-LLM&#xff09;1. \u63a8\u7406\u5f15\u64ce\u89e3\u51b3\u4ec0\u4e48\u95ee\u98982. \u4e09\u79cd\u5f15\u64ce\u7684\u5173\u952e\u6267\u884c\u8def\u5f843. \u524d\u7f00\u7f13\u5b58\u4e0e KV Cache\u4e8c\u3001\u6a21\u578b\u90e8\u7f72&#xff1a;\u73af\u5883","og_url":"https:\/\/www.wsisp.com\/helps\/104710.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-12T18:57:09+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260912185707-6aa5a0836e6a6.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"11 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/104710.html","url":"https:\/\/www.wsisp.com\/helps\/104710.html","name":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-12T18:57:09+00:00","dateModified":"2026-09-12T18:57:09+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/104710.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/104710.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/104710.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u3010\u63a8\u7406\u5f15\u64ce\u3011\u5927\u6a21\u578b\u63a8\u7406\u670d\u52a1\u90e8\u7f72\u3001\u6027\u80fd\u8c03\u4f18\u4e0e\u751f\u4ea7\u5b9e\u8df5\uff1aSGLang\u3001vLLM\u3001TensorRT-LLM"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/104710","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=104710"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/104710\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/104709"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=104710"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=104710"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=104710"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=104710"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}