{"id":31449,"date":"2026-07-04T17:01:10","date_gmt":"2026-07-04T09:01:10","guid":{"rendered":"https:\/\/cn.hostease.com\/blog\/uncategorized\/vllm-deployment-practical-gpu-vps\/"},"modified":"2026-07-04T17:01:10","modified_gmt":"2026-07-04T09:01:10","slug":"vllm-deployment-practical-gpu-vps","status":"publish","type":"post","link":"https:\/\/cn.hostease.com\/blog\/guides\/vllm-deployment-practical-gpu-vps\/","title":{"rendered":"vLLM \u9ad8\u6027\u80fd\u63a8\u7406\u90e8\u7f72\u5b9e\u6218\uff1aGPU VPS \u4e0a\u7684\u5927\u6a21\u578b\u52a0\u901f\u65b9\u6848"},"content":{"rendered":"<p>\u5982\u679c\u4f60\u6b63\u5728 GPU VPS\uff08\u865a\u62df\u4e13\u7528\u670d\u52a1\u5668\uff09\u4e0a\u8dd1\u5927\u6a21\u578b\u63a8\u7406\uff0c\u5927\u6982\u7387\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u56f0\u5883\uff1a\u6a21\u578b\u52a0\u8f7d\u5360\u6ee1\u663e\u5b58\u3001\u5e76\u53d1\u8bf7\u6c42\u4e00\u591a\u5c31\u6392\u961f\u3001\u541e\u5410\u91cf\u59cb\u7ec8\u4e0a\u4e0d\u53bb\u3002vLLM \u6b63\u662f\u4e3a\u89e3\u51b3\u8fd9\u7c7b\u95ee\u9898\u800c\u751f\u7684\u63a8\u7406\u5f15\u64ce\u2014\u2014\u5b83\u901a\u8fc7 PagedAttention\uff08\u5206\u9875\u6ce8\u610f\u529b\u673a\u5236\uff09\u548c\u8fde\u7eed\u6279\u5904\u7406\u6280\u672f\uff0c\u8ba9\u540c\u4e00\u5757 GPU\uff08\u56fe\u5f62\u5904\u7406\u5668\uff09\u7684\u63a8\u7406\u541e\u5410\u91cf\u63d0\u5347 2-4 \u500d\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n<p>\u672c\u6587\u5c06\u4ece\u5b9e\u9645\u90e8\u7f72\u89d2\u5ea6\u51fa\u53d1\uff0c\u5e26\u4f60\u5b8c\u6210\u4ece\u73af\u5883\u51c6\u5907\u3001vLLM \u5b89\u88c5\u3001\u6a21\u578b\u52a0\u8f7d\u5230\u6027\u80fd\u8c03\u4f18\u7684\u5168\u6d41\u7a0b\u3002\u65e0\u8bba\u4f60\u662f\u7b2c\u4e00\u6b21\u63a5\u89e6\u5927\u6a21\u578b\u63a8\u7406\uff0c\u8fd8\u662f\u5df2\u7ecf\u7528\u5176\u4ed6\u6846\u67b6\u8dd1\u8fc7\u6a21\u578b\uff0c\u90fd\u80fd\u5728\u8fd9\u91cc\u627e\u5230\u53ef\u76f4\u63a5\u590d\u7528\u7684\u914d\u7f6e\u65b9\u6848\u3002<\/p>\n<h2>vLLM \u662f\u4ec0\u4e48\uff1a\u4e3a\u4ec0\u4e48\u5b83\u80fd\u8ba9\u63a8\u7406\u5feb\u8fd9\u4e48\u591a<\/h2>\n<p>\u4f20\u7edf\u63a8\u7406\u6846\u67b6\u5728\u5904\u7406 KV Cache\uff08\u952e\u503c\u7f13\u5b58\uff09\u65f6\uff0c\u4f1a\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u9884\u5206\u914d\u4e00\u5757\u8fde\u7eed\u663e\u5b58\u3002\u95ee\u9898\u662f\uff0c\u4e0d\u540c\u8bf7\u6c42\u7684\u5e8f\u5217\u957f\u5ea6\u4e0d\u540c\uff0c\u9884\u5206\u914d\u7684\u7a7a\u95f4\u7ecf\u5e38\u6d6a\u8d39\u2014\u2014\u77ed\u5e8f\u5217\u5360\u7740\u5927\u5757\u663e\u5b58\u4e0d\u7528\uff0c\u957f\u5e8f\u5217\u53c8\u53ef\u80fd\u56e0\u4e3a\u663e\u5b58\u4e0d\u591f\u88ab\u62d2\u7edd\u3002vLLM \u7684\u6838\u5fc3\u521b\u65b0 PagedAttention \u501f\u9274\u4e86\u64cd\u4f5c\u7cfb\u7edf\u865a\u62df\u5185\u5b58\u7684\u601d\u8def\uff0c\u628a KV Cache \u5207\u5206\u6210\u56fa\u5b9a\u5927\u5c0f\u7684&#8221;\u9875&#8221;\uff0c\u6309\u9700\u5206\u914d\u3001\u52a8\u6001\u56de\u6536\u3002<\/p>\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u5e26\u6765\u7684\u76f4\u63a5\u597d\u5904\u6709\u4e09\u4e2a\uff1a<\/p>\n<ul>\n<li><strong>\u663e\u5b58\u5229\u7528\u7387\u5927\u5e45\u63d0\u5347<\/strong>\uff1a\u76f8\u540c GPU \u663e\u5b58\u4e0b\u53ef\u4ee5\u540c\u65f6\u670d\u52a1\u66f4\u591a\u8bf7\u6c42\u3002\u5b9e\u6d4b\u4e2d\uff0c7B \u53c2\u6570\u6a21\u578b\u5728 16GB \u663e\u5b58\u7684 GPU \u4e0a\uff0cvLLM \u76f8\u6bd4\u539f\u751f Hugging Face Transformers \u63a8\u7406\uff0c\u5e76\u53d1\u6570\u53ef\u4ee5\u4ece 4-5 \u4e2a\u63d0\u5347\u5230 15-20 \u4e2a\u3002<\/li>\n<li><strong>\u8fde\u7eed\u6279\u5904\u7406\uff08Continuous Batching\uff09<\/strong>\uff1a\u4f20\u7edf\u6279\u5904\u7406\u8981\u7b49\u4e00\u4e2a\u6279\u6b21\u5168\u90e8\u5b8c\u6210\u624d\u5904\u7406\u4e0b\u4e00\u6279\uff0cvLLM \u5219\u5728\u6709\u8bf7\u6c42\u5b8c\u6210\u65f6\u7acb\u5373\u63d2\u5165\u65b0\u8bf7\u6c42\uff0cGPU \u7a7a\u95f2\u65f6\u95f4\u5927\u5e45\u7f29\u77ed\u3002<\/li>\n<li><strong>\u5f20\u91cf\u5e76\u884c\uff08Tensor Parallelism\uff09\u652f\u6301<\/strong>\uff1a\u5982\u679c\u4f60\u7684 VPS \u914d\u4e86\u591a\u5757 GPU\uff0cvLLM \u53ef\u4ee5\u81ea\u52a8\u628a\u6a21\u578b\u5c42\u62c6\u5206\u5230\u4e0d\u540c\u5361\u4e0a\uff0c\u4e0d\u9700\u8981\u624b\u52a8\u914d\u7f6e\u590d\u6742\u7684\u5206\u5e03\u5f0f\u903b\u8f91\u3002<\/li>\n<\/ul>\n<p>\u7b80\u5355\u6765\u8bf4\uff0cvLLM \u8ba9\u4f60\u5728\u4e0d\u5347\u7ea7\u786c\u4ef6\u7684\u524d\u63d0\u4e0b\uff0c\u628a\u624b\u5934 GPU \u7684\u63a8\u7406\u80fd\u529b\u69a8\u5f97\u66f4\u5e72\u3002\u5bf9\u4e8e\u5728 <a href=\"https:\/\/cn.hostease.com\/vps\/\">VPS \u4e3b\u673a<\/a> \u4e0a\u505a\u63a8\u7406\u90e8\u7f72\u7684\u7528\u6237\u6765\u8bf4\uff0c\u8fd9\u610f\u5473\u7740\u540c\u6837\u7684\u6210\u672c\u80fd\u625b\u4f4f\u66f4\u591a\u5e76\u53d1\u3002<\/p>\n<h2>\u73af\u5883\u51c6\u5907\uff1aGPU VPS \u9700\u8981\u6ee1\u8db3\u54ea\u4e9b\u6761\u4ef6<\/h2>\n<p>\u5728\u5f00\u59cb\u5b89\u88c5 vLLM \u4e4b\u524d\uff0c\u5148\u786e\u8ba4\u4f60\u7684 VPS \u73af\u5883\u6ee1\u8db3\u4ee5\u4e0b\u57fa\u672c\u8981\u6c42\uff1a<\/p>\n<ul>\n<li><strong>\u64cd\u4f5c\u7cfb\u7edf<\/strong>\uff1aUbuntu 20.04 \u6216 22.04 LTS\uff0c\u63a8\u8350 22.04<\/li>\n<li><strong>GPU \u663e\u5b58<\/strong>\uff1a\u81f3\u5c11 16GB\uff08\u5982 NVIDIA T4 \u7ea7\u522b\uff09\uff1b\u8dd1 7B \u6a21\u578b\u5efa\u8bae 24GB\uff0c\u8dd1 13B \u6a21\u578b\u5efa\u8bae 40GB \u4ee5\u4e0a<\/li>\n<li><strong>CUDA \u7248\u672c<\/strong>\uff1a11.8 \u6216 12.1+\uff0c\u901a\u8fc7 <code>nvidia-smi<\/code> \u547d\u4ee4\u786e\u8ba4<\/li>\n<li><strong>Python \u7248\u672c<\/strong>\uff1a3.9 &#8211; 3.11<\/li>\n<li><strong>\u7cfb\u7edf\u5185\u5b58<\/strong>\uff1a\u81f3\u5c11 32GB\uff0c\u6a21\u578b\u52a0\u8f7d\u9636\u6bb5\u4f1a\u5360\u7528\u5927\u91cf RAM\uff08\u968f\u673a\u5b58\u53d6\u5b58\u50a8\u5668\uff09<\/li>\n<\/ul>\n<p>\u68c0\u67e5 GPU \u72b6\u6001\u7684\u547d\u4ee4\uff1a<\/p>\n<pre><code class=\"language-bash\">nvidia-smi\n<\/code><\/pre>\n<p>\u786e\u8ba4 CUDA \u7248\u672c\uff1a<\/p>\n<pre><code class=\"language-bash\">nvcc --version\n<\/code><\/pre>\n<p>\u5982\u679c\u4f60\u7684 VPS \u8fd8\u6ca1\u6709\u5b89\u88c5 NVIDIA \u9a71\u52a8\u548c CUDA\uff0c\u5efa\u8bae\u5148\u5b8c\u6210\u8fd9\u4e00\u6b65\u3002\u4ee5 Ubuntu 22.04 \u4e3a\u4f8b\uff1a<\/p>\n<pre><code class=\"language-bash\">sudo apt update\nsudo apt install -y nvidia-driver-535 nvidia-cuda-toolkit\nsudo reboot\n<\/code><\/pre>\n<p>\u91cd\u542f\u540e\u518d\u6b21\u8fd0\u884c <code>nvidia-smi<\/code>\uff0c\u80fd\u770b\u5230 GPU \u4fe1\u606f\u5c31\u8bf4\u660e\u9a71\u52a8\u5b89\u88c5\u6210\u529f\u3002\u5173\u4e8e <a href=\"https:\/\/cn.hostease.com\/blog\/server\/\">\u670d\u52a1\u5668\u914d\u7f6e<\/a> \u7684\u66f4\u591a\u7ec6\u8282\uff0c\u53ef\u4ee5\u53c2\u8003\u6211\u4eec\u4e4b\u524d\u7684\u6574\u7406\u3002<\/p>\n<h2>\u5b89\u88c5 vLLM \u4e0e\u542f\u52a8\u63a8\u7406\u670d\u52a1<\/h2>\n<p>vLLM \u7684\u5b89\u88c5\u975e\u5e38\u76f4\u63a5\uff0c\u63a8\u8350\u4f7f\u7528 pip \u5b89\u88c5\uff1a<\/p>\n<pre><code class=\"language-bash\">pip install vllm\n<\/code><\/pre>\n<p>\u5982\u679c\u4f60\u9700\u8981\u7279\u5b9a CUDA \u7248\u672c\u5bf9\u5e94\u7684 wheel \u5305\uff0c\u53ef\u4ee5\u4ece\u5b98\u65b9 GitHub Releases \u9875\u9762\u4e0b\u8f7d\u3002\u5b89\u88c5\u5b8c\u6210\u540e\uff0c\u7528\u4ee5\u4e0b\u547d\u4ee4\u9a8c\u8bc1\uff1a<\/p>\n<pre><code class=\"language-bash\">python -c &quot;import vllm; print(vllm.__version__)&quot;\n<\/code><\/pre>\n<p>\u63a5\u4e0b\u6765\u542f\u52a8\u4e00\u4e2a OpenAI \u517c\u5bb9\u7684 API \u670d\u52a1\u3002\u4ee5\u52a0\u8f7d Qwen2-7B \u6a21\u578b\u4e3a\u4f8b\uff1a<\/p>\n<pre><code class=\"language-bash\">python -m vllm.entrypoints.openai.api_server \\\n    --model Qwen\/Qwen2-7B-Instruct \\\n    --host 0.0.0.0 \\\n    --port 8000 \\\n    --gpu-memory-utilization 0.9 \\\n    --max-model-len 4096 \\\n    --tensor-parallel-size 1\n<\/code><\/pre>\n<p>\u51e0\u4e2a\u5173\u952e\u53c2\u6570\u89e3\u91ca\uff1a<\/p>\n<ul>\n<li><code>--gpu-memory-utilization 0.9<\/code>\uff1a\u544a\u8bc9 vLLM \u6700\u591a\u4f7f\u7528 90% \u7684 GPU \u663e\u5b58\uff0c\u7559\u4e00\u70b9\u4f59\u91cf\u9632\u6b62 OOM\uff08\u663e\u5b58\u6ea2\u51fa\uff09<\/li>\n<li><code>--max-model-len 4096<\/code>\uff1a\u9650\u5236\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 4096 tokens\uff0c\u503c\u8d8a\u5c0f\u663e\u5b58\u5360\u7528\u8d8a\u5c11<\/li>\n<li><code>--tensor-parallel-size 1<\/code>\uff1a\u5355 GPU \u8fd0\u884c\u8bbe\u4e3a 1\uff0c\u591a GPU \u65f6\u8bbe\u4e3a GPU \u6570\u91cf<\/li>\n<\/ul>\n<p>\u670d\u52a1\u542f\u52a8\u540e\uff0c\u53ef\u4ee5\u7528 curl \u6d4b\u8bd5\uff1a<\/p>\n<pre><code class=\"language-bash\">curl http:\/\/localhost:8000\/v1\/chat\/completions \\\n    -H &quot;Content-Type: application\/json&quot; \\\n    -d '{\n        &quot;model&quot;: &quot;Qwen\/Qwen2-7B-Instruct&quot;,\n        &quot;messages&quot;: [{&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: &quot;\u4f60\u597d\uff0c\u8bf7\u7b80\u5355\u4ecb\u7ecd\u4e00\u4e0b\u81ea\u5df1&quot;}],\n        &quot;max_tokens&quot;: 200\n    }'\n<\/code><\/pre>\n<p>\u5982\u679c\u8fd4\u56de\u6b63\u5e38\u7684 JSON \u54cd\u5e94\uff0c\u8bf4\u660e\u63a8\u7406\u670d\u52a1\u5df2\u7ecf\u8dd1\u8d77\u6765\u4e86\u3002<\/p>\n<h2>\u6027\u80fd\u8c03\u4f18\uff1a\u8ba9 GPU \u8dd1\u51fa\u6700\u5927\u541e\u5410\u91cf<\/h2>\n<p>\u90e8\u7f72\u53ea\u662f\u7b2c\u4e00\u6b65\uff0c\u771f\u6b63\u7684\u5dee\u8ddd\u5728\u4e8e\u8c03\u4f18\u3002\u4ee5\u4e0b\u662f\u7ecf\u8fc7\u5b9e\u6d4b\u9a8c\u8bc1\u7684\u51e0\u4e2a\u5173\u952e\u8c03\u4f18\u65b9\u5411\u3002<\/p>\n<h3>\u663e\u5b58\u5206\u914d\u7b56\u7565<\/h3>\n<p><code>gpu-memory-utilization<\/code> \u662f\u6700\u76f4\u63a5\u7684\u8c03\u4f18\u6760\u6746\u3002\u8bbe\u592a\u4f4e\u6d6a\u8d39 GPU\uff0c\u8bbe\u592a\u9ad8\u5bb9\u6613\u89e6\u53d1 OOM\u3002\u5b9e\u6d4b\u5efa\u8bae\uff1a<\/p>\n<ul>\n<li>\u7eaf\u63a8\u7406\u573a\u666f\uff1a0.85 &#8211; 0.92<\/li>\n<li>\u9700\u8981\u7559\u663e\u5b58\u7ed9\u5176\u4ed6\u8fdb\u7a0b\uff1a0.7 &#8211; 0.8<\/li>\n<li>\u6a21\u578b\u8f83\u5927\u63a5\u8fd1\u663e\u5b58\u4e0a\u9650\uff1a0.95\uff08\u914d\u5408 swap \u7a7a\u95f4\uff09<\/li>\n<\/ul>\n<p>\u53ef\u4ee5\u901a\u8fc7 <code>nvidia-smi -l 1<\/code> \u5b9e\u65f6\u89c2\u5bdf\u663e\u5b58\u4f7f\u7528\u60c5\u51b5\uff0c\u627e\u5230\u6700\u4f73\u5e73\u8861\u70b9\u3002<\/p>\n<h3>\u91cf\u5316\u90e8\u7f72\uff1a\u7528\u7cbe\u5ea6\u6362\u5bb9\u91cf<\/h3>\n<p>\u5982\u679c\u4f60\u7684 GPU \u663e\u5b58\u4e0d\u591f\u52a0\u8f7d\u5b8c\u6574\u7cbe\u5ea6\u6a21\u578b\uff0cvLLM \u652f\u6301\u591a\u79cd\u91cf\u5316\u65b9\u6848\uff1a<\/p>\n<ul>\n<li><strong>AWQ \u91cf\u5316<\/strong>\uff1a4-bit \u91cf\u5316\uff0c\u663e\u5b58\u5360\u7528\u7ea6\u4e3a FP16 \u7684 1\/3\uff0c\u7cbe\u5ea6\u635f\u5931\u8f83\u5c0f<\/li>\n<li><strong>GPTQ \u91cf\u5316<\/strong>\uff1a\u540c\u6837\u662f 4-bit\uff0c\u9700\u8981\u9884\u5148\u91cf\u5316\u597d\u7684\u6a21\u578b\u6587\u4ef6<\/li>\n<li><strong>FP8 \u91cf\u5316<\/strong>\uff1a\u5728\u652f\u6301 FP8 \u7684 GPU \u4e0a\u53ef\u7528\uff0c\u7cbe\u5ea6\u635f\u5931\u6700\u5c0f<\/li>\n<\/ul>\n<p>\u4ee5 AWQ \u91cf\u5316\u4e3a\u4f8b\uff0c\u52a0\u8f7d\u547d\u4ee4\u53ea\u9700\u52a0\u4e00\u4e2a\u53c2\u6570\uff1a<\/p>\n<pre><code class=\"language-bash\">python -m vllm.entrypoints.openai.api_server \\\n    --model Qwen\/Qwen2-7B-Instruct-AWQ \\\n    --quantization awq \\\n    --gpu-memory-utilization 0.9\n<\/code><\/pre>\n<p>\u91cf\u5316\u540e\uff0c\u539f\u672c\u9700\u8981 14GB \u663e\u5b58\u7684 7B \u6a21\u578b\uff0c\u5927\u7ea6\u53ea\u9700\u8981 5-6GB\uff0c\u4e00\u5f20 16GB \u663e\u5b58\u7684\u5361\u5c31\u80fd\u8f7b\u677e\u8dd1\u8d77\u6765\u3002<\/p>\n<h3>\u6279\u5904\u7406\u53c2\u6570\u4f18\u5316<\/h3>\n<p>vLLM \u7684\u8fde\u7eed\u6279\u5904\u7406\u867d\u7136\u81ea\u52a8\u8fd0\u884c\uff0c\u4f46\u6709\u51e0\u4e2a\u53c2\u6570\u53ef\u4ee5\u8fdb\u4e00\u6b65\u8c03\u4f18\uff1a<\/p>\n<ul>\n<li><code>--max-num-seqs<\/code>\uff1a\u5355\u6b65\u6700\u5927\u5e76\u53d1\u5e8f\u5217\u6570\uff0c\u9ed8\u8ba4 256\u3002\u5982\u679c\u8bf7\u6c42\u5ef6\u8fdf\u654f\u611f\uff0c\u53ef\u4ee5\u964d\u4f4e\u5230 64-128<\/li>\n<li><code>--max-num-batched-tokens<\/code>\uff1a\u5355\u6b65\u6700\u5927 token \u6570\uff0c\u589e\u5927\u53ef\u4ee5\u63d0\u5347\u541e\u5410\u4f46\u589e\u52a0\u5ef6\u8fdf<\/li>\n<li><code>--swap-space<\/code>\uff1aCPU swap \u7a7a\u95f4\u5927\u5c0f\uff08GB\uff09\uff0c\u5f53 GPU \u663e\u5b58\u4e0d\u591f\u65f6\u7528\u4e8e\u4e34\u65f6\u5b58\u653e KV Cache<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e\u5ef6\u8fdf\u654f\u611f\u7684\u5728\u7ebf\u670d\u52a1\uff0c\u5efa\u8bae <code>--max-num-seqs<\/code> \u8bbe\u4e3a 32-64\uff1b\u5bf9\u4e8e\u6279\u91cf\u79bb\u7ebf\u63a8\u7406\uff0c\u8bbe\u4e3a 256 \u6216\u66f4\u9ad8\u3002<\/p>\n<p>\u60f3\u4e86\u89e3\u66f4\u591a <a href=\"https:\/\/cn.hostease.com\/blog\/optimization\/\">\u7f51\u7ad9\u6027\u80fd\u4f18\u5316<\/a> \u7684\u601d\u8def\uff0c\u63a8\u7406\u670d\u52a1\u548c Web \u670d\u52a1\u5728\u8d44\u6e90\u8c03\u5ea6\u4e0a\u6709\u4e0d\u5c11\u5171\u901a\u4e4b\u5904\u3002<\/p>\n<h2>\u751f\u4ea7\u73af\u5883\u6ce8\u610f\u4e8b\u9879<\/h2>\n<p>\u628a vLLM \u4ece&#8221;\u80fd\u8dd1&#8221;\u53d8\u6210&#8221;\u80fd\u7528&#8221;\uff0c\u8fd8\u9700\u8981\u8003\u8651\u51e0\u4e2a\u751f\u4ea7\u73af\u5883\u7684\u5173\u952e\u95ee\u9898\u3002<\/p>\n<h3>\u53cd\u5411\u4ee3\u7406\u4e0e\u8d1f\u8f7d\u5747\u8861<\/h3>\n<p>vLLM \u81ea\u5e26\u7684 API \u670d\u52a1\u662f\u5355\u8fdb\u7a0b\u7684\uff0c\u751f\u4ea7\u73af\u5883\u5efa\u8bae\u5728\u524d\u9762\u52a0\u4e00\u5c42 Nginx\uff08\u53cd\u5411\u4ee3\u7406\u670d\u52a1\u5668\uff09\u505a\u8d1f\u8f7d\u5747\u8861\u548c SSL\uff08\u5b89\u5168\u4f20\u8f93\u534f\u8bae\uff09\u7ec8\u6b62\uff1a<\/p>\n<pre><code class=\"language-nginx\">upstream vllm_backend {\n    server 127.0.0.1:8000;\n    server 127.0.0.1:8001;\n}\n\nserver {\n    listen 443 ssl;\n    server_name your-domain.com;\n\n    ssl_certificate \/path\/to\/cert.pem;\n    ssl_certificate_key \/path\/to\/key.pem;\n\n    location \/v1\/ {\n        proxy_pass http:\/\/vllm_backend;\n        proxy_set_header Host $host;\n        proxy_read_timeout 300s;\n    }\n}\n<\/code><\/pre>\n<p>\u6ce8\u610f <code>proxy_read_timeout<\/code> \u8981\u8bbe\u5f97\u8db3\u591f\u957f\uff0c\u5927\u6a21\u578b\u63a8\u7406\u7684\u54cd\u5e94\u65f6\u95f4\u53ef\u80fd\u5728 10-30 \u79d2\u751a\u81f3\u66f4\u4e45\u3002<\/p>\n<h3>\u76d1\u63a7\u4e0e\u544a\u8b66<\/h3>\n<p>\u63a8\u7406\u670d\u52a1\u4e0a\u7ebf\u540e\uff0c\u5efa\u8bae\u76d1\u63a7\u4ee5\u4e0b\u6307\u6807\uff1a<\/p>\n<ul>\n<li><strong>GPU \u663e\u5b58\u4f7f\u7528\u7387<\/strong>\uff1a\u6301\u7eed\u8d85\u8fc7 95% \u8bf4\u660e\u9700\u8981\u6269\u5bb9\u6216\u964d\u4f4e\u5e76\u53d1<\/li>\n<li><strong>\u8bf7\u6c42\u961f\u5217\u6df1\u5ea6<\/strong>\uff1a\u901a\u8fc7 vLLM \u7684 <code>\/metrics<\/code> \u7aef\u70b9\u83b7\u53d6 Prometheus \u6307\u6807<\/li>\n<li><strong>P99 \u5ef6\u8fdf<\/strong>\uff1a\u8d85\u8fc7\u4e1a\u52a1 SLA\uff08\u670d\u52a1\u7b49\u7ea7\u534f\u8bae\uff09\u9608\u503c\u65f6\u89e6\u53d1\u544a\u8b66<\/li>\n<li><strong>\u9519\u8bef\u7387<\/strong>\uff1aOOM \u9519\u8bef\u610f\u5473\u7740\u663e\u5b58\u4e0d\u8db3\uff0c\u9700\u8981\u8c03\u6574\u53c2\u6570\u6216\u5347\u7ea7\u786c\u4ef6<\/li>\n<\/ul>\n<p>vLLM \u539f\u751f\u652f\u6301 Prometheus \u6307\u6807\u66b4\u9732\uff0c\u914d\u5408 Grafana \u53ef\u4ee5\u5feb\u901f\u642d\u5efa\u76d1\u63a7\u9762\u677f\u3002<\/p>\n<h3>\u591a\u5b9e\u4f8b\u90e8\u7f72<\/h3>\n<p>\u5982\u679c\u4f60\u7684\u4e1a\u52a1\u9700\u8981\u9ad8\u53ef\u7528\uff0c\u53ef\u4ee5\u5728\u540c\u4e00\u53f0 VPS \u4e0a\u542f\u52a8\u591a\u4e2a vLLM \u5b9e\u4f8b\uff0c\u901a\u8fc7 Nginx \u8f6e\u8be2\u5206\u53d1\u8bf7\u6c42\u3002\u6bcf\u4e2a\u5b9e\u4f8b\u7ed1\u5b9a\u4e0d\u540c\u7684\u7aef\u53e3\uff0c\u4f7f\u7528\u540c\u4e00\u4efd\u6a21\u578b\u6587\u4ef6\uff08\u6a21\u578b\u6743\u91cd\u4f1a\u88ab\u64cd\u4f5c\u7cfb\u7edf\u81ea\u52a8\u5171\u4eab\u5185\u5b58\u9875\uff0c\u4e0d\u4f1a\u91cd\u590d\u5360\u7528\u7269\u7406\u5185\u5b58\uff09\u3002<\/p>\n<p>\u5bf9\u4e8e\u9700\u8981\u66f4\u9ad8\u7b97\u529b\u7684\u573a\u666f\uff0c\u53ef\u4ee5\u8003\u8651\u5728 <a href=\"https:\/\/cn.hostease.com\/dedicated-server\/\">\u72ec\u7acb\u670d\u52a1\u5668<\/a> \u4e0a\u90e8\u7f72\uff0c\u72ec\u4eab\u6574\u673a\u8d44\u6e90\u4e0d\u4f1a\u53d7\u5230\u90bb\u5c45\u5e72\u6270\u3002<\/p>\n<h2>\u603b\u7ed3\u4e0e\u4e0b\u4e00\u6b65\u5efa\u8bae<\/h2>\n<p>vLLM \u7684\u90e8\u7f72\u6d41\u7a0b\u5e76\u4e0d\u590d\u6742\uff0c\u6838\u5fc3\u6b65\u9aa4\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a\u786e\u8ba4 GPU \u73af\u5883 \u2192 \u5b89\u88c5 vLLM \u2192 \u542f\u52a8 API \u670d\u52a1 \u2192 \u8c03\u4f18\u53c2\u6570 \u2192 \u751f\u4ea7\u52a0\u56fa\u3002\u771f\u6b63\u8ba9\u5b83\u8131\u9896\u800c\u51fa\u7684\u662f PagedAttention \u548c\u8fde\u7eed\u6279\u5904\u7406\u5e26\u6765\u7684\u663e\u5b58\u6548\u7387\u548c\u541e\u5410\u91cf\u63d0\u5347\u2014\u2014\u540c\u6837\u7684\u786c\u4ef6\uff0c\u80fd\u670d\u52a1\u7684\u5e76\u53d1\u8bf7\u6c42\u6570\u91cf\u53ef\u4ee5\u7ffb\u500d\u3002<\/p>\n<p>\u5982\u679c\u4f60\u6b63\u51c6\u5907\u5728 GPU VPS \u4e0a\u642d\u5efa\u63a8\u7406\u670d\u52a1\uff0c\u5efa\u8bae\u4ece\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u5411\u5f00\u59cb\uff1a<\/p>\n<ol>\n<li><strong>\u5148\u7528\u5c0f\u6a21\u578b\u9a8c\u8bc1<\/strong>\uff1a\u9009\u4e00\u4e2a 7B \u53c2\u6570\u7684\u6a21\u578b\u8dd1\u901a\u5168\u6d41\u7a0b\uff0c\u786e\u8ba4\u73af\u5883\u6ca1\u95ee\u9898\u540e\u518d\u52a0\u8f7d\u66f4\u5927\u7684\u6a21\u578b<\/li>\n<li><strong>\u91cf\u5316\u662f\u6027\u4ef7\u6bd4\u6700\u9ad8\u7684\u4f18\u5316<\/strong>\uff1aAWQ 4-bit \u91cf\u5316\u5728\u7cbe\u5ea6\u635f\u5931\u53ef\u63a7\u7684\u524d\u63d0\u4e0b\uff0c\u80fd\u628a\u663e\u5b58\u9700\u6c42\u964d\u5230\u539f\u6765\u7684 1\/3<\/li>\n<li><strong>\u76d1\u63a7\u5148\u884c<\/strong>\uff1a\u5728\u670d\u52a1\u4e0a\u7ebf\u524d\u5c31\u914d\u597d GPU \u663e\u5b58\u548c\u8bf7\u6c42\u5ef6\u8fdf\u7684\u76d1\u63a7\uff0c\u95ee\u9898\u51fa\u73b0\u65f6\u624d\u80fd\u5feb\u901f\u5b9a\u4f4d<\/li>\n<li><strong>\u6839\u636e\u4e1a\u52a1\u573a\u666f\u9009\u62e9\u786c\u4ef6<\/strong>\uff1a\u5ef6\u8fdf\u654f\u611f\u7684\u5728\u7ebf\u670d\u52a1\u9009\u9ad8\u4e3b\u9891 GPU\uff0c\u6279\u91cf\u79bb\u7ebf\u63a8\u7406\u9009\u5927\u663e\u5b58 GPU<\/li>\n<\/ol>\n<p>\u5728 Hostease \u7684 GPU VPS \u4e0a\u90e8\u7f72 vLLM\uff0c\u53ef\u4ee5\u83b7\u5f97\u7075\u6d3b\u7684\u7b97\u529b\u914d\u7f6e\u548c\u7a33\u5b9a\u7684\u7f51\u7edc\u73af\u5883\uff0c\u914d\u5408\u672c\u6587\u7684\u8c03\u4f18\u65b9\u6848\uff0c\u8db3\u4ee5\u652f\u6491\u4e2d\u5c0f\u89c4\u6a21\u7684\u5927\u6a21\u578b\u63a8\u7406\u4e1a\u52a1\u3002\u5982\u679c\u4f60\u5728\u90e8\u7f72\u8fc7\u7a0b\u4e2d\u9047\u5230\u5177\u4f53\u95ee\u9898\uff0c\u6b22\u8fce\u5728\u8bc4\u8bba\u533a\u4ea4\u6d41\u8ba8\u8bba\u3002<\/p>\n<p>\u603b\u7ed3\u6765\u8bf4\uff0cvLLM \u662f\u76ee\u524d GPU VPS \u4e0a\u505a\u5927\u6a21\u578b\u63a8\u7406\u6700\u5177\u6027\u4ef7\u6bd4\u7684\u5f00\u6e90\u65b9\u6848\u4e4b\u4e00\u3002\u5efa\u8bae\u4ece 7B \u6a21\u578b + AWQ \u91cf\u5316\u7ec4\u5408\u5f00\u59cb\u9a8c\u8bc1\uff0c\u786e\u8ba4\u4e1a\u52a1\u6d41\u7a0b\u8dd1\u901a\u540e\u518d\u9010\u6b65\u6269\u5c55\u5230\u66f4\u5927\u6a21\u578b\u548c\u591a GPU \u5e76\u884c\u90e8\u7f72\u3002\u5982\u679c\u4f60\u9700\u8981\u66f4\u4e13\u4e1a\u7684 GPU \u7b97\u529b\u652f\u6301\uff0c\u53ef\u4ee5\u8003\u8651\u9009\u62e9\u914d\u5907\u9ad8\u6027\u80fd\u663e\u5361\u7684\u72ec\u7acb\u670d\u52a1\u5668\u65b9\u6848\uff0c\u83b7\u5f97\u66f4\u7a33\u5b9a\u7684\u63a8\u7406\u6027\u80fd\u4fdd\u969c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5982\u679c\u4f60\u6b63\u5728 GPU VPS\uff08\u865a\u62df\u4e13\u7528\u670d\u52a1\u5668\uff09\u4e0a\u8dd1\u5927\u6a21\u578b\u63a8\u7406\uff0c\u5927\u6982\u7387\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u56f0\u5883\uff1a\u6a21\u578b\u52a0\u8f7d\u5360\u6ee1\u663e\u5b58\u3001\u5e76\u53d1\u8bf7\u6c42 &#8230; <a title=\"vLLM \u9ad8\u6027\u80fd\u63a8\u7406\u90e8\u7f72\u5b9e\u6218\uff1aGPU VPS \u4e0a\u7684\u5927\u6a21\u578b\u52a0\u901f\u65b9\u6848\" class=\"read-more\" href=\"https:\/\/cn.hostease.com\/blog\/guides\/vllm-deployment-practical-gpu-vps\/\" aria-label=\"\u9605\u8bfb vLLM \u9ad8\u6027\u80fd\u63a8\u7406\u90e8\u7f72\u5b9e\u6218\uff1aGPU VPS \u4e0a\u7684\u5927\u6a21\u578b\u52a0\u901f\u65b9\u6848\">\u9605\u8bfb\u66f4\u591a<\/a><\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[3822],"tags":[4491,4490,4492,4493],"class_list":["post-31449","post","type-post","status-publish","format-standard","hentry","category-guides","tag-gpu","tag-vllm-2","tag-4492","tag-4493"],"aioseo_notices":[],"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/posts\/31449","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/comments?post=31449"}],"version-history":[{"count":0,"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/posts\/31449\/revisions"}],"wp:attachment":[{"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/media?parent=31449"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/categories?post=31449"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cn.hostease.com\/blog\/wp-json\/wp\/v2\/tags?post=31449"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}