{"id":132,"date":"2026-08-29T15:49:49","date_gmt":"2026-08-29T07:49:49","guid":{"rendered":"https:\/\/www.xiaoxindong.cloud\/?p=132"},"modified":"2026-08-29T15:49:49","modified_gmt":"2026-08-29T07:49:49","slug":"gpu-%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2vllm%e3%80%81open-webui","status":"publish","type":"post","link":"https:\/\/www.xiaoxindong.cloud\/?p=132","title":{"rendered":"GPU \u5927\u6a21\u578b\u672c\u5730\u90e8\u7f72(vllm\u3001open webUI)"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u4e00\u3001vLLM \u4f01\u4e1a\u7ea7\u63a8\u7406\u6846\u67b6<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u6846\u67b6\u5b9a\u4f4d<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vLLM\uff08virtual Large Language Model\uff09\u662f\u9762\u5411\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u7684\u9ad8\u6027\u80fd\u670d\u52a1\u6846\u67b6\uff0c\u6838\u5fc3\u89e3\u51b3\u4f20\u7edf Transformers \u90e8\u7f72\u7684\u9ad8\u5ef6\u8fdf\u3001\u4f4e\u541e\u5410\u3001\u663e\u5b58\u6d6a\u8d39\u95ee\u9898\uff0c\u662f\u751f\u4ea7\u7ea7\u63a8\u7406\u670d\u52a1\u7684\u4e3b\u6d41\u65b9\u6848\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6838\u5fc3\u80fd\u529b\uff1a\u6a21\u578b\u52a0\u8f7d\u3001GPU \u8d44\u6e90\u8c03\u5ea6\u3001KV Cache \u7ba1\u7406\u3001API \u670d\u52a1<\/li>\n\n\n\n<li>\u6838\u5fc3\u6280\u672f\uff1a<strong>PagedAttention + Continuous Batching<\/strong><\/li>\n\n\n\n<li>\u517c\u5bb9\u6027\uff1aHuggingFace \u6a21\u578b\u65e0\u7f1d\u63a5\u5165\u3001OpenAI API \u6807\u51c6\u63a5\u53e3\u3001\u591a\u54c1\u724c GPU\/TPU<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u4e24\u5927\u6838\u5fc3\u6280\u672f<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">2.1 PagedAttention\uff08\u5206\u9875\u6ce8\u610f\u529b\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8bbe\u8ba1\u601d\u60f3<\/strong>\uff1a\u501f\u9274\u64cd\u4f5c\u7cfb\u7edf\u865a\u62df\u5185\u5b58\u5206\u9875\u673a\u5236\uff0c\u5c06 KV Cache \u5207\u5206\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684 Block\uff08\u9875\uff09\uff0c\u901a\u8fc7\u9875\u8868\u6620\u5c04\u7ba1\u7406\uff0c\u7269\u7406\u4e0a\u65e0\u9700\u8fde\u7eed\uff0c\u6309\u9700\u5206\u914d\u3001\u7528\u5b8c\u5373\u56de\u6536\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u51b3\u75db\u70b9<\/strong>\uff1a\u4f20\u7edf KV Cache \u6309\u6700\u5927\u4e0a\u4e0b\u6587\u9884\u7559\u6574\u5757\u8fde\u7eed\u663e\u5b58\uff0c\u5b9e\u9645\u4f7f\u7528\u7387\u4f4e\uff0c\u663e\u5b58\u788e\u7247\u5316\u4e25\u91cd\uff0c\u5bfc\u81f4\u5e76\u53d1\u6570\u4e0a\u4e0d\u53bb\u3002<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>\uff1a\u663e\u5b58\u5229\u7528\u7387\u63d0\u5347\u81f3 96% \u4ee5\u4e0a\uff0c\u652f\u6301\u66f4\u591a\u5e76\u53d1\u8bf7\u6c42\uff0c\u541e\u5410\u91cf\u6700\u9ad8\u8fbe\u4f20\u7edf\u6846\u67b6\u7684 24 \u500d\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2.2 Continuous Batching\uff08\u8fde\u7eed\u6279\u5904\u7406\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8bbe\u8ba1\u601d\u60f3<\/strong>\uff1a\u65b0\u8bf7\u6c42\u52a8\u6001\u63d2\u5165\u6b63\u5728\u6267\u884c\u7684\u6279\u6b21\uff0c\u65e0\u9700\u7b49\u5f85\u6574\u6279\u5904\u7406\u5b8c\u6210\uff1b\u8c01\u5148\u751f\u6210\u7ed3\u675f\uff0c\u7acb\u523b\u91ca\u653e\u4f4d\u7f6e\u8865\u5165\u65b0\u8bf7\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u51b3\u75db\u70b9<\/strong>\uff1a\u4f20\u7edf\u9759\u6001\u6279\u5904\u7406\u5fc5\u987b\u7b49\u6700\u957f\u8bf7\u6c42\u5b8c\u6210\u624d\u80fd\u5904\u7406\u4e0b\u4e00\u6279\uff0cGPU \u7a7a\u95f2\u65f6\u95f4\u591a\uff0c\u77ed\u8bf7\u6c42\u88ab\u957f\u8bf7\u6c42\u963b\u585e\u3002<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u4f18\u52bf<\/strong>\uff1aGPU \u59cb\u7ec8\u9ad8\u8d1f\u8f7d\u8fd0\u884c\uff0c\u8bf7\u6c42\u7b49\u5f85\u65f6\u95f4\u77ed\uff0c\u541e\u5410\u91cf\u5927\u5e45\u63d0\u5347\uff0c\u9ad8\u5e76\u53d1\u573a\u666f\u4f18\u52bf\u663e\u8457\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. vLLM \u89e3\u51b3\u7684\u6838\u5fc3\u95ee\u9898<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8868\u683c<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u4f20\u7edf\u90e8\u7f72\u75db\u70b9<\/th><th>vLLM \u89e3\u51b3\u65b9\u6848<\/th><\/tr><\/thead><tbody><tr><td>\u663e\u5b58\u788e\u7247\u4e25\u91cd\u3001\u6d6a\u8d39\u5927<\/td><td>PagedAttention \u5206\u5757\u7ba1\u7406 KV \u7f13\u5b58\uff0c\u663e\u5b58\u5229\u7528\u7387\u63a5\u8fd1\u6700\u4f18<\/td><\/tr><tr><td>\u63a8\u7406\u541e\u5410\u91cf\u4f4e\u3001GPU \u7a7a\u8f6c<\/td><td>\u8fde\u7eed\u6279\u5904\u7406 + \u4f18\u5316 CUDA \u5185\u6838\uff0c\u5927\u5e45\u63d0\u5347\u63a8\u7406\u901f\u5ea6<\/td><\/tr><tr><td>\u591a\u4efb\u52a1\u8c03\u5ea6\u6df7\u4e71<\/td><td>\u5148\u8fdb\u8c03\u5ea6\u7b56\u7565\uff08\u5148\u6765\u5148\u670d\u52a1 + \u62a2\u5360\u673a\u5236\uff09\uff0c\u8d44\u6e90\u5408\u7406\u5206\u914d<\/td><\/tr><tr><td>\u591a GPU \u90e8\u7f72\u590d\u6742<\/td><td>\u539f\u751f\u652f\u6301\u5f20\u91cf\u5e76\u884c\u3001\u6d41\u6c34\u7ebf\u5e76\u884c\uff0c\u7b80\u5316\u591a\u8bbe\u5907\u90e8\u7f72<\/td><\/tr><tr><td>\u4e1a\u52a1\u96c6\u6210\u96be\u5ea6\u9ad8<\/td><td>\u517c\u5bb9\u4e3b\u6d41\u6a21\u578b\u4e0e OpenAI API\uff0c\u964d\u4f4e\u63a5\u5165\u95e8\u69db<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u4e3b\u8981\u5e94\u7528\u573a\u666f<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5b9e\u65f6\u63a8\u7406\u670d\u52a1\uff1a\u667a\u80fd\u5ba2\u670d\u3001\u804a\u5929\u673a\u5668\u4eba\u7b49\u9ad8\u5e76\u53d1\u4f4e\u5ef6\u8fdf\u573a\u666f<\/li>\n\n\n\n<li>\u957f\u6587\u672c\u751f\u6210\uff1a\u6587\u7ae0\u5199\u4f5c\u3001\u6458\u8981\u751f\u6210\u3001\u521b\u610f\u5185\u5bb9\u521b\u4f5c<\/li>\n\n\n\n<li>\u4f01\u4e1a\u5185\u90e8 AI \u52a9\u624b\uff1a\u7a33\u5b9a\u627f\u63a5\u591a\u7528\u6237\u5e76\u53d1\u8bf7\u6c42<\/li>\n\n\n\n<li>RAG \u4e0e\u77e5\u8bc6\u5e93\u7cfb\u7edf\uff1a\u9ad8\u5e76\u53d1\u4e0b\u6027\u80fd\u4f18\u52bf\u660e\u663e<\/li>\n\n\n\n<li>\u4e1a\u52a1\u96c6\u6210\uff1a\u4f5c\u4e3a\u63a8\u7406\u540e\u7aef\u5bf9\u63a5\u524d\u7aef\u5e94\u7528\uff0c\u652f\u6301\u5206\u5e03\u5f0f\u6269\u5c55<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001vLLM \u90e8\u7f72\u5b9e\u6218<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u73af\u5883\u51c6\u5907<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">1.1 \u5305\u7ba1\u7406\u5668 uv \u5b89\u88c5<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">uv \u662f Python \u9ad8\u6027\u80fd\u5305\u4e0e\u865a\u62df\u73af\u5883\u7ba1\u7406\u5668\uff0c\u901f\u5ea6\u8fdc\u5feb\u4e8e pip\uff0c\u7528\u4e8e\u9694\u79bb\u73af\u5883\u548c\u5feb\u901f\u5b89\u88c5\u4f9d\u8d56\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># \u65b9\u5f0f1\uff1a\u5b98\u65b9\u811a\u672c\u5b89\u88c5\ncurl -LsSf https:\/\/astral.sh\/uv\/install.sh | sh\n\n# \u65b9\u5f0f2\uff1apip\u5b89\u88c5\npip install uv\n<\/code><\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9a8c\u8bc1\u547d\u4ee4\uff1a<code>uv --version<\/code><\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">1.2 \u521b\u5efa\u865a\u62df\u73af\u5883<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u8981\u6c42 Python \u7248\u672c\u22653.10<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>mkdir \/root\/uv-vllm-qwen\ncd \/root\/uv-vllm-qwen\nuv venv --python 3.10\nsource .venv\/bin\/activate  # \u6fc0\u6d3b\u540e\u7ec8\u7aef\u524d\u7f00\u663e\u793a (.venv)\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u5b89\u88c5 vLLM<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code># \u57fa\u7840\u5b89\u88c5\nuv pip install \"vllm==0.7.2\" --torch-backend=auto\n\n# \u56fd\u5185\u52a0\u901f\uff1a\u4f7f\u7528\u6e05\u534e\u6e90\nUV_INDEX_URL=https:\/\/pypi.tuna.tsinghua.edu.cn\/simple uv pip install \"vllm==0.7.2\" --torch-backend=auto\n<\/code><\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5b89\u88c5\u9a8c\u8bc1\uff1a<code>python -c \"import vllm; print(vllm.__version__)\"<\/code><\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u56fd\u5185\u6a21\u578b\u4e0b\u8f7d\u4f18\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vLLM \u9ed8\u8ba4\u4ece HuggingFace \u4e0b\u8f7d\uff0c\u56fd\u5185\u7f51\u7edc\u53d7\u9650\uff0c\u5207\u6362\u4e3a ModelScope\uff08\u9b54\u642d\uff09\u52a0\u901f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># 1. \u5b89\u88c5modelscope\u5de5\u5177\nuv pip install modelscope==1.25.0\n\n# 2. \u5f00\u542f\u73af\u5883\u53d8\u91cf\uff0c\u4f18\u5148\u4eceModelScope\u62c9\u53d6\u6a21\u578b\nexport VLLM_USE_MODELSCOPE=True\n\n# 3. \u5347\u7ea7\u914d\u5957\u4f9d\u8d56\u5e93\nuv pip install \"transformers==4.45.2\" \"tokenizers==0.20.1\"\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u542f\u52a8\u6a21\u578b\u670d\u52a1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u7840\u542f\u52a8\u547d\u4ee4\u793a\u4f8b\uff08Qwen2-1.5B-Instruct\uff09\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>vllm serve Qwen\/Qwen2-1.5B-Instruct \\\n  --max-model-len 4096 \\\n  --gpu-memory-utilization 0.9 \\\n  --tensor-parallel-size 1 \\\n  --max-num-seqs 16 \\\n  --dtype half\n<\/code><\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9ed8\u8ba4\u76d1\u542c\u5730\u5740\uff1a<code>http:\/\/localhost:8000<\/code><\/li>\n\n\n\n<li>\u9996\u6b21\u8fd0\u884c\u81ea\u52a8\u4e0b\u8f7d\u6a21\u578b\uff0c\u4e0b\u8f7d\u5b8c\u6210\u540e\u542f\u52a8\u63a8\u7406\u670d\u52a1<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">5. \u6838\u5fc3\u53c2\u6570\u8be6\u89e3<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">5.1 \u57fa\u7840\u670d\u52a1\u53c2\u6570<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u8868\u683c<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u53c2\u6570<\/th><th>\u8bf4\u660e<\/th><th>\u9ed8\u8ba4\u503c<\/th><\/tr><\/thead><tbody><tr><td><code>--host<\/code><\/td><td>\u670d\u52a1\u7ed1\u5b9a\u7684 IP \u5730\u5740<\/td><td>0.0.0.0<\/td><\/tr><tr><td><code>--port<\/code><\/td><td>\u670d\u52a1\u76d1\u542c\u7aef\u53e3<\/td><td>8000<\/td><\/tr><tr><td><code>--served-model-name<\/code><\/td><td>API \u5bf9\u5916\u66b4\u9732\u7684\u6a21\u578b\u540d\u79f0<\/td><td>\u4e0e\u539f\u6a21\u578b\u540d\u4e00\u81f4<\/td><\/tr><tr><td><code>--api-key<\/code><\/td><td>API \u8bbf\u95ee\u9274\u6743\u5bc6\u94a5<\/td><td>\u65e0<\/td><\/tr><tr><td><code>--trust-remote-code<\/code><\/td><td>\u5141\u8bb8\u6267\u884c\u6a21\u578b\u4ed3\u5e93\u81ea\u5b9a\u4e49\u4ee3\u7801<\/td><td>\u5173\u95ed<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">5.2 \u6027\u80fd\u4e0e\u663e\u5b58\u8c03\u4f18\u53c2\u6570<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u8868\u683c<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u53c2\u6570<\/th><th>\u8bf4\u660e<\/th><th>\u8c03\u4f18\u5efa\u8bae<\/th><\/tr><\/thead><tbody><tr><td><code>--gpu-memory-utilization<\/code><\/td><td>GPU \u663e\u5b58\u6700\u5927\u4f7f\u7528\u6bd4\u4f8b<\/td><td>\u9ed8\u8ba4 0.9\uff08\u4f7f\u7528 90% \u663e\u5b58\uff0c\u9884\u7559 10% \u7ed9\u7cfb\u7edf\u8fdb\u7a0b\uff09<\/td><\/tr><tr><td><code>--max-model-len<\/code><\/td><td>\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\uff08\u8f93\u5165 + \u8f93\u51fa Token \u603b\u6570\uff09<\/td><td>\u6839\u636e\u4e1a\u52a1\u573a\u666f\u8bbe\u7f6e\uff0c\u4e0d\u5fc5\u76f2\u76ee\u8ffd\u5927\uff0c\u8282\u7701\u663e\u5b58\u63d0\u5347\u5e76\u53d1<\/td><\/tr><tr><td><code>--max-num-seqs<\/code><\/td><td>\u540c\u65f6\u5904\u7406\u7684\u6700\u5927\u8bf7\u6c42\u6570\uff08\u5e76\u53d1\u5e8f\u5217\u6570\uff09<\/td><td>\u76f4\u63a5\u51b3\u5b9a\u5e76\u53d1\u80fd\u529b\uff0c\u8fc7\u5927\u6613\u663e\u5b58\u6ea2\u51fa\uff0c\u8fc7\u5c0f\u6d6a\u8d39 GPU \u8d44\u6e90<\/td><\/tr><tr><td><code>--dtype<\/code><\/td><td>\u6a21\u578b\u63a8\u7406\u6570\u636e\u7cbe\u5ea6<\/td><td>FP16 (half)\/BF16\/INT8\/INT4\uff1b\u7cbe\u5ea6\u8d8a\u4f4e\uff0c\u663e\u5b58\u8d8a\u5c11\u3001\u901f\u5ea6\u8d8a\u5feb<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u663e\u5b58\u4f30\u7b97\u53c2\u8003\uff08\u6a21\u578b\u6743\u91cd\u90e8\u5206\uff09\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>FP32\uff1a\u53c2\u6570\u91cf \u00d7 4 \u5b57\u8282<\/li>\n\n\n\n<li>FP16\uff1a\u53c2\u6570\u91cf \u00d7 2 \u5b57\u8282<\/li>\n\n\n\n<li>INT8\uff1a\u53c2\u6570\u91cf \u00d7 1 \u5b57\u8282<\/li>\n\n\n\n<li>INT4\uff1a\u53c2\u6570\u91cf \u00d7 0.5 \u5b57\u8282 \u603b\u663e\u5b58\u5360\u7528 = \u6a21\u578b\u6743\u91cd + KV Cache + CUDA \u8fd0\u884c\u65f6\u5f00\u9500<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">6. \u591a GPU \u5e76\u884c\u7b56\u7565<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">6.1 \u5f20\u91cf\u5e76\u884c\uff08&#8211;tensor-parallel-size, -tp\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u539f\u7406<\/strong>\uff1a\u5c06\u6a21\u578b\u6743\u91cd\u77e9\u9635\u6a2a\u5411\u5207\u5206\uff0c\u5206\u6563\u5230\u591a\u5f20 GPU\uff0c\u6bcf\u5f20\u5361\u8ba1\u7b97\u4e00\u90e8\u5206\uff0c\u6700\u540e\u62fc\u63a5\u7ed3\u679c\u3002<\/li>\n\n\n\n<li><strong>\u7279\u70b9<\/strong>\uff1a\u964d\u4f4e\u5355\u5361\u663e\u5b58\u5360\u7528\uff0c\u9002\u5408\u5355\u6a21\u578b\u88c5\u4e0d\u4e0b\u5355\u5f20\u5361\u7684\u573a\u666f\uff1b\u5bf9\u901a\u4fe1\u5e26\u5bbd\u8981\u6c42\u9ad8\uff0c\u4ec5\u9002\u5408\u5355\u673a\u5185\u591a\u5361\uff08NVLink \u6700\u4f73\uff09\u3002<\/li>\n\n\n\n<li>\u793a\u4f8b\uff1a4 \u5f20 GPU \u5f20\u91cf\u5e76\u884c\u8fd0\u884c 70B \u6a21\u578b <code>vllm serve Llama-3-70B-Instruct --tensor-parallel-size 4<\/code><\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">6.2 \u6d41\u6c34\u7ebf\u5e76\u884c\uff08&#8211;pipeline-parallel-size, -pp\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u539f\u7406<\/strong>\uff1a\u6309\u6a21\u578b\u5c42\u7eb5\u5411\u5207\u5206\uff0c\u6bcf\u7ec4 GPU \u8d1f\u8d23\u4e00\u90e8\u5206\u5c42\uff0c\u6570\u636e\u4f9d\u6b21\u6d41\u8fc7\u5404\u6bb5\u5b8c\u6210\u8ba1\u7b97\u3002<\/li>\n\n\n\n<li><strong>\u7279\u70b9<\/strong>\uff1a\u9002\u5408\u8de8\u8282\u70b9\u591a\u673a\u90e8\u7f72\uff1b\u5b58\u5728 \u201c\u6c14\u6ce1\u201d\uff08\u524d\u540e\u7ea7\u7b49\u5f85\u5bfc\u81f4\u7684\u7b97\u529b\u6d6a\u8d39\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">6.3 \u6570\u636e\u5e76\u884c\uff08Data Parallelism, DP\uff09<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u539f\u7406<\/strong>\uff1a\u6bcf\u5f20 GPU \u590d\u5236\u5b8c\u6574\u6a21\u578b\uff0c\u6570\u636e\u5207\u5206\u540e\u72ec\u7acb\u8ba1\u7b97\uff0c\u6700\u540e\u6c47\u603b\u68af\u5ea6\u66f4\u65b0\u6a21\u578b\u3002<\/li>\n\n\n\n<li><strong>\u7279\u70b9<\/strong>\uff1a\u5b9e\u73b0\u7b80\u5355\uff0c\u901a\u7528\u6027\u5f3a\uff1b\u65e0\u6cd5\u89e3\u51b3\u6a21\u578b\u8fc7\u5927\u5355\u5361\u88c5\u4e0d\u4e0b\u7684\u95ee\u9898\uff0c\u591a\u7528\u4e8e\u8bad\u7ec3\u573a\u666f\u3002<\/li>\n<\/ul>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">\u4e00\u53e5\u8bdd\u8bb0\u5fc6\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>TP \u6a2a\u5207\u6743\u91cd\uff0c\u5355\u673a\u591a\u5361\u964d\u663e\u5b58<\/li>\n\n\n\n<li>PP \u7ad6\u5207\u5c42\u7ea7\uff0c\u591a\u673a\u90e8\u7f72\u8dd1\u5927\u6a21\u578b<\/li>\n\n\n\n<li>DP \u590d\u5236\u6a21\u578b\uff0c\u6570\u636e\u5206\u7247\u63d0\u541e\u5410<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">7. \u9ad8\u7ea7\u4f18\u5316\u53c2\u6570<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>--enable-prefix-caching<\/code>\uff1a\u5f00\u542f\u524d\u7f00\u7f13\u5b58\uff0c\u76f8\u540c\u7cfb\u7edf\u63d0\u793a\u8bcd\u590d\u7528 KV Cache\uff0c\u663e\u8457\u964d\u4f4e\u9996 Token \u5ef6\u8fdf\uff0c\u5bf9\u8bdd\u7cfb\u7edf\u63a8\u8350\u5f00\u542f\u3002<\/li>\n\n\n\n<li><code>--enable-chunked-prefill<\/code>\uff1a\u5206\u5757\u9884\u586b\u5145\uff0c\u957f\u6587\u672c\u8f93\u5165\u62c6\u5206\u6210\u5c0f\u5757\u4e0e\u751f\u6210\u4ea4\u66ff\u6267\u884c\uff0c\u907f\u514d\u957f\u8bf7\u6c42\u72ec\u5360 GPU\u3002<\/li>\n\n\n\n<li><code>--speculative-model<\/code>\uff1a\u6295\u673a\u89e3\u7801\uff0c\u7528\u5c0f\u6a21\u578b\u751f\u6210\u5019\u9009 Token\uff0c\u5927\u6a21\u578b\u9a8c\u8bc1\uff0c\u5728\u4e0d\u635f\u5931\u8d28\u91cf\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u751f\u6210\u901f\u5ea6\u3002<\/li>\n\n\n\n<li><code>--swap-space<\/code>\uff1aKV Cache \u6362\u51fa\u5230 CPU \u5185\u5b58\u7684\u5927\u5c0f\uff0c\u663e\u5b58\u4e0d\u8db3\u65f6\u4e34\u65f6\u515c\u5e95\uff0c\u4e0d\u5b9c\u8bbe\u7f6e\u8fc7\u5927\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001Open WebUI \u524d\u7aef\u90e8\u7f72<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u6982\u8ff0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Open WebUI \u662f\u529f\u80fd\u4e30\u5bcc\u7684\u81ea\u6258\u7ba1 AI \u524d\u7aef\u754c\u9762\uff0c\u5b8c\u5168\u79bb\u7ebf\u8fd0\u884c\uff0c\u652f\u6301 Ollama\u3001OpenAI \u517c\u5bb9\u534f\u8bae\uff0c\u662f\u672c\u5730\u5927\u6a21\u578b\u7684\u53ef\u89c6\u5316\u4ea4\u4e92\u5165\u53e3\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7279\u70b9\uff1a\u534f\u8bae\u4f18\u5148\u3001\u4e0e\u540e\u7aef\u89e3\u8026\u3001\u652f\u6301\u6a21\u578b\u7ba1\u7406\u3001RAG \u77e5\u8bc6\u5e93\u3001\u63d2\u4ef6\u6269\u5c55<\/li>\n\n\n\n<li>\u9ed8\u8ba4\u7aef\u53e3\uff1a8080<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u90e8\u7f72\u6b65\u9aa4<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">2.1 \u73af\u5883\u51c6\u5907<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>mkdir \/root\/open-webui\ncd \/root\/open-webui\nuv venv --python 3.11  # \u8981\u6c42Python\u22653.11\nsource .venv\/bin\/activate\n<\/code><\/pre>\n\n\n\n<h4 class=\"wp-block-heading\">2.2 \u5b89\u88c5 Open WebUI<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code># \u907f\u514d\u4e0b\u8f7d\u8d85\u65f6\nexport UV_HTTP_TIMEOUT=600\n\n# \u56fd\u5185\u6e90\u5b89\u88c5\nuv pip install open-webui -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple\n<\/code><\/pre>\n\n\n\n<h4 class=\"wp-block-heading\">2.3 \u542f\u52a8\u670d\u52a1<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>export HF_ENDPOINT=https:\/\/hf-mirror.com    # \u6a21\u578b\u4e0b\u8f7d\u955c\u50cf\u52a0\u901f\nexport ENABLE_RAG_WEB_SEARCH=false      # \u5173\u95ed\u8054\u7f51\u641c\u7d22\u529f\u80fd\nopen-webui serve\n<\/code><\/pre>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bbf\u95ee\u5730\u5740\uff1a<code>http:\/\/\u670d\u52a1\u5668IP:8080<\/code><\/li>\n\n\n\n<li>\u9996\u6b21\u8bbf\u95ee\u9700\u6ce8\u518c\u7ba1\u7406\u5458\u8d26\u53f7<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u5bf9\u63a5\u540e\u7aef<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">3.1 \u5bf9\u63a5 Ollama<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u8fdb\u5165\u8bbe\u7f6e \u2192 \u8fde\u63a5\u914d\u7f6e<\/li>\n\n\n\n<li>\u6dfb\u52a0 Ollama \u63a5\u53e3\u5730\u5740\uff1a<code>http:\/\/localhost:11434<\/code><\/li>\n\n\n\n<li>\u542f\u7528\u8fde\u63a5\uff0c\u5728\u6a21\u578b\u4e0b\u62c9\u6846\u9009\u62e9\u5bf9\u5e94\u6a21\u578b\u5373\u53ef\u5bf9\u8bdd<\/li>\n<\/ol>\n\n\n\n<h4 class=\"wp-block-heading\">3.2 \u5bf9\u63a5 vLLM<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u8fdb\u5165\u8bbe\u7f6e \u2192 \u5916\u90e8\u8fde\u63a5<\/li>\n\n\n\n<li>\u6dfb\u52a0 OpenAI \u517c\u5bb9\u63a5\u53e3\uff1a<code>http:\/\/vLLM\u670d\u52a1IP:8000\/v1<\/code><\/li>\n\n\n\n<li>\u586b\u5199 API \u5bc6\u94a5\uff08\u5982\u6709\uff09\uff0c\u4fdd\u5b58\u540e\u5373\u53ef\u9009\u62e9 vLLM \u52a0\u8f7d\u7684\u6a21\u578b<\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001API \u8c03\u7528\u65b9\u5f0f\uff08\u4e86\u89e3\uff09<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">vLLM \u539f\u751f\u517c\u5bb9 OpenAI API \u6807\u51c6\uff0c\u652f\u6301 curl\u3001Python \u7b49\u591a\u79cd\u65b9\u5f0f\u8c03\u7528\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">1. curl \u8c03\u7528\u793a\u4f8b<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>curl http:\/\/localhost:8000\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\n    \"model\": \"Qwen\/Qwen2-1.5B-Instruct\",\n    \"messages\": &#91;{\"role\": \"user\", \"content\": \"\u4ecb\u7ecd\u4e00\u4e0bLinux\"}],\n    \"stream\": false\n  }'\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">2. Python SDK \u8c03\u7528<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">2.1 \u5b89\u88c5\u4f9d\u8d56<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>uv pip install openai\n<\/code><\/pre>\n\n\n\n<h4 class=\"wp-block-heading\">2.2 \u5355\u8f6e\u8c03\u7528\u811a\u672c<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>from openai import OpenAI\n\nclient = OpenAI(\n    base_url=\"http:\/\/localhost:8000\/v1\",\n    api_key=\"token-abc123\"  # \u4e0e\u542f\u52a8\u65f6--api-key\u53c2\u6570\u4e00\u81f4\n)\n\nresp = client.chat.completions.create(\n    model=\"Qwen\/Qwen2-1.5B-Instruct\",\n    messages=&#91;{\"role\": \"user\", \"content\": \"\u4ecb\u7ecd\u4e00\u4e0bLinux\"}]\n)\n\nprint(resp.choices&#91;0].message.content)\n<\/code><\/pre>\n\n\n\n<h4 class=\"wp-block-heading\">2.3 \u5e76\u53d1\u6279\u91cf\u8c03\u7528<\/h4>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f7f\u7528\u7ebf\u7a0b\u6c60\u5b9e\u73b0\u591a\u8bf7\u6c42\u5e76\u53d1\uff0c\u5e38\u7528\u4e8e\u538b\u6d4b\u573a\u666f\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>from openai import OpenAI\nfrom concurrent.futures import ThreadPoolExecutor\n\nclient = OpenAI(base_url=\"http:\/\/localhost:8000\/v1\", api_key=\"token-abc123\")\n\nquestions = &#91;\n    \"\u4ecb\u7ecd\u4e00\u4e0bLinux\",\n    \"\u4ec0\u4e48\u662fDocker\",\n    \"\u4ec0\u4e48\u662fKubernetes\",\n    \"\u89e3\u91ca\u4e00\u4e0bTCP\u4e09\u6b21\u63e1\u624b\"\n]\n\ndef ask(q):\n    resp = client.chat.completions.create(\n        model=\"Qwen\/Qwen2-1.5B-Instruct\",\n        messages=&#91;{\"role\": \"user\", \"content\": q}]\n    )\n    return q, resp.choices&#91;0].message.content\n\nwith ThreadPoolExecutor(max_workers=5) as executor:\n    results = executor.map(ask, questions)\n    for q, ans in results:\n        print(f\"\u95ee\u9898: {q}\")\n        print(f\"\u56de\u7b54: {ans}\")\n        print(\"-\" * 50)\n<\/code><\/pre>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001\u9ad8\u9891\u6c47\u603b<\/h2>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>PagedAttention \u7684\u539f\u7406\u548c\u4f18\u52bf\uff1f<\/strong> \u7b54\uff1a\u501f\u9274\u64cd\u4f5c\u7cfb\u7edf\u5206\u9875\u601d\u60f3\uff0c\u5c06 KV Cache \u5207\u5206\u4e3a\u56fa\u5b9a\u5927\u5c0f\u7684 Block\uff0c\u901a\u8fc7\u9875\u8868\u6620\u5c04\u7ba1\u7406\uff0c\u7269\u7406\u4e0a\u65e0\u9700\u8fde\u7eed\uff0c\u6309\u9700\u5206\u914d\u3001\u7528\u5b8c\u56de\u6536\u3002\u89e3\u51b3\u4e86\u4f20\u7edf KV Cache \u663e\u5b58\u6d6a\u8d39\u3001\u788e\u7247\u5316\u4e25\u91cd\u7684\u95ee\u9898\uff0c\u663e\u5b58\u5229\u7528\u7387\u63d0\u5347\u81f3 96% \u4ee5\u4e0a\uff0c\u53ef\u652f\u6301\u66f4\u591a\u5e76\u53d1\u8bf7\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u8fde\u7eed\u6279\u5904\u7406\u76f8\u6bd4\u4f20\u7edf\u9759\u6001\u6279\u5904\u7406\u7684\u4f18\u52bf\uff1f<\/strong> \u7b54\uff1a\u4f20\u7edf\u9759\u6001\u6279\u5904\u7406\u5fc5\u987b\u7b49\u6574\u6279\u8bf7\u6c42\u5168\u90e8\u5b8c\u6210\u624d\u80fd\u5904\u7406\u4e0b\u4e00\u6279\uff0cGPU \u7a7a\u8f6c\u65f6\u95f4\u591a\u3001\u6574\u4f53\u5ef6\u8fdf\u9ad8\u3002\u8fde\u7eed\u6279\u5904\u7406\u52a8\u6001\u5c06\u65b0\u8bf7\u6c42\u63d2\u5165\u8fd0\u884c\u4e2d\u7684\u6279\u6b21\uff0c\u8bf7\u6c42\u5b8c\u6210\u7acb\u5373\u91ca\u653e\u4f4d\u7f6e\u8865\u5165\u65b0\u8bf7\u6c42\uff0cGPU \u59cb\u7ec8\u4fdd\u6301\u9ad8\u8d1f\u8f7d\uff0c\u541e\u5410\u91cf\u66f4\u9ad8\u3001\u5e73\u5747\u5ef6\u8fdf\u66f4\u4f4e\u3002<\/li>\n\n\n\n<li><strong>vLLM \u7684\u591a GPU \u5e76\u884c\u7b56\u7565\u6709\u54ea\u4e9b\uff0c\u533a\u522b\u662f\u4ec0\u4e48\uff1f<\/strong> \u7b54\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5f20\u91cf\u5e76\u884c\uff08TP\uff09\uff1a\u6a2a\u5411\u5207\u5206\u6743\u91cd\u77e9\u9635\uff0c\u5355\u673a\u591a\u5361\u90e8\u7f72\uff0c\u964d\u4f4e\u5355\u5361\u663e\u5b58\u538b\u529b\uff0c\u5bf9\u901a\u4fe1\u5e26\u5bbd\u8981\u6c42\u9ad8<\/li>\n\n\n\n<li>\u6d41\u6c34\u7ebf\u5e76\u884c\uff08PP\uff09\uff1a\u7eb5\u5411\u5207\u5206\u6a21\u578b\u5c42\u7ea7\uff0c\u9002\u5408\u8de8\u591a\u673a\u90e8\u7f72\uff0c\u5b58\u5728\u7b97\u529b\u6c14\u6ce1\u6d6a\u8d39<\/li>\n\n\n\n<li>\u6570\u636e\u5e76\u884c\uff08DP\uff09\uff1a\u6bcf\u5f20\u5361\u590d\u5236\u5b8c\u6574\u6a21\u578b\uff0c\u6570\u636e\u5206\u7247\u8ba1\u7b97\uff0c\u591a\u7528\u4e8e\u8bad\u7ec3\uff0c\u4e0d\u80fd\u89e3\u51b3\u5355\u5361\u88c5\u4e0d\u4e0b\u6a21\u578b\u7684\u95ee\u9898<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5f71\u54cd\u5927\u6a21\u578b\u63a8\u7406\u663e\u5b58\u5360\u7528\u7684\u56e0\u7d20\u6709\u54ea\u4e9b\uff1f<\/strong> \u7b54\uff1a\u6a21\u578b\u53c2\u6570\u91cf\u3001\u63a8\u7406\u6570\u636e\u7cbe\u5ea6\uff08dtype\uff09\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u5e76\u53d1\u8bf7\u6c42\u6570\u91cf\u3001KV Cache \u5927\u5c0f\u3001CUDA \u8fd0\u884c\u65f6\u56fa\u5b9a\u5f00\u9500\u3002<\/li>\n\n\n\n<li><strong>\u5982\u4f55\u63d0\u5347\u63a8\u7406\u670d\u52a1\u7684\u5e76\u53d1\u80fd\u529b\uff1f<\/strong> \u7b54\uff1a\u4f7f\u7528 PagedAttention \u4f18\u5316\u663e\u5b58\u5229\u7528\u7387\uff1b\u5f00\u542f\u8fde\u7eed\u6279\u5904\u7406\u63d0\u5347 GPU \u4f7f\u7528\u7387\uff1b\u9002\u5f53\u964d\u4f4e\u63a8\u7406\u7cbe\u5ea6\uff08\u5982 INT4\/INT8\uff09\uff1b\u5408\u7406\u8bbe\u7f6e max-model-len\uff0c\u907f\u514d\u76f2\u76ee\u5f00\u5927\u4e0a\u4e0b\u6587\uff1b\u4f7f\u7528\u5f20\u91cf\u5e76\u884c\u6269\u5c55\u5355\u5361\u663e\u5b58\u5bb9\u91cf\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001vLLM \u4f01\u4e1a\u7ea7\u63a8\u7406\u6846\u67b6 1. \u6846\u67b6\u5b9a\u4f4d vLLM\uff08virtual Large Language Mode [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[],"class_list":["post-132","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/132","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=132"}],"version-history":[{"count":1,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/132\/revisions"}],"predecessor-version":[{"id":133,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/132\/revisions\/133"}],"wp:attachment":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=132"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=132"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=132"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}