vLLM \u662f\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u63a8\u7406\u548c\u670d\u52a1\u5f15\u64ce\uff0c\u652f\u6301 PagedAttention \u6280\u672f\uff0c\u80fd\u591f\u4ee5\u6781\u9ad8\u7684\u541e\u5410\u91cf\u8fd0\u884c\u5404\u79cd\u5f00\u6e90 LLM\uff0c\u5982 Qwen\u3001Llama\u3001Gemma \u7b49\u3002<\/p>\n
![\"vLLM](\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\")
<\/div><\/p>\n
3.1 \u5b89\u88c5 uv \u5305\u7ba1\u7406\u5668\uff08\u63a8\u8350\uff09<\/h3>\n
uv \u662f\u4e00\u4e2a\u6781\u901f\u7684 Python \u5305\u7ba1\u7406\u5de5\u5177\uff0c\u6bd4\u4f20\u7edf\u7684 pip \u5feb 10-100 \u500d\uff1a<\/p>\n
# \u5b89\u88c5 uv\nwget -qO- https:\/\/astral.sh\/uv\/install.sh | sh\nsource $HOME\/.cargo\/env<\/code><\/pre>\n3.2 \u521b\u5efa\u865a\u62df\u73af\u5883\u5e76\u5b89\u88c5 vLLM<\/h3>\n# \u521b\u5efa\u4e13\u7528\u76ee\u5f55\nmkdir ~\/vllm-service && cd ~\/vllm-service\n\n# \u4f7f\u7528 uv \u521b\u5efa\u865a\u62df\u73af\u5883\nuv venv --python 3.12\nsource .venv\/bin\/activate\n\n# \u5b89\u88c5 vLLM\uff08\u81ea\u52a8\u68c0\u6d4b CUDA \u7248\u672c\uff09\nuv pip install vllm --torch-backend=auto<\/code><\/pre>\n\u5982\u679c\u6ca1\u6709\u5b89\u88c5 uv\uff0c\u4e5f\u53ef\u4ee5\u7528\u4f20\u7edf\u7684 pip \u65b9\u5f0f\uff1a<\/p>\n
python3 -m venv ~\/vllm-service\/.venv\nsource ~\/vllm-service\/.venv\/bin\/activate\npip install vllm<\/code><\/pre>\n3.3 \u9a8c\u8bc1\u5b89\u88c5<\/h3>\npython3 -c \"import vllm; print(f'vLLM version: {vllm.__version__}')\"<\/code><\/pre>\n3.4 \u542f\u52a8 vLLM \u670d\u52a1<\/h3>\n
vLLM \u53ef\u4ee5\u542f\u52a8\u4e00\u4e2a\u517c\u5bb9 OpenAI API \u683c\u5f0f\u7684\u63a8\u7406\u670d\u52a1\u5668\uff1a<\/p>\n
# \u4ee5 Qwen2.5-7B \u6a21\u578b\u4e3a\u4f8b\nvllm serve Qwen\/Qwen2.5-7B-Instruct \\\n --host 0.0.0.0 \\\n --port 8000 \\\n --gpu-memory-utilization 0.85 \\\n --max-model-len 8192<\/code><\/pre>\n\u542f\u52a8\u540e\uff0c\u4f60\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u6d4b\u8bd5 API\uff1a<\/p>\n
# \u6d4b\u8bd5\u5bf9\u8bdd\u63a5\u53e3\ncurl http:\/\/localhost:8000\/v1\/chat\/completions \\\n -H \"Content-Type: application\/json\" \\\n -d '{\n \"model\": \"Qwen\/Qwen2.5-7B-Instruct\",\n \"messages\": [\n {\"role\": \"user\", \"content\": \"\u4f60\u597d\uff0c\u8bf7\u4ecb\u7ecd\u4e00\u4e0b\u4f60\u81ea\u5df1\"}\n ],\n \"temperature\": 0.7,\n \"max_tokens\": 512\n }'<\/code><\/pre>\n3.5 vLLM \u5e38\u7528\u53c2\u6570\u8bf4\u660e<\/h3>\n
\n\n\n| \u53c2\u6570<\/th>\n | \u8bf4\u660e<\/th>\n | \u793a\u4f8b\u503c<\/th>\n<\/tr>\n<\/thead>\n |
\n\n--host<\/code><\/td>\n| \u76d1\u542c\u5730\u5740<\/td>\n | 0.0.0.0<\/td>\n<\/tr>\n | \n--port<\/code><\/td>\n| \u76d1\u542c\u7aef\u53e3<\/td>\n | 8000<\/td>\n<\/tr>\n | \n--gpu-memory-utilization<\/code><\/td>\n| GPU \u663e\u5b58\u4f7f\u7528\u6bd4\u4f8b<\/td>\n | 0.85<\/td>\n<\/tr>\n | \n--max-model-len<\/code><\/td>\n| \u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6<\/td>\n | 8192<\/td>\n<\/tr>\n | \n--quantization<\/code><\/td>\n| \u91cf\u5316\u65b9\u5f0f\uff08\u8282\u7701\u663e\u5b58\uff09<\/td>\n | awq \/ gptq \/ fp8<\/td>\n<\/tr>\n | \n--tensor-parallel-size<\/code><\/td>\n| \u591a GPU \u5e76\u884c\u6570<\/td>\n | 2\uff08\u53cc\u5361\u65f6\uff09<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n3.6 \u8bbe\u7f6e\u5f00\u673a\u81ea\u542f\uff08\u53ef\u9009\uff09<\/h3>\nsudo tee \/etc\/systemd\/system\/vllm.service << 'EOF'\n[Unit]\nDescription=vLLM Inference Server\nAfter=network.target\n\n[Service]\nType=simple\nUser=\u4f60\u7684\u7528\u6237\u540d\nWorkingDirectory=\/home\/\u4f60\u7684\u7528\u6237\u540d\/vllm-service\nExecStart=\/home\/\u4f60\u7684\u7528\u6237\u540d\/vllm-service\/.venv\/bin\/vllm serve Qwen\/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.85\nRestart=on-failure\nRestartSec=10\nEnvironment=HUGGING_FACE_HUB_TOKEN=\u4f60\u7684token\n\n[Install]\nWantedBy=multi-user.target\nEOF\n\nsudo systemctl daemon-reload\nsudo systemctl enable vllm\nsudo systemctl start vllm<\/code><\/pre>\n\u2699\ufe0f \u7b2c\u56db\u6b65\uff1aComfyUI + vLLM \u8054\u52a8\uff08\u8fdb\u9636\u73a9\u6cd5\uff09<\/h2>\n\u4f60\u53ef\u4ee5\u5c06 vLLM \u548c ComfyUI \u7ed3\u5408\u4f7f\u7528\uff0c\u5b9e\u73b0 “\u6587\u5b57\u63cf\u8ff0 \u2192 AI \u81ea\u52a8\u6269\u5199 Prompt \u2192 \u751f\u6210\u56fe\u7247”<\/strong> \u7684\u5168\u81ea\u52a8\u5de5\u4f5c\u6d41\u3002<\/p>\n\u601d\u8def\u5982\u4e0b\uff1a<\/p>\n \n- \u7528 vLLM \u90e8\u7f72\u4e00\u4e2a LLM\uff08\u5982 Qwen2.5\uff09\uff0c\u901a\u8fc7 API \u63a5\u6536\u7b80\u77ed\u63cf\u8ff0\uff0c\u81ea\u52a8\u6269\u5c55\u4e3a\u8be6\u7ec6\u7684\u56fe\u7247\u751f\u6210 Prompt<\/li>\n
- \u5c06\u6269\u5c55\u540e\u7684 Prompt \u4f20\u9012\u7ed9 ComfyUI \u7684 API \u63a5\u53e3<\/li>\n
- ComfyUI \u6839\u636e Prompt \u81ea\u52a8\u751f\u6210\u9ad8\u8d28\u91cf\u56fe\u7247<\/li>\n<\/ol>\n
\u8fd9\u79cd\u7ec4\u5408\u5728\u6279\u91cf\u751f\u6210\u7d20\u6750\u3001\u8bbe\u8ba1 AI \u521b\u4f5c\u5de5\u4f5c\u6d41\u7b49\u573a\u666f\u4e2d\u975e\u5e38\u5b9e\u7528\uff01<\/p>\n \ud83d\udee1\ufe0f \u5e38\u89c1\u95ee\u9898\u6392\u67e5<\/h2>\nQ1: nvidia-smi \u62a5\u9519 “NVIDIA-SMI has failed”<\/h3>\n\u89e3\u51b3\u65b9\u6848\uff1a\u91cd\u65b0\u5b89\u88c5\u9a71\u52a8\uff0c\u6216\u8005\u68c0\u67e5\u5185\u6838\u7248\u672c\u662f\u5426\u5339\u914d\uff1a<\/p>\n sudo apt install --reinstall nvidia-driver-580-server\nsudo reboot<\/code><\/pre>\nQ2: ComfyUI \u542f\u52a8\u540e CUDA out of memory<\/h3>\n\u4f7f\u7528\u4f4e\u663e\u5b58\u6a21\u5f0f\u542f\u52a8\uff1a<\/p>\n python main.py --listen 0.0.0.0 --lowvram<\/code><\/pre>\nQ3: vLLM \u5b89\u88c5\u65f6\u62a5\u9519 torch \u7248\u672c\u4e0d\u517c\u5bb9<\/h3>\n\u6307\u5b9a CUDA \u540e\u7aef\u5b89\u88c5\uff1a<\/p>\n uv pip install vllm --torch-backend=cu126<\/code><\/pre>\nQ4: \u5982\u4f55\u8ba9 ComfyUI \u548c vLLM \u540c\u65f6\u8fd0\u884c\u4e0d\u51b2\u7a81\uff1f<\/h3>\n\u5982\u679c\u53ea\u6709\u4e00\u5f20 GPU\uff0c\u5efa\u8bae\uff1a<\/p>\n \n- \u4e3a vLLM \u9650\u5236\u663e\u5b58\u4f7f\u7528\uff1a
--gpu-memory-utilization 0.5<\/code><\/li>\n- \u4e3a ComfyUI \u4f7f\u7528
--lowvram<\/code> \u6a21\u5f0f<\/li>\n- \u6216\u8005\u53ea\u5728\u9700\u8981\u65f6\u542f\u52a8\u5176\u4e2d\u4e00\u4e2a\u670d\u52a1<\/li>\n<\/ul>\n
\ud83d\udcdd \u603b\u7ed3<\/h2>\n\u901a\u8fc7\u672c\u6559\u7a0b\uff0c\u4f60\u5df2\u7ecf\u6210\u529f\u5728 Ubuntu 24.04 \u4e0a\u642d\u5efa\u4e86\uff1a<\/p>\n \n- \u2705 ComfyUI<\/strong>\uff1a\u5f3a\u5927\u7684 AI \u56fe\u50cf\u751f\u6210\u5de5\u4f5c\u6d41\u5e73\u53f0<\/li>\n
- \u2705 vLLM<\/strong>\uff1a\u9ad8\u6027\u80fd\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u5f15\u64ce<\/li>\n<\/ul>\n
\u8fd9\u4e24\u4e2a\u5de5\u5177\u7684\u7ec4\u5408\uff0c\u8ba9\u4f60\u5728\u672c\u5730\u62e5\u6709\u4e86\u4e00\u4e2a\u529f\u80fd\u5b8c\u5907\u7684 AI \u5de5\u4f5c\u7ad9\u3002\u65e0\u8bba\u662f\u751f\u6210\u7cbe\u7f8e\u7684 AI \u827a\u672f\u4f5c\u54c1\uff0c\u8fd8\u662f\u90e8\u7f72\u81ea\u5df1\u7684\u804a\u5929\u673a\u5668\u4eba\uff0c\u90fd\u53ef\u4ee5\u8f7b\u677e\u5b9e\u73b0\u3002\u795d\u4f60\u73a9\u5f97\u6109\u5feb\uff01\ud83c\udf89<\/p>\n","protected":false},"excerpt":{"rendered":" \u524d\u8a00 \u968f\u7740 AI \u6280\u672f\u7684\u98de\u901f\u53d1\u5c55\uff0c\u672c\u5730\u90e8\u7f72 AI \u6a21\u578b\u5df2\u7ecf\u6210\u4e3a\u8d8a\u6765\u8d8a\u591a\u5f00\u53d1\u8005\u548c\u521b\u4f5c\u8005\u7684\u521a\u9700\u3002ComfyUI \u662f […]<\/p>\n","protected":false},"author":1,"featured_media":284,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-281","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-linux-tinkering"],"_links":{"self":[{"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/posts\/281","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/comments?post=281"}],"version-history":[{"count":4,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/posts\/281\/revisions"}],"predecessor-version":[{"id":301,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/posts\/281\/revisions\/301"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/media\/284"}],"wp:attachment":[{"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/media?parent=281"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/categories?post=281"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.nickywin.com\/wp-json\/wp\/v2\/tags?post=281"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
| | | | | | |
![\"Ubuntu](\"https:\/\/blog.nickywin.com\/wp-content\/uploads\/2026\/04\/tutorial_header_1776194185670.jpg\")
<\/div><\/p>\n![\"NVIDIA](\"https:\/\/blog.nickywin.com\/wp-content\/uploads\/2026\/04\/nvidia_gpu_setup_1776194201327.jpg\")
<\/div><\/p>\n![\"ComfyUI](\"https:\/\/blog.nickywin.com\/wp-content\/uploads\/2026\/04\/comfyui_interface_1776194214642.jpg\")
<\/div><\/p>\n![\"vLLM](\"https:\/\/blog.nickywin.com\/wp-content\/uploads\/2026\/04\/vllm_architecture_1776194228757.jpg\")
<\/div><\/p>\n