{"id":53817,"date":"2025-02-16T14:44:33","date_gmt":"2025-02-16T06:44:33","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53817\/"},"modified":"2025-02-16T14:44:33","modified_gmt":"2025-02-16T06:44:33","slug":"vllm-ollama%e7%bb%bc%e5%90%88%e5%af%b9%e6%af%94","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53817\/","title":{"rendered":"vLLM\/ollama\u7efc\u5408\u5bf9\u6bd4"},"content":{"rendered":"<p>\u6b22\u8fce\u6765\u5230\u6211\u4eec\u6df1\u5165\u7814\u7a76 LLM \u63a8\u7406\u6846\u67b6\u7684\u6700\u540e\u4e00\u90e8\u5206\uff01\u5728\u7b2c\u4e00\u90e8\u5206\u548c\u7b2c\u4e8c\u90e8\u5206\u4e2d\uff0c\u6211\u4eec\u5206\u522b\u63a2\u8ba8\u4e86 Ollama \u548c vLLM\uff0c\u4e86\u89e3\u4e86\u5b83\u4eec\u7684\u67b6\u6784\u3001\u529f\u80fd\u548c\u57fa\u672c\u6027\u80fd\u7279\u5f81\u3002\u73b0\u5728\u5230\u4e86\u51b3\u5b9a\u6027\u7684\u4e00\u8f6e\uff1a\u9762\u5bf9\u9762\u7684\u6bd4\u8f83\uff0c\u4ee5\u5e2e\u52a9\u60a8\u6839\u636e\u7279\u5b9a\u9700\u6c42\u9009\u62e9\u5408\u9002\u7684\u6846\u67b6\u3002<\/p>\n<p>\u8fd9\u6b21\u6bd4\u8f83\u5e76\u4e0d\u662f\u8981\u5ba3\u5e03\u7edd\u5bf9\u7684\u8d62\u5bb6\u2014\u2014\u800c\u662f\u8981\u4e86\u89e3\u54ea\u79cd\u6846\u67b6\u5728\u4e0d\u540c\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u6211\u4eec\u5c06\u91cd\u70b9\u5173\u6ce8\uff1a<\/p>\n<ul>\n<li>\u8d44\u6e90\u5229\u7528\u7387\u548c\u6548\u7387<\/li>\n<li>\u90e8\u7f72\u548c\u7ef4\u62a4\u7684\u7b80\u6613\u6027<\/li>\n<li>\u5177\u4f53\u7528\u4f8b\u548c\u5efa\u8bae<\/li>\n<li>\u5b89\u5168\u548c\u751f\u4ea7\u51c6\u5907<\/li>\n<li>\u6587\u6863<\/li>\n<\/ul>\n<p>\u8ba9\u6211\u4eec\u6df1\u5165\u7814\u7a76\u6570\u636e\uff0c\u770b\u770b\u6211\u4eec\u7684\u6d4b\u8bd5\u63ed\u793a\u4e86\u4ec0\u4e48\uff01<\/p>\n<p>\u53ea\u6709\u4e00\u4e2a\u53ef\u4ee5\u6210\u4e3a\u51a0\u519b\uff0c\u6216\u8005\u53ef\u80fd\u4e0d\u662f\uff1f \ud83e\udd14<\/p>\n<h2>1\u3001\u57fa\u51c6\u6d4b\u8bd5\u8bbe\u7f6e<\/h2>\n<p>\u4e3a\u4e86\u786e\u4fdd\u516c\u5e73\u6bd4\u8f83\uff0c\u6211\u4eec\u5c06\u5bf9\u4e24\u4e2a\u6846\u67b6\u4f7f\u7528\u76f8\u540c\u7684\u786c\u4ef6\u548c\u6a21\u578b\uff1a<\/p>\n<p>\u786c\u4ef6\u914d\u7f6e\uff1a<\/p>\n<ul>\n<li>GPU\uff1aNVIDIA RTX 4060 16GB Ti<\/li>\n<li>RAM\uff1a64GB RAM<\/li>\n<li>CPU\uff1aAMD Ryzen 7<\/li>\n<li>\u5b58\u50a8\uff1aNVMe SSD<\/li>\n<\/ul>\n<p>\u578b\u53f7\uff1a<\/p>\n<ul>\n<li>Qwen2.5\u201314B-Instruct\uff084 \u4f4d\u91cf\u5316\uff09<\/li>\n<li>\u4e0a\u4e0b\u6587\u957f\u5ea6\uff1a8192 \u4e2a\u6807\u8bb0<\/li>\n<li>\u6279\u5904\u7406\u5927\u5c0f\uff1a1\uff08\u5355\u7528\u6237\u573a\u666f\uff09<\/li>\n<\/ul>\n<h2>2\u3001\u975e\u5e38\u516c\u5e73\u7684\u6bd4\u8f83<\/h2>\n<p>\u8ba9\u6211\u4eec\u5206\u6790\u4e00\u4e0b\u8fd9\u4e24\u4e2a\u6846\u67b6\u5982\u4f55\u4ee5\u4e0d\u540c\u7684\u65b9\u5f0f\u7ba1\u7406\u7cfb\u7edf\u8d44\u6e90\uff0c\u91cd\u70b9\u5173\u6ce8\u5b83\u4eec\u7684\u6838\u5fc3\u67b6\u6784\u65b9\u6cd5\u548c\u5b9e\u9645\u5f71\u54cd\u3002<\/p>\n<h3>2.1 Ollama<\/h3>\n<p>\u6211\u4e3e\u4e86\u4e00\u4e2a\u95ee\u9898\u201c\u7ed9\u6211\u8bb2\u4e00\u4e2a 1000 \u5b57\u7684\u6545\u4e8b\u201d\u7684\u4f8b\u5b50\u3002\u6211\u4e00\u4e2a\u8bf7\u6c42\u7684 tok\/sec \u4e3a 25.59\u3002\u6ca1\u6709\u5e76\u884c\u8bf7\u6c42<\/p>\n<p>  \u95ee\u9898\uff1a\u201c\u7ed9\u6211\u8bb2\u4e00\u4e2a 1000 \u5b57\u7684\u6545\u4e8b\u201d\u7528\u4e8e Ollama <\/p>\n<p>\u5bf9\u4e8e\u5e76\u884c\u8bf7\u6c42\uff0c\u7528\u6237\u5fc5\u987b\u4fee\u6539\u4f4d\u4e8e <code>\/etc\/systemd\/system\/ollama.service<\/code> \u4e2d\u7684\u6587\u4ef6\uff08\u5bf9\u4e8e Ubuntu\uff09\u5e76\u6dfb\u52a0\u4e00\u884c <code>Environment=\"OLLAMA_NUM_PARALLEL=4\"<\/code>\uff0c\u4f60\u5c06\u88ab\u5141\u8bb8\u6267\u884c\u6700\u591a 4 \u4e2a\u5e76\u884c\u8bf7\u6c42<\/p>\n<pre><code>[Unit]\nDescription=Ollama Service\nAfter=network-online.target\n\n[Service]\nExecStart=\/usr\/local\/bin\/ollama serve\nUser=ollama\nGroup=ollama\nRestart=always\nRestartSec=3\nEnvironment=\"PATH=\/home\/henry\/.local\/bin:\/usr\/local\/cuda\/bin\/:\/usr\/local\/sbin:\/usr\/local\/bin:\/usr\/sbin:\/usr\/bin:\/sbin:\/bin\"\nEnvironment=\"OLLAMA_HOST=0.0.0.0:11434\"\nEnvironment=\"OLLAMA_DEBUG=1\"\nEnvironment=\"OLLAMA_NUM_PARALLEL=4\"\nEnvironment=\"OPENAI_BASE_URL=http:\/\/0.0.0.0:11434\/api\"\n\n[Install]\nWantedBy=multi-user.target<\/code><\/pre>\n<p>\u8fd9\u91cc\u662f\u6211\u5b8c\u5168\u4e0d\u559c\u6b22 Ollama \u7684\u5730\u65b9\uff0c\u6211\u8ba4\u4e3a\u5b83\u4e0d\u662f\u4e00\u4e2a\u597d\u7684\u751f\u4ea7\u6846\u67b6\u3002 Ollama \u4fdd\u7559\u4e86\u6240\u9700\u7684\u6240\u6709\u5185\u5b58\uff0c\u5373\u4f7f\u5176\u4e2d\u53ea\u6709\u4e00\u5c0f\u90e8\u5206\u4f1a\u88ab\u4f7f\u7528\u3002\u6211\u7684\u610f\u601d\u662f\uff0c\u53ea\u6709 4 \u4e2a\u5e76\u53d1\u8bf7\u6c42\uff0c\u5c31\u4e0d\u53ef\u80fd\u5728 GPU \u4e0a\u52a0\u8f7d\u6574\u4e2a\u6a21\u578b\uff0c\u5e76\u4e14\u4e00\u4e9b\u5c42\u4f1a\u52a0\u8f7d\u5230 CPU \u4e0a\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u6216\u5728\u7ec8\u7aef\u4e2d\u8fd0\u884c <code>ollama ps<\/code> \u5373\u53ef\u770b\u5230<\/p>\n<p>  15% \u7684\u795e\u7ecf\u7f51\u7edc\u6b63\u5728 GPU \u4e2d\u52a0\u8f7d <\/p>\n<p>\u8fd9\u8fd8\u4e0d\u662f\u6700\u7cdf\u7cd5\u7684\u90e8\u5206\u3002\u6211\u770b\u5230\u7684\u662f 15% \u7684\u795e\u7ecf\u7f51\u7edc\u6b63\u5728 GPU \u4e2d\u52a0\u8f7d\uff0c\u4f46 GPU \u4e2d\u6709\u8fd1 2GB \u7684 VRAM \u53ef\u7528\uff01\u4f46 Ollama \u4e3a\u4ec0\u4e48\u8981\u8fd9\u6837\u505a\uff1f<\/p>\n<p>\u5728\u6211\u5199\u8fd9\u4e9b\u884c\u65f6\uff0cGitHub \u4e0a\uff0c\u4f46 Ollama \u5f00\u53d1\u4eba\u5458\u5e76\u672a\u5bf9\u6b64\u4e88\u4ee5\u5173\u6ce8\u3002\u51e0\u4e2a\u7528\u6237\u90fd\u9762\u4e34\u7740\u540c\u6837\u7684\u95ee\u9898\uff0c\u52a0\u8f7d\u6574\u4e2a\u795e\u7ecf\u7f51\u7edc\u4f3c\u4e4e\u975e\u5e38\u56f0\u96be\uff0c\u5373\u4f7f\u6211\u4eec\u8c08\u8bba\u7684\u662f\u4ec5\u5e76\u884c 4 \u4e2a\u8bf7\u6c42\u3002Ollama \u6ca1\u6709\u63d0\u4f9b\u4efb\u4f55\u6587\u6863\u3002<\/p>\n<p>\u77e5\u9053\u8fd9\u4e00\u70b9\u540e\uff0cOllama \u53ef\u4ee5\u652f\u6301\u7684\u6700\u5927\u4e0a\u4e0b\u6587\u91cf\u662f\u591a\u5c11\uff0c\u624d\u80fd\u5728 GPU \u4e2d\u52a0\u8f7d 100% \u7684\u6a21\u578b\uff1f\u6211\u5c1d\u8bd5\u901a\u8fc7\u8bbe\u7f6e PARAMETER num_ctx 24576\uff08\u7a0d\u540e\u4f60\u5c06\u770b\u5230\u4e3a\u4ec0\u4e48\u662f\u8fd9\u4e2a\u6570\u5b57\uff09\u6765\u4fee\u6539\u6211\u7684\u6a21\u578b\u6587\u4ef6\uff0c\u6211\u6ce8\u610f\u5230\u51fa\u73b0\u4e86\u540c\u6837\u7684\u95ee\u9898\uff1a\u5c3d\u7ba1 GPU \u4e2d\u6709\u8fd1 2GB \u7684 VRAM \u53ef\u7528\uff0c\u4f46 CPU \u7684\u4f7f\u7528\u7387\u4e3a 4%\u3002<\/p>\n<p>  Ollama \u5728 CPU \u4e2d\u52a0\u8f7d\u4e86 4% \u7684\u6a21\u578b \ud83d\ude41 <\/p>\n<h3>2.2 vLLM<\/h3>\n<p>vLLM \u91c7\u7528\u7eaf GPU \u4f18\u5316\u65b9\u6cd5\uff0c\u6b63\u5982\u6211\u4eec\u5728\u672c\u7cfb\u5217\u7684\u7b2c\u4e8c\u90e8\u5206\u4e2d\u770b\u5230\u7684\uff0cGGUF \u91cf\u5316\u4ecd\u5904\u4e8e\u5b9e\u9a8c\u9636\u6bb5\u3002\u6211\u5fc5\u987b\u8fdb\u884c\u540c\u7c7b\u6bd4\u8f83\uff0c\u6240\u4ee5\u6211\u60f3\u4e3a\u6211\u7684 GPU \u83b7\u5f97\u6700\u5927\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002\u7ecf\u8fc7\u51e0\u6b21\u5c1d\u8bd5\uff0c\u6211\u7684 RTX 4060 Ti \u652f\u6301 24576 \u4e2a\u4ee4\u724c\u3002\u6240\u4ee5\u6211\u8fd0\u884c\u4e86\u8fd9\u4e2a\u4fee\u6539\u540e\u7684 docker\uff08\u76f8\u5bf9\u4e8e\u672c\u7cfb\u5217\u7684\u7b2c\u4e8c\u90e8\u5206\uff09\uff1a<\/p>\n<pre><code># Run the container with GPU support\ndocker run -it \\\n    --runtime nvidia \\\n    --gpus all \\\n    --network=\"host\" \\\n    --ipc=host \\\n    -v .\/models:\/vllm-workspace\/models \\\n    -v .\/config:\/vllm-workspace\/config \\\n    vllm\/vllm-openai:latest \\\n    --model models\/Qwen2.5-14B-Instruct\/Qwen2.5-14B-Instruct-Q4_K_M.gguf \\\n    --tokenizer Qwen\/Qwen2.5-14B-Instruct \\\n    --host \"0.0.0.0\" \\\n    --port 5000 \\\n    --gpu-memory-utilization 1.0 \\\n    --served-model-name \"VLLMQwen2.5-14B\" \\\n    --max-num-batched-tokens 24576 \\\n    --max-num-seqs 256 \\\n    --max-model-len 8192 \\\n    --generation-config config<\/code><\/pre>\n<p>\u6211\u53ef\u4ee5\u540c\u65f6\u8fd0\u884c\u591a\u8fbe 20 \u4e2a\u8bf7\u6c42\uff01\uff01\u592a\u75af\u72c2\u4e86\uff01\uff01\u3002\u4e3a\u4e86\u6d4b\u8bd5\u8fd9\u4e2a\u6846\u67b6\uff0c\u6211\u4f7f\u7528\u4e86\u4ee5\u4e0b\u4ee3\u7801\uff1a<\/p>\n<pre><code>import requests\nimport concurrent.futures\n\nBASE_URL = \"http:\/\/&lt;your_vLLM_server_ip&gt;:5000\/v1\"\nAPI_TOKEN = \"sk-1234\"\nMODEL = \"VLLMQwen2.5-14B\"\n\ndef create_request_body():\n    return {\n        \"model\": MODEL,\n        \"messages\": [\n            {\"role\": \"user\", \"content\": \"Tell me a story of 1000 words.\"}\n        ]\n    }\n\ndef make_request(request_body):\n    headers = {\n        \"Authorization\": f\"Bearer {API_TOKEN}\",\n        \"Content-Type\": \"application\/json\"\n    }\n    response = requests.post(f\"{BASE_URL}\/chat\/completions\", json=request_body, headers=headers, verify=False)\n    return response.json()\n\ndef parallel_requests(num_requests):\n    request_body = create_request_body()\n    with concurrent.futures.ThreadPoolExecutor(max_workers=num_requests) as executor:\n        futures = [executor.submit(make_request, request_body) for _ in range(num_requests)]\n        results = [future.result() for future in concurrent.futures.as_completed(futures)]\n    return results\n\nif __name__ == \"__main__\":\n    num_requests = 50  # Example: Set the number of parallel requests\n    responses = parallel_requests(num_requests)\n    for i, response in enumerate(responses):\n        print(f\"Response {i+1}: {response}\")<\/code><\/pre>\n<p>\u6211\u83b7\u5f97\u4e86\u8d85\u8fc7 100 \u4e2a\u4ee4\u724c\/\u79d2\uff01\u6211\u4e0d\u6562\u76f8\u4fe1\u8fd9\u662f\u4f7f\u7528\u6e38\u620f GPU \u53ef\u4ee5\u5b9e\u73b0\u7684\u3002 GPU \u5229\u7528\u7387\u8fbe\u5230 100%\uff0c\u8fd9\u6b63\u662f\u6211\u60f3\u8981\u7684\uff1a\u83b7\u5f97\u6700\u5927\u6570\u91cf\u7684 GPU\uff08\u56e0\u4e3a\u6211\u652f\u4ed8\u4e86 100% \u7684 GPU \ud83e\udd23\uff09\u3002<\/p>\n<p>  \u5e76\u884c 20 \u4e2a\u8bf7\u6c42\u8fdb\u884c\u63a8\u7406\uff01\uff01\uff01 <\/p>\n<p>\u8fd9\u8fd8\u4e0d\u662f\u6700\u597d\u7684\u90e8\u5206\uff0c\u6211\u4eec\u8bbe\u7f6e\u4e86 <code>--max-num-seq 256<\/code>\uff0c\u6240\u4ee5\u7406\u8bba\u4e0a\u6211\u4eec\u53ef\u4ee5\u5e76\u884c\u53d1\u9001 256 \u4e2a\u8bf7\u6c42\uff01\uff01\u6211\u4e0d\u6562\u76f8\u4fe1\uff0c\u4e5f\u8bb8\u6211\u4ee5\u540e\u4f1a\u5c1d\u8bd5\u8fd9\u4e9b\u6d4b\u8bd5\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u6211\u7684\u6700\u7ec8\u60f3\u6cd5<\/p>\n<h2>3\u3001\u6700\u7ec8\u51b3\u5b9a\u2026\u2026\ufe0f<\/h2>\n<ul>\n<li>\u6027\u80fd\u6982\u8ff0\uff1a\u83b7\u80dc\u8005\u663e\u7136\u662f vLLM\u3002\u6b63\u5982\u6211\u4eec\u5728\u672c\u6587\u7b2c\u4e8c\u90e8\u5206\u4e2d\u770b\u5230\u7684\u90a3\u6837\uff0c\u901a\u8fc7 1 \u4e2a\u8bf7\u6c42\uff0c\u6211\u83b7\u5f97\u4e86 11% \u7684\u6539\u8fdb\uff08Ollama 26 tok\/\u79d2 vs vLLM 29 tok\/\u79d2\uff09\u3002<\/li>\n<li>\u8d44\u6e90\u7ba1\u7406\uff1a\u6beb\u65e0\u7591\u95ee\uff0cvLLM \u662f\u8fd9\u91cc\u7684\u738b\u8005\u3002\u5f53\u6211\u770b\u5230 Ollama \u65e0\u6cd5\u5e76\u884c\u5904\u7406\u8bb8\u591a\u8bf7\u6c42\u65f6\uff0c\u6211\u611f\u5230\u975e\u5e38\u5931\u671b\uff0c\u7531\u4e8e\u8d44\u6e90\u7ba1\u7406\u6548\u7387\u4f4e\u4e0b\uff0c\u5b83\u751a\u81f3\u65e0\u6cd5\u5e76\u884c\u5904\u7406 4 \u4e2a\u8bf7\u6c42\u3002<\/li>\n<li>\u6613\u7528\u6027\u548c\u5f00\u53d1\u6027\uff1a\u6ca1\u6709\u4ec0\u4e48\u6bd4 Ollama \u66f4\u5bb9\u6613\u7684\u4e86\u3002\u5373\u4f7f\u4f60\u4e0d\u662f\u4e13\u5bb6\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528\u4e00\u884c\u4ee3\u7801\u8f7b\u677e\u4e0e LLM \u804a\u5929\u3002\u540c\u65f6\uff0cvLLM \u9700\u8981\u4e00\u4e9b\u77e5\u8bc6\uff0c\u4f8b\u5982 docker \u548c\u66f4\u591a\u53c2\u6570\u3002<\/li>\n<li>\u751f\u4ea7\u5c31\u7eea\u6027\uff1avLLM \u5c31\u662f\u4e3a\u6b64\u800c\u521b\u5efa\u7684\uff0c\u751a\u81f3\u8bb8\u591a\u65e0\u670d\u52a1\u5668\u7aef\u70b9\u63d0\u4f9b\u5546\u516c\u53f8\uff08\u6211\u6709\u6211\u7684\u6765\u6e90\ud83e\udd23\uff09\u90fd\u5728\u5c06\u6b64\u6846\u67b6\u7528\u4e8e\u4ed6\u4eec\u7684\u7aef\u70b9\u3002<\/li>\n<li>\u5b89\u5168\u6027\uff1avLLM \u51fa\u4e8e\u5b89\u5168\u76ee\u7684\u652f\u6301\u4ee4\u724c\u6388\u6743\uff0c\u800c Ollama \u5219\u4e0d\u652f\u6301\u3002\u56e0\u6b64\uff0c\u5982\u679c\u4f60\u6ca1\u6709\u5f88\u597d\u5730\u4fdd\u62a4\u5b83\uff0c\u4efb\u4f55\u4eba\u90fd\u53ef\u4ee5\u8bbf\u95ee\u4f60\u7684\u7aef\u70b9\u3002<\/li>\n<li>\u6587\u6863\uff1a\u8fd9\u4e24\u4e2a\u6846\u67b6\u91c7\u7528\u4e0d\u540c\u7684\u6587\u6863\u65b9\u6cd5\uff1aOllama \u7684\u6587\u6863\u7b80\u5355\u4e14\u9002\u5408\u521d\u5b66\u8005\uff0c\u4f46\u7f3a\u4e4f\u6280\u672f\u6df1\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u6027\u80fd\u548c\u200b\u200b\u5e76\u884c\u5904\u7406\u65b9\u9762\u3002\u4ed6\u4eec\u7684 GitHub \u8ba8\u8bba\u7ecf\u5e38\u7559\u4e0b\u5173\u952e\u95ee\u9898\u672a\u5f97\u5230\u89e3\u7b54\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cvLLM \u63d0\u4f9b\u5168\u9762\u7684\u6280\u672f\u6587\u6863\uff0c\u5176\u4e2d\u5305\u542b\u8be6\u7ec6\u7684 API \u53c2\u8003\u548c\u6307\u5357\u3002\u4ed6\u4eec\u7684 GitHub \u7ef4\u62a4\u826f\u597d\uff0c\u5f00\u53d1\u4eba\u5458\u53cd\u5e94\u8fc5\u901f\uff0c\u6709\u52a9\u4e8e\u6392\u9664\u6545\u969c\u548c\u7406\u89e3\uff0c\u4ed6\u4eec\u751a\u81f3\u4e3a\u6b64\u4e13\u95e8\u8bbe\u7acb\u4e86\u4e00\u4e2a\u7f51\u7ad9\u3002<\/li>\n<\/ul>\n<p>\u6240\u4ee5\uff0c\u4ece\u6211\u7684\u89d2\u5ea6\u6765\u770b\uff0c\u8d62\u5bb6\u662f\u2026\u2026 \u6ca1\u6709\u4e00\u4e2a\uff01<\/p>\n<p>\u5728\u6211\u770b\u6765\uff0c\u5982\u679c\u4f60\u7684\u76ee\u6807\u662f\u5728\u672c\u5730\u73af\u5883\u751a\u81f3\u8fdc\u7a0b\u670d\u52a1\u5668\u4e0a\u5feb\u901f\u8bd5\u9a8c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff0c\u800c\u65e0\u9700\u592a\u591a\u8bbe\u7f6e\u9ebb\u70e6\uff0cOllama \u65e0\u7591\u662f\u4f60\u7684\u9996\u9009\u89e3\u51b3\u65b9\u6848\u3002\u5b83\u7684\u7b80\u5355\u6027\u548c\u6613\u7528\u6027\u4f7f\u5176\u975e\u5e38\u9002\u5408\u5feb\u901f\u539f\u578b\u8bbe\u8ba1\u3001\u6d4b\u8bd5\u60f3\u6cd5\uff0c\u6216\u8005\u9002\u5408\u521a\u5f00\u59cb\u4f7f\u7528 LLM \u5e76\u5e0c\u671b\u5b66\u4e60\u66f2\u7ebf\u5e73\u7f13\u7684\u5f00\u53d1\u4eba\u5458\u3002<\/p>\n<p>\u4f46\u662f\uff0c\u5f53\u6211\u4eec\u5c06\u91cd\u70b9\u8f6c\u79fb\u5230\u6027\u80fd\u3001\u53ef\u6269\u5c55\u6027\u548c\u8d44\u6e90\u4f18\u5316\u81f3\u5173\u91cd\u8981\u7684\u751f\u4ea7\u73af\u5883\u65f6\uff0cvLLM \u663e\u7136\u5927\u653e\u5f02\u5f69\u3002\u5b83\u5bf9\u5e76\u884c\u8bf7\u6c42\u7684\u51fa\u8272\u5904\u7406\u3001\u9ad8\u6548\u7684 GPU \u5229\u7528\u7387\u548c\u5f3a\u5927\u7684\u6587\u6863\u4f7f\u5176\u6210\u4e3a\u4e25\u8083\u3001\u5927\u89c4\u6a21\u90e8\u7f72\u7684\u6709\u529b\u7ade\u4e89\u8005\u3002\u8be5\u6846\u67b6\u4ece\u53ef\u7528\u786c\u4ef6\u8d44\u6e90\u4e2d\u69a8\u53d6\u6700\u5927\u6027\u80fd\u7684\u80fd\u529b\u5c24\u5176\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\uff0c\u5e76\u4e14\u53ef\u80fd\u4f1a\u6539\u53d8\u90a3\u4e9b\u5e0c\u671b\u4f18\u5316\u5176 LLM \u57fa\u7840\u8bbe\u65bd\u7684\u516c\u53f8\u7684\u6e38\u620f\u89c4\u5219\u3002<\/p>\n<p>\u8bdd\u867d\u5982\u6b64\uff0cOllama \u548c vLLM \u4e4b\u95f4\u7684\u51b3\u5b9a\u4e0d\u5e94\u51ed\u7a7a\u800c\u6765\u3002\u5b83\u5fc5\u987b\u53d6\u51b3\u4e8e\u4f60\u7684\u7279\u5b9a\u7528\u4f8b\uff0c\u5e76\u8003\u8651\u4ee5\u4e0b\u56e0\u7d20\uff1a<\/p>\n<ul>\n<li>\u4f60\u7684\u9879\u76ee\u89c4\u6a21<\/li>\n<li>\u4f60\u56e2\u961f\u7684\u6280\u672f\u4e13\u957f<\/li>\n<li>\u5e94\u7528\u7a0b\u5e8f\u7684\u7279\u5b9a\u6027\u80fd\u8981\u6c42<\/li>\n<li>\u4f60\u7684\u5f00\u53d1\u65f6\u95f4\u8868\u548c\u8d44\u6e90<\/li>\n<li>\u5b9a\u5236\u548c\u5fae\u8c03\u7684\u9700\u6c42<\/li>\n<li>\u957f\u671f\u7ef4\u62a4\u548c\u652f\u6301\u6ce8\u610f\u4e8b\u9879<\/li>\n<\/ul>\n<p>\u672c\u8d28\u4e0a\uff0c\u867d\u7136 vLLM \u53ef\u80fd\u4e3a\u751f\u4ea7\u73af\u5883\u63d0\u4f9b\u66f4\u9ad8\u6027\u80fd\u548c\u53ef\u6269\u5c55\u6027\u63d0\u4f9b\u652f\u6301\uff0cOllama \u7684\u7b80\u5355\u6027\u5bf9\u4e8e\u67d0\u4e9b\u573a\u666f\u6765\u8bf4\u53ef\u80fd\u662f\u65e0\u4ef7\u7684\uff0c\u5c24\u5176\u662f\u5728\u5f00\u53d1\u7684\u65e9\u671f\u9636\u6bb5\u6216\u8f83\u5c0f\u89c4\u6a21\u7684\u9879\u76ee\u3002<\/p>\n<p>\u6700\u7ec8\uff0c\u6700\u597d\u7684\u9009\u62e9\u5c06\u662f\u6700\u7b26\u5408\u4f60\u9879\u76ee\u72ec\u7279\u9700\u6c42\u548c\u7ea6\u675f\u7684\u9009\u62e9\u3002\u503c\u5f97\u8003\u8651\u7684\u662f\uff0c\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u4f60\u751a\u81f3\u53ef\u80fd\u53d7\u76ca\u4e8e\u540c\u65f6\u4f7f\u7528\uff1aOllama \u7528\u4e8e\u5feb\u901f\u539f\u578b\u8bbe\u8ba1\u548c\u521d\u59cb\u5f00\u53d1\uff0c\u800c vLLM \u5219\u7528\u4e8e\u4f60\u51c6\u5907\u6269\u5c55\u548c\u4f18\u5316\u751f\u4ea7\u3002\u8fd9\u79cd\u6df7\u5408\u65b9\u6cd5\u53ef\u4ee5\u4e3a\u4f60\u63d0\u4f9b\u4e24\u5168\u5176\u7f8e\u7684\u6548\u679c\uff0c\u8ba9\u4f60\u53ef\u4ee5\u5728\u9879\u76ee\u751f\u547d\u5468\u671f\u7684\u4e0d\u540c\u9636\u6bb5\u5229\u7528\u6bcf\u4e2a\u6846\u67b6\u7684\u4f18\u52bf\u3002<\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u6b22\u8fce\u6765\u5230\u6211\u4eec\u6df1\u5165\u7814\u7a76 LLM \u63a8\u7406\u6846\u67b6\u7684\u6700\u540e\u4e00\u90e8\u5206\uff01\u5728\u7b2c\u4e00\u90e8\u5206\u548c\u7b2c\u4e8c\u90e8\u5206\u4e2d\uff0c\u6211\u4eec\u5206\u522b\u63a2\u8ba8\u4e86 Ollama \u548c vLLM\uff0c\u4e86\u89e3\u4e86\u5b83\u4eec\u7684\u67b6\u6784\u3001\u529f\u80fd\u548c\u57fa\u672c\u6027\u80fd\u7279\u5f81\u3002\u73b0\u5728\u5230\u4e86\u51b3\u5b9a\u6027\u7684\u4e00\u8f6e\uff1a\u9762\u5bf9\u9762\u7684\u6bd4\u8f83\uff0c\u4ee5\u5e2e\u52a9\u60a8\u6839\u636e\u7279\u5b9a\u9700\u6c42\u9009\u62e9\u5408\u9002\u7684\u6846\u67b6\u3002 \u8fd9\u6b21\u6bd4\u8f83\u5e76\u4e0d\u662f\u8981\u5ba3\u5e03\u7edd\u5bf9\u7684\u8d62\u5bb6\u2014\u2014\u800c\u662f\u8981\u4e86\u89e3\u54ea\u79cd\u6846\u67b6\u5728\u4e0d\u540c\u573a\u666f\u4e2d\u8868\u73b0\u51fa\u8272\u3002\u6211\u4eec\u5c06\u91cd\u70b9\u5173\u6ce8\uff1a \u8d44\u6e90\u5229\u7528\u7387\u548c\u6548\u7387 \u90e8\u7f72\u548c\u7ef4\u62a4\u7684\u7b80\u6613\u6027 \u5177\u4f53\u7528\u4f8b\u548c\u5efa\u8bae \u5b89\u5168\u548c\u751f\u4ea7\u51c6\u5907 \u6587\u6863 \u8ba9\u6211\u4eec\u6df1\u5165\u7814\u7a76\u6570\u636e\uff0c\u770b\u770b\u6211\u4eec\u7684\u6d4b\u8bd5\u63ed\u793a\u4e86\u4ec0\u4e48\uff01 \u53ea\u6709\u4e00\u4e2a\u53ef\u4ee5\u6210\u4e3a\u51a0\u519b\uff0c\u6216\u8005\u53ef\u80fd\u4e0d\u662f\uff1f \ud83e\udd14 1\u3001\u57fa\u51c6\u6d4b\u8bd5\u8bbe\u7f6e \u4e3a\u4e86\u786e\u4fdd\u516c\u5e73\u6bd4\u8f83\uff0c\u6211\u4eec\u5c06\u5bf9\u4e24\u4e2a\u6846\u67b6\u4f7f\u7528\u76f8\u540c\u7684\u786c\u4ef6\u548c\u6a21\u578b\uff1a \u786c\u4ef6\u914d\u7f6e\uff1a GPU\uff1aNVIDIA RTX 4060 16GB Ti RAM\uff1a64GB RAM CPU\uff1aAMD Ryzen 7 \u5b58\u50a8\uff1aNVMe SSD \u578b\u53f7\uff1a Qwen2.5\u201314B-Instruct\uff084 \u4f4d\u91cf\u5316\uff09 \u4e0a\u4e0b\u6587\u957f\u5ea6\uff1a8192 \u4e2a\u6807\u8bb0 \u6279\u5904\u7406\u5927\u5c0f\uff1a1\uff08\u5355\u7528\u6237\u573a\u666f\uff09 2\u3001\u975e\u5e38\u516c\u5e73\u7684\u6bd4\u8f83 \u8ba9\u6211\u4eec\u5206\u6790\u4e00\u4e0b\u8fd9\u4e24\u4e2a\u6846\u67b6\u5982\u4f55\u4ee5\u4e0d\u540c\u7684\u65b9\u5f0f\u7ba1\u7406\u7cfb\u7edf\u8d44\u6e90\uff0c\u91cd\u70b9\u5173\u6ce8\u5b83\u4eec\u7684\u6838\u5fc3\u67b6\u6784\u65b9\u6cd5\u548c\u5b9e\u9645\u5f71\u54cd\u3002 2.1 Ollama \u6211\u4e3e\u4e86\u4e00\u4e2a\u95ee\u9898\u201c\u7ed9\u6211\u8bb2\u4e00\u4e2a 1000 \u5b57\u7684\u6545\u4e8b\u201d\u7684\u4f8b\u5b50\u3002\u6211\u4e00\u4e2a\u8bf7\u6c42\u7684 tok\/sec \u4e3a 25.59\u3002\u6ca1\u6709\u5e76\u884c\u8bf7\u6c42 \u95ee\u9898\uff1a\u201c\u7ed9\u6211\u8bb2\u4e00\u4e2a 1000 \u5b57\u7684\u6545\u4e8b\u201d\u7528\u4e8e Ollama \u5bf9\u4e8e\u5e76\u884c\u8bf7\u6c42\uff0c\u7528\u6237\u5fc5\u987b\u4fee\u6539\u4f4d\u4e8e \/etc\/systemd\/system\/ollama.service \u4e2d\u7684\u6587\u4ef6\uff08\u5bf9\u4e8e Ubuntu\uff09\u5e76\u6dfb\u52a0\u4e00\u884c Environment=&#8221;OLLAMA_NUM_PARALLEL=4&#8243;\uff0c\u4f60\u5c06\u88ab\u5141\u8bb8\u6267\u884c\u6700\u591a [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53817","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53817","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53817"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53817\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53817"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53817"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53817"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}