{"id":938,"date":"2026-10-07T16:21:37","date_gmt":"2026-10-07T08:21:37","guid":{"rendered":"https:\/\/cqiot.cc\/?p=938"},"modified":"2026-10-07T16:21:37","modified_gmt":"2026-10-07T08:21:37","slug":"vllm-vs-llamacpp-32b-24gb","status":"publish","type":"post","link":"https:\/\/cqiot.cc\/index.php\/2026\/10\/07\/vllm-vs-llamacpp-32b-24gb\/","title":{"rendered":"24GB \u663e\u5361\u8dd1 32B \u91cf\u5316\u6a21\u578b\uff1avLLM \u4e0e llama.cpp \u9009\u578b\u5bf9\u6bd4"},"content":{"rendered":"

<\/p>\n

\n

\u5185\u5bb9\u6807\u8bc6\uff1a<\/strong>\u672c\u6587\u4e3a AI \u8f85\u52a9\u751f\u6210\u5185\u5bb9\u3002\u4f9d\u636e \u300a\u4eba\u5de5\u667a\u80fd\u751f\u6210\u5408\u6210\u5185\u5bb9\u6807\u8bc6\u529e\u6cd5\u300b2025-09-01\uff0c\u5df2\u6dfb\u52a0\u663e\u5f0f\u6807\u8bc6\uff08\u672c\u58f0\u660e\uff09\u4e0e\u9690\u5f0f\u6807\u8bc6\uff08\u6587\u4ef6\u5143\u6570\u636e \/ \u56fe\u7247 ComfyUI \u5143\u4fe1\u606f\uff09\u3002\u6587\u4e2d\u6280\u672f\u7ec6\u8282\u4e0e\u7ed3\u8bba\u5747\u7531\u4f5c\u8005\u590d\u6838\u786e\u8ba4\u3002<\/p>\n<\/div>\n

\u6211\u4eec\u7528 Ollama \u5728 PVE \u865a\u62df\u673a\u4e0a\u8dd1\u901a\u4e86\u672c\u5730\u5927\u6a21\u578b\uff08\u53c2\u89c1\u7528 Ollama \u5728 PVE \u865a\u62df\u673a\u91cc\u8dd1\u672c\u5730\u5927\u6a21\u578b\uff1a\u663e\u5361\u76f4\u901a\u4e0e\u663e\u5b58\u89c4\u5212<\/a>\uff09\u3002\u4f46\u9700\u6c42\u5347\u5230 32B \u7ea7\u522b\u3001\u8981\u591a\u4eba\u5e76\u53d1\u8c03\u7528\u65f6\uff0cOllama \u5c31\u4e0d\u591f\u7528\u4e86\uff0c\u4e3b\u6d41\u9009\u62e9\u662f vLLM<\/strong>\uff08\u9ad8\u5e76\u53d1\u670d\u52a1\u5316\uff09\u4e0e llama.cpp<\/strong>\uff08\u4f4e\u8d44\u6e90\u5355\u673a\uff09\u3002<\/p>\n

\u5f88\u591a\u4eba\u9ed8\u8ba4\u8fd9\u662f\u4e00\u9053\u81ea\u7531\u9009\u62e9\u9898\uff0c\u4e8c\u9009\u4e00\u5373\u53ef\u3002\u4f46\u52a8\u624b\u524d\u5fc5\u987b\u5148\u786e\u8ba4\u4e00\u4ef6\u4e8b\uff1a\u4f60\u90a3\u5f20 24GB \u5361\u5c5e\u4e8e\u54ea\u4e00\u4ee3\u67b6\u6784<\/strong>\u2014\u2014\u8fd9\u4e00\u6761\u51e0\u4e4e\u76f4\u63a5\u51b3\u5b9a\u6846\u67b6\u80fd\u4e0d\u80fd\u8dd1\u8d77\u6765\u3002<\/p>\n

\u6d4b\u8bd5\u73af\u5883\uff1a\u4e00\u53f0 HP Z440 \u5de5\u4f5c\u7ad9 + \u4e09\u5f20\u8001\u5361<\/h2>\n

\u672c\u6587\u7684\u73af\u5883\u5c31\u662f\u624b\u8fb9\u8fd9\u53f0 HP Z440 \u5854\u5f0f\u5de5\u4f5c\u7ad9\uff0c\u90a3\u5f20 24GB \u7684\u5361\u662f NVIDIA Quadro M6000\u3002<\/p>\n

\"NVIDIA
\n\u56fe 1\uff1aNVIDIA Quadro M6000 24GB \u663e\u5361\u3002\u56fe\u7247\u6765\u6e90\uff1aNVIDIA \u5b98\u65b9\u4ea7\u54c1\u8d44\u6599\uff0c\u4ec5\u4f5c\u578b\u53f7\u5916\u89c2\u793a\u610f\uff08\u975e\u672c\u673a\u5b9e\u62cd\uff09\u3002<\/em><\/p>\n

\"\u6d4b\u8bd5\u4e3b\u673a
\n\u56fe 2\uff1a\u6d4b\u8bd5\u4e3b\u673a HP Z440 \u5854\u5f0f\u5de5\u4f5c\u7ad9\u3002\u56fe\u7247\u6765\u6e90\uff1a\u4e2d\u5173\u6751\u5728\u7ebf\uff08ZOL\uff09\u4ea7\u54c1\u56fe\uff0c\u4ec5\u4f5c\u578b\u53f7\u5916\u89c2\u793a\u610f\u3002<\/em><\/p>\n\n\n\n\n\n\n\n\n
\u90e8\u4ef6<\/th>\n\u578b\u53f7<\/th>\n\u8ba1\u7b97\u80fd\u529b<\/th>\n\u89d2\u8272<\/th>\n<\/tr>\n<\/thead>\n
\u4e3b\u673a<\/td>\nHP Z440 \u5854\u5f0f\u5de5\u4f5c\u7ad9\uff08Xeon E5 v3\/v4 \u5e73\u53f0\u3001C612 \u82af\u7247\u7ec4\uff09<\/td>\n\u2014<\/td>\n\u6d4b\u8bd5\u5e73\u53f0<\/td>\n<\/tr>\n
\u4e3b\u89d2\u663e\u5361<\/td>\nNVIDIA Quadro M6000 24GB<\/td>\n5.2\uff08Maxwell\uff09<\/td>\n\u672c\u6587\u8ba8\u8bba\u7684 24GB \u5361<\/td>\n<\/tr>\n
\u540c\u673a\u663e\u5361<\/td>\nNVIDIA GTX 1050<\/td>\n6.1\uff08Pascal\uff09<\/td>\n\u4eae\u673a \/ \u8f7b\u91cf\u4efb\u52a1<\/td>\n<\/tr>\n
\u540c\u673a\u663e\u5361<\/td>\nNVIDIA Tesla P4 8GB<\/td>\n6.1\uff08Pascal\uff09<\/td>\n\u88ab\u52a8\u6563\u70ed\u8ba1\u7b97\u5361<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n


\n
\n<\/p>\n

\u8fd9\u5f20\u8868\u91cc\u57cb\u7740\u4e00\u4e2a\u5bf9\u9009\u578b\u5f88\u5173\u952e\u7684\u7ec6\u8282\uff1a\u4e09\u5f20\u5361\u7684\u8ba1\u7b97\u80fd\u529b\u5168\u90e8\u4f4e\u4e8e 7.0<\/strong>\uff085.2 \/ 6.1 \/ 6.1\uff09\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u8fd9\u53f0\u673a\u5668\u4e0a\u65e0\u8bba\u6362\u7528\u54ea\u5f20\u5361\uff0c\u90fd\u8fc8\u4e0d\u8fc7 vLLM \u7684\u95e8\u69db\u3002<\/p>\n

\u4f5c\u4e3a\u5bf9\u7167\uff0c\u4e0a\u4e00\u7bc7\u6587\u7ae0\u91cc\u7528\u4e8e\u56fe\u7247\u751f\u6210\u5b9e\u6d4b\u7684\u90a3\u53f0\u673a\u5668\u662f RTX 4070 Ti SUPER 16G\uff08Ada \u67b6\u6784\u3001\u7b97\u529b 8.9\uff09\uff0c\u5c5e\u4e8e vLLM \u652f\u6301\u7684\u8303\u56f4\u3002<\/p>\n

<\/p>\n

\u5148\u7ea0\u6b63\u4e00\u4e2a\u5e38\u89c1\u8bef\u5224\uff1aM6000 \u662f Maxwell\uff0c\u4e0d\u662f Pascal<\/h2>\n

Quadro M6000 24GB \u5e38\u88ab\u548c Pascal \u6df7\u4e3a\u4e00\u8c08\uff08\u6211\u81ea\u5df1\u4e5f\u4e00\u76f4\u8bb0\u6210 Pascal\uff09\u3002\u5b83\u7684\u771f\u5b9e\u8eab\u4efd\u662f GM200 \u6838\u5fc3\u3001Maxwell 2.0 \u67b6\u6784\uff0c\u8ba1\u7b97\u80fd\u529b\uff08Compute Capability\uff095.2<\/strong>\uff0c2016 \u5e74 3 \u6708\u53d1\u5e03\uff0c24GB GDDR5\u3001384-bit \u4f4d\u5bbd\u3001\u5e26\u5bbd\u7ea6 317 GB\/s\u3002<\/p>\n

<\/p>\n

\u8fd9\u4e2a\u5dee\u5f02\u662f\u9009\u578b\u5206\u6c34\u5cad\uff1a<\/p>\n\n\n\n\n\n\n\n
\u67b6\u6784<\/th>\n\u8ba1\u7b97\u80fd\u529b<\/th>\n\u4ee3\u8868\u5361<\/th>\n\u5bf9\u672c\u6587\u7684\u610f\u4e49<\/th>\n<\/tr>\n<\/thead>\n
Maxwell<\/td>\n5.0 \/ 5.2<\/td>\nM6000 24G\u3001GTX 9xx<\/td>\nllama.cpp \u53ef\u7528\uff0cvLLM \u4e0d\u652f\u6301<\/td>\n<\/tr>\n
Pascal<\/td>\n6.0 \/ 6.1<\/td>\nGTX 10xx\u3001Tesla P40<\/td>\nllama.cpp \u53ef\u7528\uff0cvLLM \u9700\u6539\u6e90\u7801<\/td>\n<\/tr>\n
Volta \u53ca\u4ee5\u540e<\/td>\n\u2265 7.0<\/td>\nV100\u3001T4\u3001RTX 20xx+<\/td>\n\u4e24\u5957\u6846\u67b6\u90fd\u80fd\u7528<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

vLLM \u5728 M6000 \u4e0a\u76f4\u63a5\u51fa\u5c40<\/h2>\n

vLLM \u5b98\u65b9\u5b89\u88c5\u6587\u6863\u5bf9 GPU \u7684\u8981\u6c42\u5f88\u76f4\u767d\uff1a\u8ba1\u7b97\u80fd\u529b 7.0 \u6216\u66f4\u9ad8<\/strong>\uff0c\u793a\u4f8b\u5217\u4e3e V100\u3001T4\u3001RTX 20xx\u3001A100\u3001L4\u3001H100\u3002<\/p>\n

<\/p>\n

\u4e5f\u5c31\u662f\u8bf4\uff0c\u9884\u7f16\u8bd1\u8f6e\u5b50\u8986\u76d6\u7684\u662f Volta\uff087.0\uff09\u53ca\u66f4\u65b0\u7684\u5361\uff0cPascal \u53ca\u66f4\u65e9\uff08Maxwell\uff09\u4e0d\u5728\u652f\u6301\u8303\u56f4\u3002\u793e\u533a\u786e\u5b9e\u6709\u6539\u6e90\u7801\u8ba9 vLLM \u8dd1 Pascal \u7684\u505a\u6cd5\uff0c\u4f46 M6000 \u8fd8\u8981\u518d\u5f80\u524d\u9000\u4e00\u4ee3\uff0c\u6539\u9020\u6295\u5165\u4e0e\u5b9e\u9645\u6536\u76ca\u4e0d\u6210\u6bd4\u4f8b\u3002<\/p>\n

\u7ed3\u8bba\uff1a\u5728 M6000 \u4e0a\uff0c\u300cvLLM \u4e0e llama.cpp \u5bf9\u6bd4\u300d\u8fd9\u4e2a\u547d\u9898\u5e76\u4e0d\u6210\u7acb<\/strong>\uff0cvLLM \u5728\u7b2c\u4e00\u5173\u5c31\u88ab\u786c\u4ef6\u95e8\u69db\u62e6\u4e0b\u3002\u975e\u8981\u7528 vLLM\uff0c\u5c31\u5f97\u6362\u4e00\u5f20\u8ba1\u7b97\u80fd\u529b \u2265 7.0 \u7684\u5361\uff08\u5982\u4e8c\u624b T4 \/ RTX 2080Ti\uff09\u3002<\/p>\n

llama.cpp \u662f M6000 \u4e0a\u552f\u4e00\u73b0\u5b9e\u7684\u9009\u62e9<\/h2>\n

llama.cpp \u7684 CUDA \u540e\u7aef\u8981\u6c42\u5bbd\u677e\u5f97\u591a\uff1a\u8ba1\u7b97\u80fd\u529b \u2265 5.0\uff08Maxwell \u6216\u66f4\u65b0\uff09<\/strong>\u5373\u53ef\uff0cMaxwell \u5728\u5b98\u65b9\u652f\u6301\u8868\u4e2d\u88ab\u6807\u6ce8\u4e3a Legacy\uff08\u4f20\u7edf\u652f\u6301\uff09\u3002<\/p>\n

<\/p>\n

\u5173\u952e\u662f\u663e\u5f0f\u6307\u5b9a\u7f16\u8bd1\u67b6\u6784<\/strong>\uff0c\u5426\u5219\u9ed8\u8ba4 fatbin \u91cc\u53ef\u80fd\u6ca1\u6709 sm_52 \u539f\u751f\u4ee3\u7801\uff0c\u8fd0\u884c\u65f6\u9760 PTX JIT \u7ffb\u8bd1\uff0c\u901f\u5ea6\u4f1a\u660e\u663e\u5403\u4e8f\uff1a<\/p>\n

git clone https:\/\/github.com\/ggml-org\/llama.cpp\ncd llama.cpp\ncmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="52"\ncmake --build build --config Release -j$(nproc)<\/code><\/pre>\n
<\/p>\n
\u8fd8\u6709\u4e00\u6761\u6613\u88ab\u5ffd\u7565\u7684\u786c\u7ea6\u675f\uff1aCUDA \u5de5\u5177\u94fe\u5fc5\u987b\u505c\u7559\u5728 12.x<\/strong>\u3002Maxwell\/Pascal\/Volta \u5df2\u88ab\u5217\u4e3a\u5f03\u7528\u67b6\u6784\uff0cCUDA 13.x \u7684\u6700\u4f4e\u8981\u6c42\u662f Turing\uff08sm_75\uff09\uff0c\u4e3a M6000 \u7f16\u8bd1\u4e0d\u80fd\u518d\u5347\u5230 13\u3002<\/p>\n
<\/p>\n
\u542f\u52a8\u670d\u52a1\u53ea\u9700\u4e00\u6761\u547d\u4ee4\uff08\u81ea\u5e26 OpenAI \u517c\u5bb9\u63a5\u53e3\uff09\uff1a<\/p>\n
.\/build\/bin\/llama-server \n  -m \/data\/models\/qwen2.5-32b-instruct-q4_k_m.gguf \n  -ngl 99 -c 8192 --port 8080<\/code><\/pre>\n
<\/p>\n
\u6027\u80fd\u9884\u671f\uff1a\u5e26\u5bbd\u662f\u786c\u7ea6\u675f<\/h2>\n
\u81ea\u56de\u5f52\u89e3\u7801\u662f\u5178\u578b\u7684\u663e\u5b58\u5e26\u5bbd\u74f6\u9888<\/strong>\u4efb\u52a1\u3002M6000 \u5e26\u5bbd\u7ea6 317 GB\/s\uff0cRTX 3090 \u7ea6 936 GB\/s\uff0c\u524d\u8005\u53ea\u6709\u540e\u8005\u4e09\u5206\u4e4b\u4e00\u3002\u6309\u6b64\u7c97\u4f30\uff0c32B Q4_K_M \u5728 M6000 \u4e0a\u7684\u51fa\u8bcd\u901f\u5ea6\u5927\u7ea6\u5728\u6bcf\u79d2\u51e0\u4e2a token \u5230\u5341\u4e2a token \u4e4b\u95f4<\/strong>\uff0c\u9996 token \u5ef6\u8fdf\u4e5f\u66f4\u957f\u3002<\/p>\n
\u8981\u5f3a\u8c03\uff1a\u4ee5\u4e0a\u662f\u6309\u5e26\u5bbd\u6bd4\u63a8\u7b97\u7684\u4f30\u7b97\u503c\uff0c\u5e76\u975e\u672c\u673a\u5b9e\u6d4b\u6570\u5b57<\/strong>\u3002CUDA \u7248\u672c\u3001\u662f\u5426\u547d\u4e2d\u539f\u751f sm_52 \u4ee3\u7801\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u90fd\u4f1a\u663e\u8457\u5f71\u54cd\u7ed3\u679c\uff0c\u8bf7\u4ee5\u672c\u673a llama-bench<\/code> \u5b9e\u6d4b\u4e3a\u51c6\u3002<\/p>\n
<\/p>\n
\u9009\u578b\u5efa\u8bae<\/h2>\n\n\n\n\n\n\n\n
\u573a\u666f<\/th>\n\u63a8\u8350<\/th>\n\u7406\u7531<\/th>\n<\/tr>\n<\/thead>\n
M6000\uff08Maxwell 5.2\uff0924G<\/td>\nllama.cpp<\/td>\nvLLM \u6709 7.0 \u7b97\u529b\u95e8\u69db\uff0c\u8dd1\u4e0d\u8d77\u6765<\/td>\n<\/tr>\n
\u5355\u4eba\u81ea\u7528\u3001\u8ffd\u6c42\u7701\u4e8b<\/td>\nllama.cpp<\/td>\n\u90e8\u7f72\u7b80\u5355\uff0cGGUF \u751f\u6001\u4e30\u5bcc<\/td>\n<\/tr>\n
\u591a\u4eba\u5e76\u53d1 API \u670d\u52a1<\/td>\nvLLM<\/td>\n\u9700\u8ba1\u7b97\u80fd\u529b \u2265 7.0 \u7684\u5361\uff0c\u8fde\u7eed\u6279\u5904\u7406\u5e26\u6765\u6570\u500d\u541e\u5410<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
<\/p>\n
\u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a\u6846\u67b6\u6ca1\u6709\u7edd\u5bf9\u4f18\u52a3\uff0c\u53ea\u6709\u573a\u666f\u5339\u914d<\/strong>\u3002M6000 \u8fd9\u7c7b\u8001\u5361\u7684\u4ef7\u503c\u662f\u300c24GB \u663e\u5b58\u4fbf\u5b9c\u300d\uff0c\u4ee3\u4ef7\u662f\u67b6\u6784\u4ee3\u5dee\u5e26\u6765\u7684\u517c\u5bb9\u6027\u6536\u7a84\u2014\u2014\u5148\u786e\u8ba4\u67b6\u6784\u4ee3\u6b21\uff0c\u518d\u9009\u6846\u67b6\uff0c\u80fd\u7701\u4e0b\u5927\u91cf\u6298\u817e\u3002\u56fe\u7247\u751f\u6210\u73af\u5883\u642d\u5efa\u53ef\u53c2\u8003\u672c\u7ad9\u7684 ComfyUI \u672c\u5730\u90e8\u7f72\u6307\u5357<\/a>\uff0c\u601d\u8def\u76f8\u901a\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u624b\u4e0a\u6709 24GB \u663e\u5b58\u7684 Quadro M6000\uff0c\u60f3\u8dd1 32B \u91cf\u5316\u6a21\u578b\uff0c\u8be5\u9009 vLLM \u8fd8\u662f llama.cpp\uff1f\u672c\u6587\u6838\u5b9e\u4e86\u4e24\u5957\u6846\u67b6\u7684\u786c\u4ef6\u95e8\u69db\uff0c\u7ed9\u51fa M6000\uff08Maxwell \u67b6\u6784\uff09\u4e0a\u7684\u53ef\u884c\u6027\u7ed3\u8bba\u4e0e\u53ef\u590d\u73b0\u7684\u7f16\u8bd1\u6b65\u9aa4\u3002<\/p>\n","protected":false},"author":1,"featured_media":940,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_kad_post_transparent":"","_kad_post_title":"","_kad_post_layout":"","_kad_post_sidebar_id":"","_kad_post_content_style":"","_kad_post_vertical_padding":"","_kad_post_feature":"","_kad_post_feature_position":"","_kad_post_header":false,"_kad_post_footer":false,"_kad_post_classname":"","footnotes":""},"categories":[46],"tags":[64,63,65],"class_list":["post-938","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-local-ai","tag-llama-cpp","tag-vllm","tag-65"],"_links":{"self":[{"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/posts\/938","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/comments?post=938"}],"version-history":[{"count":5,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/posts\/938\/revisions"}],"predecessor-version":[{"id":946,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/posts\/938\/revisions\/946"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/media\/940"}],"wp:attachment":[{"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/media?parent=938"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/categories?post=938"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/cqiot.cc\/index.php\/wp-json\/wp\/v2\/tags?post=938"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}