{"id":125,"date":"2026-08-26T23:42:51","date_gmt":"2026-08-26T15:42:51","guid":{"rendered":"https:\/\/www.xiaoxindong.cloud\/?p=125"},"modified":"2026-09-06T19:55:07","modified_gmt":"2026-09-06T11:55:07","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2","status":"publish","type":"post","link":"https:\/\/www.xiaoxindong.cloud\/?p=125","title":{"rendered":"\u5927\u6a21\u578b\u672c\u5730\u90e8\u7f72"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u4e00\u3001GPU \u9a71\u52a8\u4e0e CUDA \u5b89\u88c5<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. GPU \u9a71\u52a8\u6838\u5fc3\u8ba4\u77e5<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u672c\u8d28<\/strong>\uff1a\u64cd\u4f5c\u7cfb\u7edf\u4e0e GPU \u786c\u4ef6\u4e4b\u95f4\u7684 \u201c\u7ffb\u8bd1\u5b98\u201d\uff0c\u8d1f\u8d23\u8ba9\u7cfb\u7edf\u8bc6\u522b\u3001\u8c03\u7528\u5e76\u63a7\u5236\u663e\u5361<\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u529f\u80fd<\/strong>\uff1a\u8bc6\u522b\u786c\u4ef6\u4fe1\u606f\uff08\u6e29\u5ea6 \/ \u663e\u5b58 \/ \u529f\u8017\uff09\u3001\u7ba1\u7406\u663e\u5b58\u5206\u914d\u56de\u6536\u3001\u63a7\u5236\u8fd0\u884c\u72b6\u6001\uff08\u98ce\u6247 \/ \u65f6\u949f \/ ECC\/MIG\/ \u591a\u5361\u901a\u4fe1\uff09<\/li>\n\n\n\n<li><strong>\u65e0\u9a71\u52a8\u540e\u679c<\/strong>\uff1a<code>nvidia-smi<\/code> \u4e0d\u53ef\u7528\u3001CUDA \u5931\u6548\u3001PyTorch \u68c0\u6d4b\u4e0d\u5230 GPU\u3001\u4ec5\u57fa\u7840\u663e\u793a\u65e0\u52a0\u901f\uff0c\u663e\u5361\u6027\u80fd\u5b8c\u5168\u65e0\u6cd5\u53d1\u6325<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u9a71\u52a8\u5b89\u88c5\u65b9\u5f0f\u5bf9\u6bd4<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u5b89\u88c5\u65b9\u5f0f<\/th><th>\u662f\u5426\u8054\u7f51<\/th><th>\u5305\u542b\u5185\u5bb9<\/th><th>\u7248\u672c\u7ba1\u7406<\/th><th>\u9002\u7528\u573a\u666f<\/th><\/tr><\/thead><tbody><tr><td>APT \u5728\u7ebf\u5b89\u88c5<\/td><td>\u662f<\/td><td>\u4ec5\u9a71\u52a8\u53ca\u7cfb\u7edf\u4f9d\u8d56<\/td><td>\u4ed3\u5e93\u7edf\u4e00\u7ba1\u7406\uff0c\u6613\u5347\u7ea7 \/ \u56de\u6eda<\/td><td>\u6709\u5916\u7f51\u7684\u751f\u4ea7 \/ \u5f00\u53d1\u73af\u5883\u3001\u6279\u91cf\u81ea\u52a8\u5316\u90e8\u7f72<\/td><\/tr><tr><td>run \u79bb\u7ebf\u5b89\u88c5<\/td><td>\u5426<\/td><td>\u9a71\u52a8 + CUDA Toolkit + \u8fd0\u884c\u5e93 + \u793a\u4f8b<\/td><td>\u624b\u52a8\u7ba1\u7406\uff0c\u9700\u91cd\u65b0\u4e0b\u8f7d\u5b89\u88c5\u5305<\/td><td>\u65e0\u5916\u7f51\u3001\u4e34\u65f6\u73af\u5883\u3001\u5feb\u901f\u5b89\u88c5\u6307\u5b9a\u7248\u672c<\/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\">\u56fd\u5185\u751f\u4ea7\u73af\u5883\u8865\u5145\u65b9\u6848<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u81ea\u5efa\u5185\u7f51 APT \u4ed3\u5e93<\/strong>\uff1a\u4e0b\u8f7d\u4e00\u6b21\u5185\u7f51\u590d\u7528\uff0c\u5b89\u5168\u53ef\u63a7\u3001\u53ef\u5ba1\u8ba1\uff0c\u9002\u5408\u5927\u89c4\u6a21\u96c6\u7fa4\u7edf\u4e00\u90e8\u7f72<\/li>\n\n\n\n<li><strong>\u6d77\u5916\u4e13\u7ebf\u5b89\u88c5<\/strong>\uff1a\u65e0\u9700\u81ea\u5efa\u4ed3\u5e93\uff0c\u5b9e\u65f6\u83b7\u53d6\u6700\u65b0\u7248\uff0c\u9002\u5408\u5c0f\u89c4\u6a21\u3001\u6709\u66f4\u65b0\u9700\u6c42\u7684\u73af\u5883<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<h3 class=\"wp-block-heading\">3. APT \u5728\u7ebf\u5b89\u88c5\u6b65\u9aa4\uff08Ubuntu\uff09<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u786c\u4ef6\u68c0\u67e5<\/strong>\uff1a<code>lspci | grep -i tesla<\/code> \u786e\u8ba4\u7269\u7406\u663e\u5361\u5b58\u5728<\/li>\n\n\n\n<li><strong>\u67e5\u770b\u652f\u6301\u9a71\u52a8<\/strong>\uff1a<code>ubuntu-drivers devices<\/code><\/li>\n\n\n\n<li><strong>\u5b89\u88c5\u9a71\u52a8<\/strong>\uff1a\u6dfb\u52a0 NVIDIA \u5b98\u65b9\u6e90 \u2192 <code>apt install -y nvidia-driver-535<\/code><\/li>\n\n\n\n<li><strong>\u91cd\u542f\u751f\u6548<\/strong>\uff1a<code>reboot<\/code><\/li>\n\n\n\n<li><strong>\u5b89\u88c5\u9a8c\u8bc1<\/strong>\uff1a<code>nvidia-smi<\/code><\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">4. SXM \u673a\u578b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">SXM \u591a\u5361 NVSwitch \u67b6\u6784\u670d\u52a1\u5668\uff0c\u5b89\u88c5\u9a71\u52a8\u540e<strong>\u5fc5\u987b\u5b89\u88c5\u5bf9\u5e94\u7248\u672c\u7684 Fabric Manager<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f5c\u7528\uff1a\u521d\u59cb\u5316\u548c\u7ba1\u7406 NVSwitch\/NVLink Fabric<\/li>\n\n\n\n<li>\u8981\u6c42\uff1a\u7248\u672c\u5fc5\u987b\u4e0e\u9a71\u52a8\u7248\u672c\u4e25\u683c\u5339\u914d<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">5. CUDA Toolkit<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5b9a\u4e49<\/strong>\uff1aNVIDIA \u5b98\u65b9 GPU \u5f00\u53d1\u73af\u5883\uff0c\u6838\u5fc3\u5305\u542b NVCC \u7f16\u8bd1\u5668\u3001CUDA \u8fd0\u884c\u65f6\u5e93\u3001cuDNN \u6df1\u5ea6\u5b66\u4e60\u5e93\u3001\u8c03\u8bd5\u5206\u6790\u5de5\u5177<\/li>\n\n\n\n<li><strong>\u7528\u9014<\/strong>\uff1a\u5e76\u884c\u8ba1\u7b97\u3001\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u63a8\u7406\u3001\u56fe\u5f62\u6e32\u67d3\u3001\u89c6\u9891\u56fe\u50cf\u5904\u7406<\/li>\n\n\n\n<li><strong>\u5b89\u88c5\u6b65\u9aa4<\/strong>\uff1a\n<ol class=\"wp-block-list\">\n<li>\u4e0b\u8f7d\u5bf9\u5e94\u7248\u672c <code>.run<\/code> \u5b89\u88c5\u5305<\/li>\n\n\n\n<li><code>chmod +x cuda_xxx.run<\/code> \u6dfb\u52a0\u6267\u884c\u6743\u9650<\/li>\n\n\n\n<li>\u8fd0\u884c\u5b89\u88c5\uff0c<strong>\u53d6\u6d88 Driver \u52fe\u9009<\/strong>\uff08\u9a71\u52a8\u5df2\u5355\u72ec\u5b89\u88c5\uff09\uff0c\u4ec5\u5b89\u88c5 CUDA Toolkit<\/li>\n\n\n\n<li>\u914d\u7f6e\u73af\u5883\u53d8\u91cf\uff08<code>~\/.bashrc<\/code> \u4e2d\u6dfb\u52a0 PATH \u548c LD_LIBRARY_PATH\uff09<\/li>\n\n\n\n<li>\u9a8c\u8bc1\uff1a<code>source ~\/.bashrc<\/code> \u540e\u6267\u884c <code>nvcc --version<\/code><\/li>\n<\/ol>\n<\/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\">\u9a71\u52a8\u4e0e CUDA \u7248\u672c\u5bf9\u5e94<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9a71\u52a8 535 \u2192 CUDA 12.2\uff1b\u9a71\u52a8 570 \u2192 CUDA 12.8\uff1b\u9a71\u52a8 580 \u2192 CUDA 13.0<\/li>\n\n\n\n<li>\u517c\u5bb9\u539f\u5219\uff1a<strong>\u5411\u4e0b\u517c\u5bb9<\/strong>\uff0c\u6700\u7ec8\u4ee5 NVIDIA \u5b98\u65b9\u517c\u5bb9\u6027\u6587\u6863\u4e3a\u51c6<\/li>\n<\/ul>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001GPU \u76d1\u63a7\u4e0e\u7ba1\u7406\u547d\u4ee4<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. nvidia-smi\uff08\u2b50\uff09<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u57fa\u7840\u547d\u4ee4<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u67e5\u770b\u6240\u6709 GPU \u72b6\u6001\uff1a<code>nvidia-smi<\/code><\/li>\n\n\n\n<li>\u67e5\u770b\u6307\u5b9a GPU\uff1a<code>nvidia-smi -i 0<\/code><\/li>\n\n\n\n<li><strong>\u6838\u5fc3\u5173\u6ce8 4 \u9879\u6307\u6807<\/strong>\uff1a\u6e29\u5ea6\u3001\u529f\u8017\u3001GPU \u5229\u7528\u7387\u3001\u663e\u5b58\u5360\u7528\u4e0e\u8fdb\u7a0b<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u5b9e\u65f6\u76d1\u63a7<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6bcf\u79d2\u81ea\u52a8\u5237\u65b0\uff1a<code>nvidia-smi -l 1<\/code><\/li>\n\n\n\n<li>\u9ad8\u4eae\u53d8\u5316\u503c\u5237\u65b0\uff1a<code>watch -n 2 -d nvidia-smi<\/code><\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u9ad8\u7ea7\u7ba1\u7406\u547d\u4ee4\u6c47\u603b<\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u529f\u80fd<\/th><th>\u547d\u4ee4<\/th><th><\/th><th><\/th><th><\/th><\/tr><\/thead><tbody><tr><td>\u67e5\u770b GPU \u4e0a\u8fd0\u884c\u7684\u8fdb\u7a0b<\/td><td><code>nvidia-smi pmon<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u91cd\u7f6e\u6307\u5b9a GPU\uff08\u89e3\u51b3\u5361\u6b7b\uff09<\/td><td><code>nvidia-smi -i 0 -r<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u67e5\u770b GPU \u62d3\u6251\u77e9\u9635<\/td><td><code>nvidia-smi topo -m<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u67e5\u770b NVLink \u94fe\u8def\u72b6\u6001<\/td><td><code>nvidia-smi nvlink -s<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u67e5\u770b\u529f\u8017\u8be6\u7ec6\u4fe1\u606f<\/td><td><code>nvidia-smi -q -d POWER<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u8bbe\u7f6e GPU 0 \u529f\u8017\u4e0a\u9650\u4e3a 200W<\/td><td><code>nvidia-smi -i 0 -pl 200<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u67e5\u770b ECC \u62a5\u9519\u7edf\u8ba1<\/td><td><code>nvidia-smi -q -d ECC<\/code><\/td><td><\/td><td><\/td><td><\/td><\/tr><tr><td>\u67e5\u770b\u7cfb\u7edf\u5c42\u9762\u786c\u4ef6\u62a5\u9519<\/td><td>`dmesg -T<\/td><td>grep -iE &#8220;NVRM<\/td><td>Xid<\/td><td>ECC&#8221;`<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">GPU \u6389\u5361\u6392\u67e5\u601d\u8def<\/h4>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5148\u6267\u884c <code>lspci | grep -i nvidia<\/code> \u4e0e <code>nvidia-smi<\/code> \u4ea4\u53c9\u5224\u65ad\n<ul class=\"wp-block-list\">\n<li>lspci \u53ef\u89c1\u3001nvidia-smi \u4e0d\u53ef\u89c1\uff1a\u4f18\u5148\u6000\u7591\u9a71\u52a8\u3001GPU \u521d\u59cb\u5316\u3001XID\u3001Fabric Manager \u7b49<strong>\u8f6f\u4ef6 \/ \u56fa\u4ef6\u95ee\u9898<\/strong><\/li>\n\n\n\n<li>lspci \u4e5f\u4e0d\u53ef\u89c1\uff1a\u504f\u5411 PCIe \u94fe\u8def\u3001GPU \u6a21\u7ec4\u3001SXM \u57fa\u677f\u3001\u4f9b\u7535\u7b49<strong>\u786c\u4ef6\u95ee\u9898<\/strong><\/li>\n<\/ul>\n<\/li>\n\n\n\n<li>\u7ed3\u5408\u4ea4\u53c9\u6d4b\u8bd5\uff08\u6362\u69fd\u3001\u6362\u5361\uff09\u5224\u65ad\u6545\u969c\u662f\u5426\u8f6c\u79fb<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">2. nvitop\uff08\u4ea4\u4e92\u5f0f\u76d1\u63a7\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7279\u70b9<\/strong>\uff1a\u4fe1\u606f\u66f4\u4e30\u5bcc\u3001\u5f69\u8272\u754c\u9762\u76f4\u89c2\u3001\u652f\u6301\u4ea4\u4e92\u5f0f\u64cd\u4f5c\u3001\u5e26\u5386\u53f2\u8d8b\u52bf\u66f2\u7ebf<\/li>\n\n\n\n<li><strong>\u5b89\u88c5<\/strong>\uff1a<code>pip3 install nvitop<\/code><\/li>\n\n\n\n<li><strong>\u542f\u52a8<\/strong>\uff1a<code>nvitop<\/code><\/li>\n\n\n\n<li><strong>\u5e38\u7528\u5feb\u6377\u952e<\/strong>\uff1aq \u9000\u51fa\u3001h \u5e2e\u52a9\u3001p \u6309\u8fdb\u7a0b\u6392\u5e8f\u3001m \u6309\u663e\u5b58\u6392\u5e8f\u3001u \u6309\u5229\u7528\u7387\u6392\u5e8f\u3001k \u6740\u6b7b\u8fdb\u7a0b<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">3. gpustat\uff08\u8f7b\u91cf\u76d1\u63a7\uff09<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u7279\u70b9<\/strong>\uff1a\u7eaf Python \u5b9e\u73b0\u3001\u8f7b\u91cf\u7b80\u6d01\u3001\u652f\u6301 JSON \u8f93\u51fa\u3001\u6613\u96c6\u6210<\/li>\n\n\n\n<li><strong>\u5b89\u88c5<\/strong>\uff1a<code>pip3 install gpustat<\/code><\/li>\n\n\n\n<li><strong>\u5e38\u7528\u547d\u4ee4<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li><code>gpustat<\/code>\uff1a\u5feb\u901f\u67e5\u770b\u72b6\u6001<\/li>\n\n\n\n<li><code>gpustat --watch<\/code>\uff1a\u5b9e\u65f6\u76d1\u63a7<\/li>\n\n\n\n<li><code>gpustat --json<\/code>\uff1aJSON \u683c\u5f0f\u8f93\u51fa\uff0c\u4fbf\u4e8e\u7a0b\u5e8f\u96c6\u6210<\/li>\n<\/ul>\n<\/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\u3001\u5927\u6a21\u578b\u57fa\u7840\u6982\u5ff5\u4e0e\u5206\u7c7b<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u5927\u6a21\u578b\u5206\u7c7b<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u6309\u6a21\u6001\u5212\u5206<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6587\u672c\u5927\u6a21\u578b<\/strong>\uff1a\u5904\u7406\u6587\u672c\u8f93\u5165\u8f93\u51fa\uff0c\u64c5\u957f\u95ee\u7b54\u3001\u5199\u4f5c\u3001\u7ffb\u8bd1\uff1b\u4ee3\u8868\uff1aLlama 3.1\u3001Qwen3<\/li>\n\n\n\n<li><strong>\u591a\u6a21\u6001\u5927\u6a21\u578b<\/strong>\uff1a\u540c\u65f6\u7406\u89e3\u6587\u672c\u3001\u56fe\u7247\uff0c\u90e8\u5206\u652f\u6301\u97f3\u9891\u89c6\u9891\uff1b\u4ee3\u8868\uff1aGPT-4o\u3001Qwen3.5-Omni<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u5927\u6a21\u578b<\/strong>\uff1a\u5f3a\u8c03\u591a\u6b65\u601d\u8003\u4e0e\u590d\u6742\u63a8\u7406\uff0c\u64c5\u957f\u6570\u5b66\u3001\u4ee3\u7801\u3001\u5206\u6790\uff1b\u4ee3\u8868\uff1aDeepSeek-R1<\/li>\n\n\n\n<li><strong>\u751f\u6210\u5f0f\u5a92\u4f53\u6a21\u578b<\/strong>\uff1a\u4e13\u95e8\u751f\u6210\u56fe\u7247 \/ \u89c6\u9891 \/ \u97f3\u9891\uff1b\u4ee3\u8868\uff1aSora\u3001Kling\u3001Midjourney<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u6309\u80fd\u529b\u4e0e\u7528\u9014\u5212\u5206<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u57fa\u5ea7\u6a21\u578b\uff08Base\uff09<\/strong>\uff1a\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\uff0c\u5b66\u4e60\u901a\u7528\u8bed\u8a00\u89c4\u5f8b\uff0c\u662f\u6240\u6709\u80fd\u529b\u7684\u57fa\u7840<\/li>\n\n\n\n<li><strong>\u6307\u4ee4 \/ \u5bf9\u8bdd\u6a21\u578b<\/strong>\uff1a\u5728\u57fa\u5ea7\u4e0a\u505a\u4eba\u7c7b\u504f\u597d\u5bf9\u9f50\uff0c\u66f4\u61c2\u6307\u4ee4\u3001\u64c5\u957f\u5bf9\u8bdd<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u6a21\u578b<\/strong>\uff1a\u5f3a\u5316\u601d\u8003\u80fd\u529b\uff0c\u590d\u6742\u95ee\u9898\u62c6\u89e3\u80fd\u529b\u5f3a<\/li>\n\n\n\n<li><strong>\u4ee3\u7801 \/ Agent \u6a21\u578b<\/strong>\uff1a\u652f\u6301\u4ee3\u7801\u751f\u6210\u3001\u5de5\u5177\u8c03\u7528\u3001\u4efb\u52a1\u6267\u884c<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u6309\u751f\u6001\u4e0e\u5546\u4e1a\u6a21\u5f0f\u5212\u5206<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u95ed\u6e90\u5546\u7528<\/strong>\uff1aAPI \u8c03\u7528\uff0c\u4ea7\u54c1\u6210\u719f\u751f\u6001\u5b8c\u5584\uff1b\u4ee3\u8868\uff1aGPT\u3001Claude\u3001Gemini<\/li>\n\n\n\n<li><strong>\u5f00\u6e90\u5f00\u653e<\/strong>\uff1a\u6743\u91cd\u516c\u5f00\uff0c\u53ef\u672c\u5730\u90e8\u7f72\u3001\u5fae\u8c03\u3001\u4e8c\u6b21\u5f00\u53d1\uff1b\u4ee3\u8868\uff1aLlama\u3001Qwen\u3001DeepSeek<\/li>\n\n\n\n<li><strong>\u5782\u76f4\u884c\u4e1a\u6a21\u578b<\/strong>\uff1a\u5728\u901a\u7528\u6a21\u578b\u57fa\u7840\u4e0a\uff0c\u9762\u5411\u533b\u7597 \/ \u91d1\u878d \/ \u5de5\u4e1a\u7b49\u573a\u666f\u505a\u4e13\u4e1a\u589e\u5f3a<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u6838\u5fc3\u672f\u8bed\uff08\u5fc5\u8bb0\uff09<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u672f\u8bed<\/th><th>\u542b\u4e49<\/th><\/tr><\/thead><tbody><tr><td>Token<\/td><td>\u5927\u6a21\u578b\u5904\u7406\u6587\u672c\u7684\u6700\u5c0f\u5355\u4f4d\uff1b\u4e2d\u6587 100 \u5b57\u2248120-180 Token<\/td><\/tr><tr><td>\u4e0a\u4e0b\u6587\u957f\u5ea6<\/td><td>\u6a21\u578b\u4e00\u6b21\u6700\u591a\u80fd\u8bb0\u4f4f\u7684 Token \u6570\uff0c\u5982 8K\u300132K\u3001128K<\/td><\/tr><tr><td>\u8bad\u7ec3<\/td><td>\u8ba9\u6a21\u578b\u901a\u8fc7\u6d77\u91cf\u6570\u636e\u5b66\u4e60\u77e5\u8bc6\u4e0e\u89c4\u5f8b\u7684\u8fc7\u7a0b<\/td><\/tr><tr><td>\u63a8\u7406<\/td><td>\u4f7f\u7528\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u56de\u7b54\u95ee\u9898\u3001\u751f\u6210\u5185\u5bb9\u7684\u8fc7\u7a0b<\/td><\/tr><tr><td>\u5fae\u8c03\uff08Fine-tuning\uff09<\/td><td>\u5728\u5df2\u6709\u6a21\u578b\u4e0a\u7528\u5c11\u91cf\u6570\u636e\u7ee7\u7eed\u8bad\u7ec3\uff0c\u9002\u914d\u7279\u5b9a\u9886\u57df \/ \u4efb\u52a1<\/td><\/tr><tr><td>TFLOPS<\/td><td>GPU \u7b97\u529b\u5355\u4f4d\uff0c\u6bcf\u79d2\u4e07\u4ebf\u6b21\u6d6e\u70b9\u8fd0\u7b97<\/td><\/tr><tr><td>7B\/70B<\/td><td>\u6a21\u578b\u53c2\u6570\u91cf\u7ea7\uff1b1B = 10 \u4ebf\u53c2\u6570\uff0c\u53c2\u6570\u91cf\u8d8a\u5927\u80fd\u529b\u8d8a\u5f3a\u3001\u663e\u5b58\u8981\u6c42\u8d8a\u9ad8<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u5e38\u89c1\u6a21\u578b\u6587\u4ef6\u683c\u5f0f<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u683c\u5f0f<\/th><th>\u7279\u70b9<\/th><th>\u5178\u578b\u573a\u666f<\/th><\/tr><\/thead><tbody><tr><td><code>.safetensors<\/code><\/td><td>\u5b89\u5168\uff08\u65e0\u53ef\u6267\u884c\u4ee3\u7801\uff09\u3001\u52a0\u8f7d\u901f\u5ea6\u5feb\uff0c\u5f53\u524d\u4e3b\u6d41<\/td><td>\u670d\u52a1\u7aef\u8bad\u7ec3\u4e0e\u63a8\u7406<\/td><\/tr><tr><td><code>.pt\/.pth\/.bin<\/code><\/td><td>\u4f20\u7edf PyTorch \u683c\u5f0f\uff0c\u5b58\u5728\u5b89\u5168\u6f0f\u6d1e\u3001\u52a0\u8f7d\u6162<\/td><td>\u65e7\u7248\u6a21\u578b\u3001\u8bad\u7ec3\u6d41\u7a0b<\/td><\/tr><tr><td><code>.gguf<\/code><\/td><td>\u5355\u6587\u4ef6\u96c6\u6210\uff08\u6743\u91cd + \u5206\u8bcd\u5668 + \u914d\u7f6e\uff09\u3001\u6781\u81f4\u91cf\u5316\uff0c\u786c\u4ef6\u8981\u6c42\u4f4e<\/td><td>\u672c\u5730\u90e8\u7f72\u3001\u7aef\u4fa7\u3001Ollama<\/td><\/tr><tr><td><code>.onnx<\/code><\/td><td>\u8de8\u5e73\u53f0\u3001\u8131\u79bb Python \u73af\u5883<\/td><td>\u7aef\u4fa7\u90e8\u7f72\u3001\u5f02\u6784\u786c\u4ef6\u52a0\u901f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">4. \u672c\u5730\u90e8\u7f72\u7684\u6838\u5fc3\u4ef7\u503c<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6570\u636e\u5b89\u5168\u5408\u89c4<\/strong>\uff1a\u6570\u636e\u6d41\u8f6c\u5b8c\u5168\u5728\u4f01\u4e1a\u5185\u7f51\uff0c\u6ee1\u8db3\u91d1\u878d\u3001\u533b\u7597\u3001\u653f\u52a1\u7b49\u5f3a\u76d1\u7ba1\u8981\u6c42<\/li>\n\n\n\n<li><strong>\u5b9a\u5236\u5316\u4e0e\u63a7\u5236\u6743<\/strong>\uff1a\u652f\u6301\u6df1\u5ea6\u5fae\u8c03\u3001RAG \u77e5\u8bc6\u5e93\u642d\u5efa\uff0c\u6a21\u578b\u7248\u672c\u5b8c\u5168\u53ef\u63a7<\/li>\n\n\n\n<li><strong>\u957f\u671f\u6210\u672c\u4f18\u52bf<\/strong>\uff1a\u524d\u671f\u4e00\u6b21\u6027\u786c\u4ef6\u6295\u5165\uff0c\u9ad8\u9891\u4f7f\u7528\u573a\u666f\u4e0b 1-2 \u5e74\u56de\u672c\uff0c\u8fb9\u9645\u6210\u672c\u8d8b\u8fd1\u4e8e\u96f6<\/li>\n\n\n\n<li><strong>\u4f4e\u5ef6\u8fdf\u4e0e\u79bb\u7ebf\u53ef\u7528<\/strong>\uff1a\u5185\u7f51\u4f20\u8f93\u5ef6\u8fdf\u6781\u4f4e\uff0c\u652f\u6301\u65ad\u7f51\u3001\u4fdd\u5bc6\u3001\u91ce\u5916\u7b49\u79bb\u7ebf\u73af\u5883<\/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\u3001\u5927\u6a21\u578b\u63a8\u7406\u6846\u67b6<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u4f20\u7edf\u90e8\u7f72\uff08Transformers+PyTorch\uff09\u7684\u5c40\u9650\u6027<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u65e0\u52a8\u6001\u6279\u5904\u7406<\/strong>\uff1a\u4e32\u884c\u6216\u9759\u6001 Batch\uff0c\u957f\u8bf7\u6c42\u963b\u585e\u77ed\u8bf7\u6c42\uff0c\u7b97\u529b\u6d6a\u8d39\u4e25\u91cd\u3001\u5ef6\u8fdf\u9ad8<\/li>\n\n\n\n<li><strong>KV Cache \u7ba1\u7406\u7c97\u7cd9<\/strong>\uff1a\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u76f2\u76ee\u9884\u7559\u6700\u5927\u663e\u5b58\uff0c\u788e\u7247\u5316\u4e25\u91cd\uff0c\u663e\u5b58\u5229\u7528\u7387\u6781\u4f4e<\/li>\n\n\n\n<li><strong>\u7f3a\u4e4f\u751f\u4ea7\u7ea7\u4f18\u5316<\/strong>\uff1a\u6d41\u5f0f\u8f93\u51fa\u3001\u591a\u5361\u5e76\u884c\u3001\u8bf7\u6c42\u8c03\u5ea6\u3001\u8d85\u65f6\u91cd\u8bd5\u90fd\u9700\u81ea\u884c\u5f00\u53d1\uff0c\u7ef4\u62a4\u6210\u672c\u9ad8<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u63a8\u7406\u6846\u67b6\u56db\u5927\u6838\u5fc3\u4f18\u5316<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8fde\u7eed\u6279\u5904\u7406\uff08Continuous Batching\uff09<\/strong> \u8bf7\u6c42\u751f\u6210\u5b8c\u6210\u7acb\u5373\u8fd4\u56de\uff0c\u7a7a\u51fa\u7684\u4f4d\u7f6e\u52a8\u6001\u63d2\u5165\u65b0\u8bf7\u6c42\uff0c\u65e0\u9700\u7b49\u5f85\u6574\u6279\u5b8c\u6210\uff0c\u541e\u5410\u91cf\u63d0\u5347\u6570\u500d\u3002<\/li>\n\n\n\n<li><strong>PagedAttention\uff08\u5206\u9875 KV \u7f13\u5b58\uff09<\/strong> \u501f\u9274\u64cd\u4f5c\u7cfb\u7edf\u865a\u62df\u5185\u5b58\u601d\u60f3\uff0c\u5c06\u663e\u5b58\u5207\u5206\u4e3a \u201c\u9875\u201d\uff0c\u6309\u9700\u52a8\u6001\u5206\u914d\u3001\u7528\u5b8c\u5373\u56de\u6536\uff0c\u6d88\u9664\u7ea6 96% \u663e\u5b58\u6d6a\u8d39\uff0c\u5927\u5e45\u63d0\u5347\u5e76\u53d1\u6570\u3002<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u91cf\u5316<\/strong> \u5c06 FP16 \u7cbe\u5ea6\u538b\u7f29\u4e3a INT4\/INT8\uff0c\u663e\u5b58\u9700\u6c42\u964d\u4f4e\u6570\u500d\uff0c\u6574\u6570\u8fd0\u7b97\u901f\u5ea6\u66f4\u5feb\uff0c\u5728\u53ef\u63a5\u53d7\u7cbe\u5ea6\u635f\u5931\u4e0b\u5927\u5e45\u964d\u4f4e\u786c\u4ef6\u95e8\u69db\u3002<\/li>\n\n\n\n<li><strong>\u5de5\u7a0b\u5316\u63a5\u53e3<\/strong> \u5f00\u7bb1\u63d0\u4f9b OpenAI \u517c\u5bb9 API\u3001\u81ea\u52a8\u591a\u5361\u5e76\u884c\u3001\u6d41\u5f0f\u8f93\u51fa\u3001\u8bf7\u6c42\u961f\u5217\u8c03\u5ea6\uff0c\u65e0\u9700\u91cd\u590d\u9020\u8f6e\u5b50\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">3. \u4e3b\u6d41\u63a8\u7406\u6846\u67b6<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">Llama.cpp<\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5b9a\u4f4d\uff1a\u8f7b\u91cf\u7ea7 C\/C++ \u63a8\u7406\u5f15\u64ce<\/li>\n\n\n\n<li>\u7279\u70b9\uff1a\u65e0\u5916\u90e8\u4f9d\u8d56\u3001\u652f\u6301 CPU\/GPU \u5168\u5e73\u53f0\u3001\u591a\u79cd\u91cf\u5316\u7cbe\u5ea6<\/li>\n\n\n\n<li>\u9002\u7528\uff1a\u8d44\u6e90\u53d7\u9650\u73af\u5883\u3001\u7aef\u4fa7\u8bbe\u5907\u3001\u5e95\u5c42\u5b9a\u5236\u5f00\u53d1<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">Ollama <strong>\uff01\uff01\uff01<\/strong><\/h4>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5b9a\u4f4d\uff1a\u96f6\u95e8\u69db\u672c\u5730\u63a8\u7406\u5e73\u53f0\uff0c\u5e95\u5c42\u57fa\u4e8e llama.cpp<\/li>\n\n\n\n<li>\u67b6\u6784\uff1a\u7ecf\u5178 Client-Server \u67b6\u6784\uff0c\u670d\u52a1\u7aef = HTTP \u63a5\u53e3\u5c42 + llama.cpp \u63a8\u7406\u5f15\u64ce<\/li>\n\n\n\n<li>\u7279\u70b9\uff1a\u4e00\u952e\u90e8\u7f72\u3001\u81ea\u52a8 GPU \u52a0\u901f\u3001\u6a21\u578b\u751f\u6001\u4e30\u5bcc\u3001\u652f\u6301\u547d\u4ee4\u884c\u4e0e API \u53cc\u6a21\u5f0f<\/li>\n\n\n\n<li>\u9002\u7528\uff1a\u4e2a\u4eba\u672c\u5730\u4f7f\u7528\u3001\u8f7b\u91cf\u79c1\u6709\u5316\u90e8\u7f72\u3001\u5feb\u901f\u9a8c\u8bc1\u573a\u666f<\/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\">\u4e94\u3001Ollama \u672c\u5730\u90e8\u7f72\u5b9e\u6218<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">1. \u5b89\u88c5\u65b9\u5f0f<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5728\u7ebf\u4e00\u952e\u5b89\u88c5<\/strong>\uff1a<code>curl -fsSL https:\/\/ollama.com\/install.sh | sh<\/code><\/li>\n\n\n\n<li><strong>\u79bb\u7ebf\u5b89\u88c5\u6b65\u9aa4<\/strong>\uff1a\n<ol class=\"wp-block-list\">\n<li>\u5b89\u88c5\u89e3\u538b\u4f9d\u8d56\uff1a<code>apt install -y zstd<\/code><\/li>\n\n\n\n<li>\u89e3\u538b\u5b89\u88c5\u5305\uff1a<code>tar --use-compress-program=unzstd -xvf ollama-linux-amd64.tar.zst -C \/root\/ollama<\/code><\/li>\n\n\n\n<li>\u914d\u7f6e\u8f6f\u94fe\u63a5\uff1a<code>ln -sf \/root\/ollama\/bin\/ollama \/usr\/local\/bin\/ollama<\/code><\/li>\n\n\n\n<li>\u7f16\u5199 systemd \u670d\u52a1\uff0c\u8bbe\u7f6e\u5f00\u673a\u81ea\u542f\uff0c\u76d1\u542c 0.0.0.0:11434<\/li>\n\n\n\n<li>\u542f\u52a8\u670d\u52a1\u5e76\u9a8c\u8bc1\uff1a<code>ollama -v<\/code><\/li>\n<\/ol>\n<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">2. \u5e38\u7528\u547d\u4ee4\u6c47\u603b<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th>\u547d\u4ee4<\/th><th>\u529f\u80fd<\/th><\/tr><\/thead><tbody><tr><td><code>ollama list<\/code><\/td><td>\u5217\u51fa\u672c\u5730\u6240\u6709\u6a21\u578b<\/td><\/tr><tr><td><code>ollama pull \u6a21\u578b\u540d<\/code><\/td><td>\u4ece\u8fdc\u7a0b\u4ed3\u5e93\u62c9\u53d6 \/ \u589e\u91cf\u66f4\u65b0\u6a21\u578b<\/td><\/tr><tr><td><code>ollama run \u6a21\u578b\u540d<\/code><\/td><td>\u8fd0\u884c\u6a21\u578b\u5e76\u8fdb\u5165\u7ec8\u7aef\u5bf9\u8bdd<\/td><\/tr><tr><td><code>ollama ps<\/code><\/td><td>\u67e5\u770b\u5f53\u524d\u6b63\u5728\u8fd0\u884c\u7684\u6a21\u578b<\/td><\/tr><tr><td><code>ollama stop \u6a21\u578b\u540d<\/code><\/td><td>\u505c\u6b62\u8fd0\u884c\u4e2d\u7684\u6a21\u578b\uff0c\u91ca\u653e\u663e\u5b58<\/td><\/tr><tr><td><code>ollama rm \u6a21\u578b\u540d<\/code><\/td><td>\u5220\u9664\u672c\u5730\u6a21\u578b\u6587\u4ef6<\/td><\/tr><tr><td><code>ollama cp \u539f\u540d\u79f0 \u65b0\u540d\u79f0<\/code><\/td><td>\u590d\u5236\u6a21\u578b\uff08\u7528\u4e8e\u81ea\u5b9a\u4e49\u4fee\u6539\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">3. API \u8c03\u7528\uff08\u9ed8\u8ba4\u7aef\u53e3\uff1a11434\uff09<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\u975e\u6d41\u5f0f\u8c03\u7528\uff08\u4e00\u6b21\u6027\u8fd4\u56de\u7ed3\u679c\uff09<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>curl http:\/\/localhost:11434\/api\/generate -d '{\n  \"model\": \"qwen3.5:4b\",\n  \"prompt\": \"\u4ece\u8fd0\u7ef4\u89d2\u5ea6\u4ecb\u7ecdDocker\u7684\u4f5c\u7528\",\n  \"stream\": false\n}'\n<\/code><\/pre>\n\n\n\n<h4 class=\"wp-block-heading\">\u6d41\u5f0f\u8c03\u7528\uff08\u9010\u5b57\u8fd4\u56de\uff09<\/h4>\n\n\n\n<pre class=\"wp-block-code\"><code>curl http:\/\/localhost:11434\/api\/generate -d '{\n  \"model\": \"qwen:7b\",\n  \"prompt\": \"\u4e3a\u4ec0\u4e48\u8349\u662f\u7eff\u7684?\"\n}'<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001GPU \u9a71\u52a8\u4e0e CUDA \u5b89\u88c5 1. GPU \u9a71\u52a8\u6838\u5fc3\u8ba4\u77e5 2. \u9a71\u52a8\u5b89\u88c5\u65b9\u5f0f\u5bf9\u6bd4 \u5b89\u88c5\u65b9\u5f0f \u662f\u5426\u8054\u7f51 \u5305 [&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-125","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/125","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=125"}],"version-history":[{"count":3,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/125\/revisions"}],"predecessor-version":[{"id":187,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/125\/revisions\/187"}],"wp:attachment":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=125"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=125"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=125"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}