{"id":166,"date":"2026-09-02T23:49:36","date_gmt":"2026-09-02T15:49:36","guid":{"rendered":"https:\/\/www.xiaoxindong.cloud\/?p=166"},"modified":"2026-09-06T19:52:41","modified_gmt":"2026-09-06T11:52:41","slug":"gpu-%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%e5%a4%8d%e4%b9%a0%e8%b5%84%e6%96%99%ef%bc%88k8s-%e9%9b%86%e7%be%a4-gpu-%e8%b0%83%e5%ba%a6%e4%b8%8e%e6%9c%8d%e5%8a%a1%e9%83%a8","status":"publish","type":"post","link":"https:\/\/www.xiaoxindong.cloud\/?p=166","title":{"rendered":"GPU \u5927\u6a21\u578b\u672c\u5730\u90e8\u7f72\u590d\u4e60\u8d44\u6599\uff08K8S \u96c6\u7fa4 GPU \u8c03\u5ea6\u4e0e\u670d\u52a1\u90e8\u7f72\uff09"},"content":{"rendered":"\n

\u672c\u7bc7\u805a\u7126 Kubernetes \u96c6\u7fa4\u4e0b\u7684 GPU \u7b97\u529b\u8c03\u5ea6\u4e0e\u5927\u6a21\u578b\u670d\u52a1\u7f16\u6392\uff0c\u638c\u63e1K8s \u96c6\u7fa4\u642d\u5efa<\/strong>\u3001NVIDIA Device Plugin \u539f\u7406\u4e0e\u90e8\u7f72<\/strong>\u3001\u4e09\u5927 AI \u6846\u67b6\u7684\u5bb9\u5668\u5316\u7f16\u6392<\/strong><\/p>\n\n\n\n

\n\n\n\n

\u4e00\u3001\u4e3a\u4ec0\u4e48\u7528 K8s \u7ba1\u7406 GPU \u96c6\u7fa4<\/h2>\n\n\n\n

1. \u65e0 K8s \u7ba1\u7406\u7684\u6838\u5fc3\u75db\u70b9<\/h3>\n\n\n\n
\u75db\u70b9<\/th>\u5177\u4f53\u8868\u73b0<\/th><\/tr><\/thead>
GPU \u8d44\u6e90\u6d6a\u8d39<\/strong><\/td>\u4eba\u5de5\u5206\u914d\u7c92\u5ea6\u7c97\uff0c\u663e\u5b58 \/ \u7b97\u529b\u788e\u7247\u5316\uff0c\u6602\u8d35 GPU \u7a7a\u8f7d\u7387\u9ad8\uff0c\u6210\u672c\u635f\u8017\u5927<\/td><\/tr>
\u4efb\u52a1\u8c03\u5ea6\u6df7\u4e71<\/strong><\/td>\u591a\u7528\u6237\u591a\u4efb\u52a1\u62a2\u5361\uff0c\u4eba\u5de5\u5206\u914d\u51b2\u7a81\u591a\u3001\u6548\u7387\u4f4e\uff0c\u65e0\u6cd5\u6309\u9700\u6c42\u5f39\u6027\u5206\u914d<\/td><\/tr>
\u670d\u52a1\u7a33\u5b9a\u6027\u5dee<\/strong><\/td>\u8fdb\u7a0b\u5d29\u6e83\u65e0\u81ea\u52a8\u6062\u590d\uff0c\u4e1a\u52a1\u4e2d\u65ad\uff1b\u6d41\u91cf\u6ce2\u52a8\u65e0\u6cd5\u5feb\u901f\u6269\u5bb9\uff0c\u9ad8\u5cf0\u671f\u670d\u52a1\u88ab\u6253\u6302<\/td><\/tr>
\u6269\u5bb9\u8fd0\u7ef4\u56f0\u96be<\/strong><\/td>\u591a\u673a\u7ba1\u7406\u6df7\u4e71\uff0c\u73af\u5883\u4e0d\u4e00\u81f4\uff0c\u90e8\u7f72\u3001\u8fc1\u79fb\u3001\u5347\u7ea7\u6210\u672c\u9ad8\u3001\u5468\u671f\u957f<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

2. K8s \u89e3\u51b3\u7684\u6838\u5fc3\u4ef7\u503c<\/h3>\n\n\n\n
\u80fd\u529b<\/th>\u8bf4\u660e<\/th><\/tr><\/thead>
\u81ea\u52a8 GPU \u8c03\u5ea6<\/strong><\/td>\u96c6\u7fa4\u7edf\u4e00\u7ba1\u7406 GPU \u8d44\u6e90\uff0c\u6309\u9700\u81ea\u52a8\u5206\u914d\uff0c\u8d44\u6e90\u5229\u7528\u7387\u6700\u5927\u5316<\/td><\/tr>
\u81ea\u52a8\u5f39\u6027\u6269\u7f29<\/strong><\/td>\u57fa\u4e8e\u6d41\u91cf\u81ea\u52a8\u589e\u51cf Pod \u526f\u672c\uff0c\u4f4e\u8c37\u7701\u8d44\u6e90\u3001\u9ad8\u5cf0\u625b\u538b\u529b<\/td><\/tr>
\u670d\u52a1\u81ea\u6108<\/strong><\/td>Pod \u5f02\u5e38\u5d29\u6e83\u81ea\u52a8\u91cd\u542f\u91cd\u5efa\uff0c\u4fdd\u8bc1\u670d\u52a1\u6301\u7eed\u53ef\u7528<\/td><\/tr>
\u591a\u673a\u7edf\u4e00\u7ba1\u7406<\/strong><\/td>\u6807\u51c6\u5316\u90e8\u7f72\u3001\u914d\u7f6e\u3001\u8fd0\u7ef4\uff0c\u652f\u6301\u5927\u89c4\u6a21\u96c6\u7fa4\u9ad8\u6548\u6269\u5c55\uff0c\u73af\u5883\u4e00\u81f4\u6027\u5f3a<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
\n\n\n\n

\u4e8c\u3001K8s \u96c6\u7fa4\u90e8\u7f72\uff08\u2b50\uff09<\/h2>\n\n\n\n

1. \u96c6\u7fa4\u89c4\u5212\u793a\u4f8b<\/h3>\n\n\n\n
\u4e3b\u673a\u540d<\/th>IP<\/th>\u89d2\u8272<\/th><\/tr><\/thead>
k8s-master<\/td>192.168.100.146<\/td>\u63a7\u5236\u9762\uff08CPU\uff09<\/td><\/tr>
k8s-node1<\/td>192.168.100.147<\/td>\u5de5\u4f5c\u8282\u70b9\uff08GPU\uff09<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

2. \u6240\u6709\u8282\u70b9\u57fa\u7840\u73af\u5883\u914d\u7f6e<\/h3>\n\n\n\n

2.1 \u914d\u7f6e APT \u963f\u91cc\u4e91\u6e90<\/h4>\n\n\n\n
# \u5907\u4efd\u539f\u6e90\ncp \/etc\/apt\/sources.list \/etc\/apt\/sources.list.backup\n# \u66ff\u6362\u4e3a\u963f\u91cc\u4e91\u955c\u50cf\nsed -i 's\/archive.ubuntu.com\/mirrors.aliyun.com\/g' \/etc\/apt\/sources.list\nsed -i 's\/security.ubuntu.com\/mirrors.aliyun.com\/g' \/etc\/apt\/sources.list\napt update\n<\/code><\/pre>\n\n\n\n
2.2 \u8bbe\u7f6e\u4e3b\u673a\u540d\u4e0e\u4e3b\u673a\u89e3\u6790<\/h4>\n\n\n\n
# Master\u6267\u884c\nhostnamectl set-hostname k8s-master && bash\n# Worker\u6267\u884c\nhostnamectl set-hostname k8s-node1 && bash\n\n# \u6240\u6709\u8282\u70b9\u6dfb\u52a0\u4e3b\u673a\u89e3\u6790\ncat <<EOF | sudo tee -a \/etc\/hosts\n192.168.32.14 k8s-master\n192.168.32.13 k8s-node1\nEOF\n<\/code><\/pre>\n\n\n\n
2.3 \u65f6\u949f\u540c\u6b65<\/h4>\n\n\n\n
apt install -y chrony\nsystemctl enable chrony && systemctl start chrony\ntimedatectl set-timezone Asia\/Shanghai\n<\/code><\/pre>\n\n\n\n
2.4 \u5173\u95ed Swap\uff08\u5fc5\u987b\uff0c\u5426\u5219 kubelet \u65e0\u6cd5\u542f\u52a8\uff09<\/h4>\n\n\n\n
# \u4e34\u65f6\u5173\u95ed\nswapoff -a\n# \u6c38\u4e45\u5173\u95ed\nsed -i '\/swap\/s\/^\/#\/' \/etc\/fstab\n# \u9a8c\u8bc1\nfree -h\n<\/code><\/pre>\n\n\n\n
2.5 \u5173\u95ed\u9632\u706b\u5899<\/h4>\n\n\n\n
ufw disable\nsystemctl stop ufw && systemctl disable ufw\n<\/code><\/pre>\n\n\n\n
2.6 \u5f00\u542f IPv4 \u8f6c\u53d1\u4e0e\u6865\u63a5\u5185\u6838\u53c2\u6570<\/h4>\n\n\n\n
cat <<EOF | sudo tee \/etc\/sysctl.d\/k8s.conf\nnet.bridge.bridge-nf-call-iptables = 1\nnet.bridge.bridge-nf-call-ip6tables = 1\nnet.ipv4.ip_forward = 1\nEOF\nsysctl --system\nmodprobe br_netfilter\n<\/code><\/pre>\n\n\n\n
3. \u5b89\u88c5 containerd \u8fd0\u884c\u65f6<\/h3>\n\n\n\n
# \u4e0b\u8f7d\u5b89\u88c5\ncurl -O https:\/\/mirrors.aliyun.com\/docker-ce\/linux\/ubuntu\/dists\/jammy\/pool\/stable\/amd64\/containerd.io_1.6.26-1_amd64.deb\ndpkg -i containerd.io_1.6.26-1_amd64.deb\n\n# \u751f\u6210\u5e76\u4fee\u6539\u914d\u7f6e\nmkdir -p \/etc\/containerd\ncontainerd config default | sudo tee \/etc\/containerd\/config.toml\n# \u542f\u7528SystemdCgroup\uff08\u4e0ekubelet\u4e00\u81f4\uff09\nsed -i 's\/SystemdCgroup = false\/SystemdCgroup = true\/g' \/etc\/containerd\/config.toml\n# \u66ff\u6362pause\u955c\u50cf\u4e3a\u56fd\u5185\u6e90\nsed -i 's#sandbox_image = \"registry.k8s.io\/pause:.*\"#sandbox_image = \"registry.aliyuncs.com\/google_containers\/pause:3.6\"#' \/etc\/containerd\/config.toml\n\n# \u542f\u52a8\u670d\u52a1\nsystemctl enable containerd && systemctl restart containerd\n<\/code><\/pre>\n\n\n\n
4. \u914d\u7f6e crictl \u5ba2\u6237\u7aef<\/h3>\n\n\n\n
# \u5b89\u88c5\nexport VERSION=\"v1.29.0\"\nwget https:\/\/github.com\/kubernetes-sigs\/cri-tools\/releases\/download\/${VERSION}\/crictl-${VERSION}-linux-amd64.tar.gz\ntar -zxvf crictl-${VERSION}-linux-amd64.tar.gz -C \/usr\/local\/bin\n\n# \u914d\u7f6e\ntee \/etc\/crictl.yaml <<EOF\nruntime-endpoint: unix:\/\/\/run\/containerd\/containerd.sock\nimage-endpoint: unix:\/\/\/run\/containerd\/containerd.sock\ntimeout: 10\nEOF\n<\/code><\/pre>\n\n\n\n
5. \u5b89\u88c5 kubeadm\/kubelet\/kubectl<\/h3>\n\n\n\n
# \u6dfb\u52a0GPG\u5bc6\u94a5\u4e0e\u963f\u91cc\u4e91K8s\u6e90\ncurl -fsSL https:\/\/mirrors.aliyun.com\/kubernetes\/apt\/doc\/apt-key.gpg | sudo apt-key add -\ncat <<EOF | sudo tee \/etc\/apt\/sources.list.d\/kubernetes.list\ndeb https:\/\/mirrors.aliyun.com\/kubernetes\/apt\/ kubernetes-xenial main\nEOF\n\n# \u5b89\u88c5\u6307\u5b9a\u7248\u672c\u5e76\u9501\u5b9a\napt update\napt install -y kubelet=1.28.2-00 kubeadm=1.28.2-00 kubectl=1.28.2-00\napt-mark hold kubelet kubeadm kubectl\nsystemctl enable kubelet\n<\/code><\/pre>\n\n\n\n
6. Master \u8282\u70b9\u521d\u59cb\u5316<\/h3>\n\n\n\n
# \u9884\u62c9\u53d6\u955c\u50cf\nkubeadm config images pull --image-repository=registry.aliyuncs.com\/google_containers --kubernetes-version=\"v1.28.2\"\n\n# \u521d\u59cb\u5316\u96c6\u7fa4\nkubeadm init \\\n  --kubernetes-version=v1.28.2 \\\n  --pod-network-cidr=10.244.0.0\/16 \\\n  --service-cidr=10.96.0.0\/12 \\\n  --image-repository=registry.aliyuncs.com\/google_containers\n<\/code><\/pre>\n\n\n\n
\u521d\u59cb\u5316\u6210\u529f\u540e\u6267\u884c\u914d\u7f6e\uff1a<\/p>\n\n\n\n
mkdir -p $HOME\/.kube\nsudo cp -i \/etc\/kubernetes\/admin.conf $HOME\/.kube\/config\nsudo chown $(id -u):$(id -g) $HOME\/.kube\/config\n<\/code><\/pre>\n\n\n\n
7. \u90e8\u7f72 Flannel \u7f51\u7edc\u63d2\u4ef6<\/h3>\n\n\n\n
kubectl apply -f kube-flannel.yml\n# \u9a8c\u8bc1\nkubectl get pod -n kube-flannel\n<\/code><\/pre>\n\n\n\n
8. Worker \u8282\u70b9\u52a0\u5165\u96c6\u7fa4<\/h3>\n\n\n\n
\u5728 Worker \u8282\u70b9\u6267\u884c Master \u521d\u59cb\u5316\u8f93\u51fa\u7684kubeadm join<\/code>\u547d\u4ee4\uff0c\u683c\u5f0f\u5982\u4e0b\uff1a<\/p>\n\n\n\n
kubeadm join 192.168.215.129:6443 \\\n  --token 97pgmv.vxjakgynflgdh4zw \\\n  --discovery-token-ca-cert-hash sha256:52163ce2f39c0ccc44558e7339c52bfd006c78944088efb2f94dcc57d424b816\n<\/code><\/pre>\n\n\n\n
9. \u96c6\u7fa4\u72b6\u6001\u9a8c\u8bc1<\/h3>\n\n\n\n
kubectl get nodes -o wide\nkubectl get pods -n kube-system\n<\/code><\/pre>\n\n\n\n
\u6240\u6709\u8282\u70b9\u72b6\u6001\u4e3aReady<\/code>\u3001\u7cfb\u7edf Pod \u5168\u90e8 Running \u5219\u96c6\u7fa4\u6b63\u5e38\u3002<\/p>\n\n\n\n
\n\n\n\n
\u4e09\u3001K8s GPU \u7b97\u529b\u8c03\u5ea6\uff08\u2b50\uff09<\/h2>\n\n\n\n
1. NVIDIA Device Plugin \u6982\u8ff0<\/h3>\n\n\n\n
    \n
  • \u5b9a\u4f4d<\/strong>\uff1aKubernetes \u96c6\u7fa4\u7ea7 GPU \u8bbe\u5907\u7ba1\u7406\u63d2\u4ef6\uff0c\u4ee5 DaemonSet \u5f62\u5f0f\u8fd0\u884c\u5728\u6bcf\u4e2a GPU \u8282\u70b9\u3002<\/li>\n\n\n\n
  • \u6838\u5fc3\u529f\u80fd<\/strong>\uff1a\u5411 kubelet \u6ce8\u518c GPU \u8d44\u6e90\u3001\u76d1\u63a7 GPU \u5065\u5eb7\u72b6\u6001\u3001\u5b9e\u73b0 GPU \u8d44\u6e90\u7684\u5206\u914d\u4e0e\u6e05\u7406\u3002<\/li>\n\n\n\n
  • \u5de5\u4f5c\u539f\u7406<\/strong>\uff1a\n
      \n
    1. \u63d2\u4ef6\u901a\u8fc7 gRPC \u63a5\u53e3\u4e0e kubelet \u901a\u4fe1\uff0c\u4e0a\u62a5\u8282\u70b9 GPU \u6570\u91cf\u3001\u5065\u5eb7\u72b6\u6001<\/li>\n\n\n\n
    2. K8s \u8c03\u5ea6\u5668\u67e5\u8be2\u8bbe\u5907\u8d44\u6e90\uff0c\u5c06 Pod \u8c03\u5ea6\u5230\u6709\u53ef\u7528 GPU \u7684\u8282\u70b9<\/li>\n\n\n\n
    3. Pod \u521b\u5efa\u65f6\uff0ckubelet \u8c03\u7528\u63d2\u4ef6\u7684 Allocate \u63a5\u53e3\uff0c\u5c06 GPU \u8bbe\u5907\u3001\u9a71\u52a8\u5e93\u52a8\u6001\u6ce8\u5165\u5bb9\u5668<\/li>\n\n\n\n
    4. \u5bb9\u5668\u5185\u5e94\u7528\u5373\u53ef\u900f\u660e\u4f7f\u7528 GPU\uff0c\u5168\u7a0b\u6807\u51c6\u5316\u8c03\u5ea6\u3001\u65e0\u9700\u624b\u52a8\u5e72\u9884<\/li>\n<\/ol>\n<\/li>\n<\/ul>\n\n\n\n
      2. \u90e8\u7f72\u524d\u7f6e\uff1aNVIDIA Container Toolkit<\/h3>\n\n\n\n
      GPU \u8282\u70b9\u5fc5\u987b\u5148\u5b89\u88c5 Container Toolkit\uff0c\u8ba9 containerd \u652f\u6301 GPU \u8fd0\u884c\u65f6\u3002<\/p>\n\n\n\n
      # 1. \u6dfb\u52a0GPG\u5bc6\u94a5\u4e0e\u8f6f\u4ef6\u6e90\ncurl -fsSL https:\/\/nvidia.github.io\/libnvidia-container\/gpgkey | sudo gpg --dearmor -o \/usr\/share\/keyrings\/nvidia-container-toolkit-keyring.gpg\ncurl -s -L https:\/\/nvidia.github.io\/libnvidia-container\/stable\/deb\/nvidia-container-toolkit.list | \\\nsed 's#deb https:\/\/#deb [signed-by=\/usr\/share\/keyrings\/nvidia-container-toolkit-keyring.gpg] https:\/\/#' | \\\nsudo tee \/etc\/apt\/sources.list.d\/nvidia-container-toolkit.list\n\n# 2. \u5b89\u88c5\u6307\u5b9a\u7248\u672c\napt update\nexport NVIDIA_CONTAINER_TOOLKIT_VERSION=1.17.8-1\napt install -y nvidia-container-toolkit=${NVIDIA_CONTAINER_TOOLKIT_VERSION}\n\n# 3. \u914d\u7f6econtainerd\u9ed8\u8ba4\u8fd0\u884c\u65f6\u4e3anvidia\nnvidia-ctk runtime configure --runtime=containerd\nsed -i 's\/default_runtime_name *= *\"runc\"\/default_runtime_name = \"nvidia\"\/g' \/etc\/containerd\/config.toml\nsystemctl restart containerd\n<\/code><\/pre>\n\n\n\n
      3. \u90e8\u7f72 NVIDIA Device Plugin<\/h3>\n\n\n\n
      kubectl apply -f nvidia-device-plugin.yml\n<\/code><\/pre>\n\n\n\n
        \n
      • \u90e8\u7f72\u540e\u4f1a\u5728\u6bcf\u4e2a GPU \u8282\u70b9\u542f\u52a8\u4e00\u4e2a\u63d2\u4ef6 Pod\uff0c\u5411\u96c6\u7fa4\u6ce8\u518cnvidia.com\/gpu<\/code>\u8d44\u6e90\u7c7b\u578b\u3002<\/li>\n<\/ul>\n\n\n\n
        4. \u90e8\u7f72\u9a8c\u8bc1<\/h3>\n\n\n\n
        # \u67e5\u770b\u63d2\u4ef6\u65e5\u5fd7\uff0c\u65e0\u62a5\u9519\u5219\u6b63\u5e38\nkubectl logs -f nvidia-device-plugin-daemonset-xxx -n kube-system\n\n# \u67e5\u770b\u8282\u70b9\u53ef\u5206\u914dGPU\u8d44\u6e90\nkubectl describe node k8s-node1 | grep nvidia.com\/gpu\n<\/code><\/pre>\n\n\n\n
        5. \u6613\u6df7\u6982\u5ff5\u5bf9\u6bd4<\/h3>\n\n\n\n
        \u5bf9\u6bd4\u9879<\/th>NVIDIA Container Toolkit<\/th>K8s NVIDIA Device Plugin<\/th><\/tr><\/thead>
        \u4f5c\u7528\u5c42\u7ea7<\/strong><\/td>\u5355\u673a\u5bb9\u5668\u8fd0\u884c\u65f6\u7ea7\u522b<\/td>K8s \u96c6\u7fa4\u8c03\u5ea6\u7ea7\u522b<\/td><\/tr>
        \u6838\u5fc3\u529f\u80fd<\/strong><\/td>\u8ba9\u5355\u4e2a\u5bb9\u5668\u80fd\u8bbf\u95ee GPU<\/td>\u8ba9 K8s \u96c6\u7fa4\u80fd\u7ba1\u7406\u3001\u8c03\u5ea6 GPU<\/td><\/tr>
        \u90e8\u7f72\u4f4d\u7f6e<\/strong><\/td>\u5bbf\u4e3b\u673a\u64cd\u4f5c\u7cfb\u7edf<\/td>K8s \u96c6\u7fa4 DaemonSet<\/td><\/tr>
        \u4f9d\u8d56\u5173\u7cfb<\/strong><\/td>\u4f9d\u8d56 containerd\/Docker<\/td>\u4f9d\u8d56 Container Toolkit + kubelet<\/td><\/tr>
        \u9002\u7528\u573a\u666f<\/strong><\/td>\u5355\u673a Docker\/containerd \u8dd1 GPU \u5bb9\u5668<\/td>K8s \u96c6\u7fa4\u591a\u8282\u70b9 GPU \u7f16\u6392\u8c03\u5ea6<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n
        \n\n\n\n
        \u56db\u3001AI \u6846\u67b6 K8s \u90e8\u7f72\uff08\u2b50\uff09<\/h2>\n\n\n\n
        1. Ollama \u90e8\u7f72<\/h3>\n\n\n\n
        \u6838\u5fc3\u914d\u7f6e\u8981\u70b9<\/h4>\n\n\n\n
          \n
        • \u8d44\u6e90\u58f0\u660e<\/strong>\uff1alimits.nvidia.com\/gpu: 1<\/code> \u58f0\u660e\u7533\u8bf7 1 \u5f20 GPU<\/li>\n\n\n\n
        • \u6570\u636e\u6301\u4e45\u5316<\/strong>\uff1ahostPath \u6302\u8f7d\u6a21\u578b\u76ee\u5f55\uff0c\u907f\u514d Pod \u91cd\u5efa\u4e22\u5931\u6a21\u578b<\/li>\n\n\n\n
        • \u670d\u52a1\u66b4\u9732<\/strong>\uff1aNodePort \u7c7b\u578b\u66b4\u9732 11434 \u7aef\u53e3<\/li>\n<\/ul>\n\n\n\n
          Deployment \u6838\u5fc3\u7247\u6bb5<\/h4>\n\n\n\n
          spec:\n  containers:\n  - name: ollama\n    image: 192.168.100.149\/gpuimages\/ollama:latest\n    ports:\n    - containerPort: 11434\n    env:\n    - name: OLLAMA_HOST\n      value: \"0.0.0.0\"\n    resources:\n      limits:\n        nvidia.com\/gpu: 1\n    volumeMounts:\n    - name: ollama-data\n      mountPath: \/root\/.ollama\n  volumes:\n  - name: ollama-data\n    hostPath:\n      path: \/data\/ollama\n<\/code><\/pre>\n\n\n\n
          \u5e38\u7528\u64cd\u4f5c<\/h4>\n\n\n\n
          # \u8fdb\u5165Pod\u62c9\u53d6\u6a21\u578b\nkubectl exec -it ollama-xxx -- bash\nollama pull qwen3.5:4b\n\n# \u5916\u90e8\u8c03\u7528API\ncurl http:\/\/\u8282\u70b9IP:30114\/api\/generate -d '{\n  \"model\": \"qwen3.5:4b\",\n  \"prompt\": \"\u5199\u4e00\u4e2aK8s\u6392\u969c\u6d41\u7a0b\",\n  \"stream\": false\n}'\n<\/code><\/pre>\n\n\n\n
          2. vLLM \u90e8\u7f72<\/h3>\n\n\n\n
          \u6838\u5fc3\u914d\u7f6e\u8981\u70b9<\/h4>\n\n\n\n
            \n
          • \u542f\u52a8\u547d\u4ee4<\/strong>\uff1a\u901a\u8fc7 command\/args \u6307\u5b9a\u6a21\u578b\u3001\u5e76\u884c\u6570\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u7b49\u53c2\u6570<\/li>\n\n\n\n
          • \u73af\u5883\u53d8\u91cf<\/strong>\uff1a\u5f00\u542f ModelScope \u52a0\u901f\u56fd\u5185\u6a21\u578b\u4e0b\u8f7d<\/li>\n\n\n\n
          • GPU \u8d44\u6e90<\/strong>\uff1a\u58f0\u660e\u5bf9\u5e94\u6570\u91cf\u7684 GPU\uff0c\u652f\u6301\u5f20\u91cf\u5e76\u884c<\/li>\n\n\n\n
          • \u6027\u80fd\u4f18\u5316<\/strong>\uff1a\u5efa\u8bae\u4f7f\u7528hostNetwork<\/code>\u51cf\u5c11\u7f51\u7edc\u8f6c\u53d1\uff0c\u914d\u7f6e\u5927\u5171\u4eab\u5185\u5b58<\/li>\n<\/ul>\n\n\n\n
            Deployment \u6838\u5fc3\u7247\u6bb5<\/h4>\n\n\n\n
            spec:\n  containers:\n  - name: vllm\n    image: 192.168.100.149\/gpuimages\/vllm:latest\n    command: [\"python3\", \"-m\", \"vllm.entrypoints.openai.api_server\"]\n    args:\n    - \"--model\"\n    - \"Qwen\/Qwen2-1.5B-Instruct\"\n    - \"--tensor-parallel-size\"\n    - \"1\"\n    - \"--max-model-len\"\n    - \"4096\"\n    - \"--port\"\n    - \"8000\"\n    - \"--host\"\n    - \"0.0.0.0\"\n    env:\n    - name: VLLM_USE_MODELSCOPE\n      value: \"true\"\n    resources:\n      limits:\n        nvidia.com\/gpu: 1\n    ports:\n    - containerPort: 8000\n<\/code><\/pre>\n\n\n\n
            API \u9a8c\u8bc1<\/h4>\n\n\n\n
            curl http:\/\/\u8282\u70b9IP:30080\/v1\/chat\/completions \\\n  -H \"Content-Type: application\/json\" \\\n  -d '{\n    \"model\": \"Qwen\/Qwen2-1.5B-Instruct\",\n    \"messages\": [{\"role\": \"user\", \"content\": \"\u4ecb\u7ecd\u4e00\u4e0bKubernetes\"}],\n    \"max_tokens\": 512\n  }'\n<\/code><\/pre>\n\n\n\n
            3. Open WebUI \u90e8\u7f72<\/h3>\n\n\n\n
            \u6838\u5fc3\u914d\u7f6e\u8981\u70b9<\/h4>\n\n\n\n
              \n
            • \u7aef\u53e3\u6620\u5c04<\/strong>\uff1a\u5bb9\u5668 8080 \u7aef\u53e3\u6620\u5c04\u5230\u8282\u70b9 3000 \u7aef\u53e3<\/li>\n\n\n\n
            • \u6570\u636e\u6301\u4e45\u5316<\/strong>\uff1a\u6302\u8f7d\u7528\u6237\u6570\u636e\u3001\u914d\u7f6e\u3001\u5bf9\u8bdd\u8bb0\u5f55<\/li>\n\n\n\n
            • \u73af\u5883\u53d8\u91cf<\/strong>\uff1a\u5173\u95ed\u8054\u7f51\u641c\u7d22\uff0c\u7eaf\u5185\u7f51\u79bb\u7ebf\u8fd0\u884c<\/li>\n<\/ul>\n\n\n\n
              Deployment \u6838\u5fc3\u7247\u6bb5<\/h4>\n\n\n\n
              spec:\n  replicas: 1\n  containers:\n  - name: open-webui\n    image: 192.168.100.149\/gpuimages\/open-webui:latest\n    ports:\n    - containerPort: 8080\n    env:\n    - name: ENABLE_RAG_WEB_SEARCH\n      value: \"false\"\n    volumeMounts:\n    - name: open-webui-data\n      mountPath: \/app\/backend\/data\n  volumes:\n  - name: open-webui-data\n    hostPath:\n      path: \/data\/open-webui\n<\/code><\/pre>\n\n\n\n
              \u540e\u7aef\u5bf9\u63a5<\/h4>\n\n\n\n
                \n
              • \u5bf9\u63a5 Ollama\uff1a\u8bbe\u7f6e\u4e2d\u586b\u5199http:\/\/ollama-service:11434<\/code><\/li>\n\n\n\n
              • \u5bf9\u63a5 vLLM\uff1a\u5916\u90e8\u8fde\u63a5\u4e2d\u6dfb\u52a0 OpenAI \u63a5\u53e3http:\/\/vllm-service:8000\/v1<\/code><\/li>\n<\/ul>\n\n\n\n
                \n\n\n\n
                \u4e94\u3001\u6c47\u603b<\/h2>\n\n\n\n
                  \n
                • K8s \u4e2d NVIDIA Device Plugin \u7684\u5de5\u4f5c\u539f\u7406\u662f\u4ec0\u4e48\uff1f<\/strong><\/li>\n<\/ul>\n\n\n\n
                  \u63d2\u4ef6\u4ee5 DaemonSet \u8fd0\u884c\u5728\u6bcf\u4e2a GPU \u8282\u70b9\uff0c\u901a\u8fc7 gRPC \u4e0e kubelet \u901a\u4fe1\uff0c\u4e0a\u62a5\u8282\u70b9 GPU \u6570\u91cf\u4e0e\u5065\u5eb7\u72b6\u6001\uff1bK8s \u8c03\u5ea6\u5668\u6839\u636envidia.com\/gpu<\/mark>\u8d44\u6e90\u8c03\u5ea6 Pod\uff1bPod \u521b\u5efa\u65f6\uff0ckubelet \u8c03\u7528\u63d2\u4ef6\u5206\u914d GPU\uff0c\u5c06\u8bbe\u5907\u8282\u70b9\u3001\u9a71\u52a8\u5e93\u52a8\u6001\u6ce8\u5165\u5bb9\u5668\uff0c\u5b9e\u73b0 GPU \u7684\u96c6\u7fa4\u7ea7\u7edf\u4e00\u8c03\u5ea6\u3002<\/p>\n\n\n\n
                    \n
                  • \u4e3a\u4ec0\u4e48\u90e8\u7f72 K8s \u5fc5\u987b\u5173\u95ed Swap\uff1f<\/strong><\/li>\n<\/ul>\n\n\n\n
                    Kubernetes \u8bbe\u8ba1\u8981\u6c42\u8282\u70b9\u5185\u5b58\u4e0d\u8db3\u65f6\u7531\u8c03\u5ea6\u5668\u9a71\u9010 Pod\uff0c\u800c\u4e0d\u662f\u4f7f\u7528 Swap\u3002Swap \u4f1a\u5bfc\u81f4\u8282\u70b9\u6027\u80fd\u4e0d\u7a33\u5b9a\u3001Pod QoS \u5931\u6548\u3001\u8c03\u5ea6\u5224\u65ad\u5931\u771f\uff0c\u56e0\u6b64 kubelet \u9ed8\u8ba4\u5f3a\u5236\u8981\u6c42\u5173\u95ed Swap \u624d\u80fd\u542f\u52a8\u3002<\/p>\n\n\n\n
                      \n
                    • containerd \u4e3a\u4ec0\u4e48\u8981\u914d\u7f6e SystemdCgroup\uff1f<\/strong> <\/li>\n<\/ul>\n\n\n\n
                      kubelet \u9ed8\u8ba4\u4f7f\u7528 SystemdCgroup \u9a71\u52a8\u7ba1\u7406\u5bb9\u5668 cgroup\uff0ccontainerd \u5fc5\u987b\u4e0e\u5176\u4fdd\u6301\u4e00\u81f4\uff0c\u5426\u5219\u4f1a\u51fa\u73b0 cgroup \u9a71\u52a8\u4e0d\u5339\u914d\u9519\u8bef\uff0c\u5bfc\u81f4\u8282\u70b9\u65e0\u6cd5\u6b63\u5e38\u52a0\u5165\u96c6\u7fa4\u3001Pod \u8fd0\u884c\u5f02\u5e38\u3002<\/p>\n\n\n\n
                        \n
                      • vLLM \u90e8\u7f72\u4e3a\u4ec0\u4e48\u9700\u8981\u5927\u5171\u4eab\u5185\u5b58\uff08–shm-size\uff09\uff1f<\/strong> <\/li>\n<\/ul>\n\n\n\n
                        \u7b54\uff1avLLM \u7684 PagedAttention\u3001\u8fde\u7eed\u6279\u5904\u7406\u673a\u5236\u4f9d\u8d56\u5927\u91cf\u8fdb\u7a0b\u95f4\u901a\u4fe1\u548c\u5171\u4eab\u5185\u5b58\u6570\u636e\u4ea4\u6362\uff1b\u5171\u4eab\u5185\u5b58\u4e0d\u8db3\u4f1a\u5bfc\u81f4\u63a8\u7406\u6027\u80fd\u4e0b\u964d\u3001\u9ad8\u5e76\u53d1\u4e0b\u62a5\u9519\uff0c\u56e0\u6b64\u901a\u5e38\u914d\u7f6e\u6570 GB \u5230\u5341\u51e0 GB \u7684\u5171\u4eab\u5185\u5b58\u3002<\/p>\n\n\n\n
                          \n
                        • K8s \u90e8\u7f72 GPU \u5927\u6a21\u578b\u670d\u52a1\u76f8\u6bd4\u88f8 Docker \u6709\u4ec0\u4e48\u4f18\u52bf\uff1f<\/strong> <\/li>\n<\/ul>\n\n\n\n
                          \u96c6\u7fa4\u7edf\u4e00\u8c03\u5ea6 GPU\uff0c\u8d44\u6e90\u5229\u7528\u7387\u66f4\u9ad8\uff1b\u670d\u52a1\u81ea\u6108 + \u81ea\u52a8\u6269\u7f29\u5bb9\uff0c\u7a33\u5b9a\u6027\u66f4\u5f3a\uff1b\u6807\u51c6\u5316\u90e8\u7f72\uff0c\u591a\u673a\u6269\u5c55\u3001\u8fc1\u79fb\u66f4\u65b9\u4fbf\uff1b\u652f\u6301\u670d\u52a1\u53d1\u73b0\u3001\u8d1f\u8f7d\u5747\u8861\u3001\u6eda\u52a8\u66f4\u65b0\uff0c\u66f4\u9002\u5408\u751f\u4ea7\u7ea7\u5927\u89c4\u6a21\u670d\u52a1\u3002<\/p>\n\n\n\n
                          <\/p>\n","protected":false},"excerpt":{"rendered":"
                          \u672c\u7bc7\u805a\u7126 Kubernetes \u96c6\u7fa4\u4e0b\u7684 GPU \u7b97\u529b\u8c03\u5ea6\u4e0e\u5927\u6a21\u578b\u670d\u52a1\u7f16\u6392\uff0c\u638c\u63e1K8s \u96c6\u7fa4\u642d\u5efa\u3001NVIDIA […]<\/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-166","post","type-post","status-publish","format-standard","hentry","category-6"],"_links":{"self":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/166","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=166"}],"version-history":[{"count":5,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/166\/revisions"}],"predecessor-version":[{"id":185,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=\/wp\/v2\/posts\/166\/revisions\/185"}],"wp:attachment":[{"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=166"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=166"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.xiaoxindong.cloud\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=166"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}