在 K8s 上部署#

本文档介绍如何在Kubernetes环境中部署Xinference企业版。

前置准备#

创建用于私有镜像拉取的 Docker Registry Secret#

kubectl create secret docker-registry xinference-regcred \
  --docker-server=registry.cn-hangzhou.aliyuncs.com \
  --docker-username=qin@qinxuye.me \
  --docker-password=cre.uwd3nyn4UDM6fzm \
  --docker-email=qin@qinxuye.me \
  --namespace=xinference

配置文件#

准备 values-xinf-enterprise.yaml 文件#

###############################################################################################
#
# Xinference Enterprise deployment configuration
# Two workers, each using one GPU
#
###############################################################################################

# Common configurations
config:
  xinference_image: "registry.cn-hangzhou.aliyuncs.com/xinference-prod/xinference-prod:0.0.11-nvidia"
  curl_image: curlimages/curl:8.8.0
  image_pull_policy: "IfNotPresent"
  imagePullSecrets:
    - name: xinference-regcred
  worker_num: 2  # 根据实际情况调整 worker 的个数
  model_src: "modelscope"
  persistence:
    enabled: true
    mountPath: "/data"
  extra_envs: {}

# Storage configuration
storageClass:
  name: local-storage
  spec:
    provisioner: kubernetes.io/no-provisioner
    volumeBindingMode: WaitForFirstConsumer

pv:
  accessModes:
    - ReadWriteMany
  capacity:
    storage: 500Gi
  hostPath:
    path: /mnt/xinference
  persistentVolumeReclaimPolicy: Retain
  storageClassName: "local-storage"

pvc:
  accessModes:
    - ReadWriteMany
  sharedVolumeClaim:
    storageRequest: 500Gi
  storageClassName: "local-storage"
  volumeMode: "Filesystem"

# Service configurations
serviceWeb:
  ports:
  - name: frontend-port
    nodePort: 30003
    port: 8000
    protocol: TCP
    targetPort: 8000
  - name: api-port
    nodePort: 30004
    port: 9997
    protocol: TCP
    targetPort: 9997
  type: NodePort

serviceSupervisor:
  ports:
  - name: service-supervisor-oscar
    port: 9999
    protocol: TCP
    targetPort: 9999
  - name: service-supervisor-web
    port: 9997
    protocol: TCP
    targetPort: 9997
  type: ClusterIP

serviceWorker:
  ports:
  - port: 30001
    protocol: TCP
    targetPort: 30001
  type: ClusterIP

xinferenceSupervisor:
  supervisor:
    command:
      - /bin/sh
      - -c
      - "/opt/projects/xinf-enterprise.sh --host $(POD_IP) --port 30004 && xinference-supervisor --host $(POD_IP) --port 9997 --log-level debug"
    ports:
      - containerPort: 9997
        name: web
      - containerPort: 9999
        name: oscar
    resources:
      requests:
        cpu: "1"
        memory: 4Gi

xinferenceWorker:
  strategy:
    type: Recreate
  worker:
    initContainers:
      command: [ 'sh', '-c', "until curl -v http://service-supervisor:9997/v1/address; do echo waiting for supervisor; sleep 1; done" ]
    args:
    - -e
    - http://service-supervisor:9997
    - --host
    - $(POD_IP)
    - --worker-port
    - "30001"
    - --log-level
    - debug
    ports:
      - containerPort: 30001
    resources:  # 根据实际情况调整资源
      requests:
        cpu: "2"
        memory: 8Gi
      limits:
        nvidia.com/gpu: "1"

配置注意点#

  • Worker 个数在 config 里配置；worker 使用的资源在 xinferenceWorker.worker.resources 里定义。包括 CPU、GPU 以及内存大小。

  • /opt/projects/xinf-enterprise.sh --host $(POD_IP) --port 30004 是 Xinference 企业版前端（内部端口 8000，对外映射为 30003）需要连接的后端地址，当 Xinference API（端口9997，默认映射到对外地址是 30004）暴露后，需要将这里的 host 和 port 指定成对外地址。

部署步骤#

创建 Xinference helm charts 服务#

# add repo
helm repo add xinference https://xorbitsai.github.io/xinference-helm-charts

# update indexes and query xinference versions
helm repo update
helm search repo xinference/xinference --devel --versions

# install xinference
helm install xinference xinference/xinference -n xinference \
  --version 0.0.2-v1.3.1.post1 \
  -f values-xinf-enterprise.yaml

验证部署#

使用如下命令看 Xinference supervisor 和 worker 的启动情况：

kubectl get pods -n xinference

访问服务#

根据 values-xinf-enterprise.yaml，默认 XInference 前端对外配置在 supervisor_ip:30003，API 地址在 supervisor_ip:30004。访问 http://supervisor_ip:30003 即可打开 Xinference 企业版服务地址。（supervisor_ip 替换成真实 IP）

相关文档#