其实坑还挺多的，尤其是 immich 还需要使用特定版本的 pgvecto 支持的数据库。加之如果需要硬件加速，还需要额外的配置，这里记录一下防止我忘了。

Kubernetes Device Plugin 与 GPU Device

负载机用的 n5095 核显直通，是个 intel 设备，在用 helm 部署 immich 的情况中，直接 /dev/dri 直通并不是很方便，不过查了一下可以通过 Device Plugin 的形式，更好地管理设备。

# https://github.com/intel/intel-device-plugins-for-kubernetes/blob/main/cmd/gpu_plugin/README.md

apiVersion: kustomize.config.k8s.io/v1beta1
kind: Kustomization
resources:
  # NFD base
  - https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/nfd?ref=v0.34.0
  # NodeFeatureRules for GPU detection
  - https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/nfd/overlays/node-feature-rules?ref=v0.34.0
  # GPU plugin daemonset
  - https://github.com/intel/intel-device-plugins-for-kubernetes/deployments/gpu_plugin/overlays/nfd_labeled_nodes?ref=v0.34.0
patches:
  - target:
      kind: DaemonSet
      name: intel-gpu-plugin
    patch: |-
      - op: replace
        path: /spec/template/spec/containers/0/args
        value:
          - "-enable-monitoring"
          - "-v=2"
          - "-shared-dev-num=5"

相较于官方的配置，我加入了 -share-dev-num=5, 使 GPU 可以在至多 5 个容器中共享，默认配置只能挂载一个。

PSQL Cluster

immich 用到特定版本的数据库，可以用这种方式部署。

apiVersion: v1
kind: Secret
metadata:
  name: immich-postgres-user
type: Opaque
stringData:
  username: immich
  password: your-strong-password
---
# https://github.com/immich-app/immich-charts/issues/149
apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: immich-postgres
spec:
  # At the time of writing, immich is only compatible with pgvecto.rs <0.4. Latest postgres image with that version is 16.5.
  imageName: ghcr.io/tensorchord/cloudnative-pgvecto.rs:16.5-v0.3.0@sha256:be3f025d79aa1b747817f478e07e71be43236e14d00d8a9eb3914146245035ba
  instances: 1

  postgresql:
    shared_preload_libraries:
      - "vectors.so"

  managed:
    roles:
      - name: immich
        superuser: true
        login: true

  bootstrap:
    initdb:
      database: immich
      owner: immich
      secret:
        name: immich-postgres-user
      postInitSQL:
        - CREATE EXTENSION IF NOT EXISTS "vectors";
        - CREATE EXTENSION IF NOT EXISTS "cube" CASCADE;
        - CREATE EXTENSION IF NOT EXISTS "earthdistance" CASCADE;

  storage:
    size: 4Gi
    storageClass: nfs-client

Values 参考

再手动创建一个 PVC 后，就可以进行部署了，注意 tag 的设置，选用合适的加速后端即可：

env:
  DB_HOSTNAME: "immich-postgres-rw"
  DB_USERNAME: "immich"
  DB_DATABASE_NAME: "immich"
  DB_PASSWORD: "your-strong-password"

image:
  tag: v1.143.1

immich:
  metrics:
    enabled: false
  persistence:
    library:
      existingClaim: immich-pvc

redis:
  enabled: true

server:
  resources:
    limits:
      cpu: 1000m
      gpu.intel.com/i915: 1
  ingress:
    main:
      enabled: true
      hosts:
        - host: immich.core
          paths:
            - path: "/"
      tls: []

machine-learning:
  image:
    tag: v1.143.1-openvino
  resources:
    limits:
      cpu: 1000m
      gpu.intel.com/i915: 1
  persistence:
    cache:
      type: pvc
      storageClass: nfs-client