HAMi 多设备使用指南

# values.yaml
devices:
  # NVIDIA GPU - 默认启用
  nvidia:
    gpuCorePolicy: default
    libCudaLogLevel: 1

  # 寒武纪 MLU - 通过 customresources 自动启用
  # (只要 Scheduler 能发现节点上的 MLU 资源即可)

  # 华为 Ascend - 需要单独启用和部署
  ascend:
    enabled: false              # 设为 true 启用
    image: ""                   # Ascend Device Plugin 镜像
    imagePullPolicy: IfNotPresent
    extraArgs: []
    nodeSelector:
      ascend: "on"              # Ascend 节点标签
    tolerations: []
    customresources:
      - huawei.com/Ascend910A
      - huawei.com/Ascend910A-memory
      - huawei.com/Ascend910B2
      - huawei.com/Ascend910B2-memory
      # ...

  # 天数智芯 - 需要显式启用
  iluvatar:
    enabled: false
    customresources:
      - iluvatar.ai/BI-V100-vgpu
      - iluvatar.ai/BI-V100.vCore
      - iluvatar.ai/BI-V100.vMem
      # ...

  # 其他默认启用的设备
  kunlun:
    enabled: true
  enflame:
    enabled: true
  mthreads:
    enabled: true

# values.yaml
mluResourceName: "cambricon.com/vmlu"
mluResourceMem: "cambricon.com/mlu.smlu.vmemory"
mluResourceCores: "cambricon.com/mlu.smlu.vcore"

apiVersion: v1
kind: Pod
metadata:
  name: mlu-inference
  annotations:
    cambricon.com/use-mlutype: "MLU370"    # 可选: 指定 MLU 型号
spec:
  containers:
  - name: mlu-app
    image: my-mlu-inference:latest
    resources:
      limits:
        cambricon.com/vmlu: 1                    # 1 个虚拟 MLU
        cambricon.com/mlu.smlu.vmemory: 16       # 4GB 显存 (16 x 256MB)
        cambricon.com/mlu.smlu.vcore: 50         # 50% 算力

# values.yaml
devices:
  ascend:
    enabled: true                     # 启用 Ascend 支持
    image: "projecthami/ascend-device-plugin:latest"
    imagePullPolicy: IfNotPresent
    nodeSelector:
      ascend: "on"                    # Ascend 节点标签
    tolerations: []
    customresources:
      - huawei.com/Ascend910A
      - huawei.com/Ascend910A-memory
      - huawei.com/Ascend910B2
      - huawei.com/Ascend910B2-memory
      - huawei.com/Ascend910B3
      - huawei.com/Ascend910B3-memory
      - huawei.com/Ascend910B4
      - huawei.com/Ascend910B4-memory
      - huawei.com/Ascend910B4-1
      - huawei.com/Ascend910B4-1-memory
      - huawei.com/Ascend310P
      - huawei.com/Ascend310P-memory

# Ascend 910B3 推理任务
apiVersion: v1
kind: Pod
metadata:
  name: ascend-inference
spec:
  containers:
  - name: mindspore-app
    image: ascendhub/ascend-mindspore:latest
    command: ["python", "inference.py"]
    resources:
      limits:
        huawei.com/Ascend910B3: "1"            # 1 个 Ascend 910B3
        huawei.com/Ascend910B3-memory: "16384"  # 16GB 显存

# Ascend 310P 边缘推理
apiVersion: v1
kind: Pod
metadata:
  name: ascend-edge-inference
spec:
  containers:
  - name: edge-app
    image: ascend-edge-inference:latest
    resources:
      limits:
        huawei.com/Ascend310P: "1"
        huawei.com/Ascend310P-memory: "4096"

resources:
  limits:
    hygon.com/dcunum: 1
    hygon.com/dcumem: 4096
    hygon.com/dcucores: 50

resources:
  limits:
    iluvatar.ai/BI-V100-vgpu: 1
    iluvatar.ai/BI-V100.vCore: 50
    iluvatar.ai/BI-V100.vMem: 4096

resources:
  limits:
    metax-tech.com/sgpu: 1
    metax-tech.com/vcore: 50
    metax-tech.com/vmemory: 4096

resources:
  limits:
    mthreads.com/vgpu: 1

resources:
  limits:
    kunlunxin.com/vxpu: 1
    kunlunxin.com/vxpu-memory: 4096

resources:
  limits:
    enflame.com/vgcu: 1
    enflame.com/vgcu-percentage: 50

resources:
  limits:
    amd.com/gpu: 1
    amd.com/gpu-memory: 8192

resources:
  limits:
    aws.amazon.com/neuroncore: 2

type Devices interface {
    // 设备标识
    CommonWord() string

    // Webhook 阶段 - 准入变更
    MutateAdmission(ctr *corev1.Container, pod *corev1.Pod) (bool, error)

    // 健康检查
    CheckHealth(devType string, n *corev1.Node) (bool, bool)

    // 节点设备发现
    GetNodeDevices(n corev1.Node) ([]*DeviceInfo, error)

    // 节点锁管理
    LockNode(n *corev1.Node, p *corev1.Pod) error
    ReleaseNodeLock(n *corev1.Node, p *corev1.Pod) error

    // 资源请求解析
    GenerateResourceRequests(ctr *corev1.Container) ContainerDeviceRequest

    // 注解管理
    PatchAnnotations(pod *corev1.Pod, annoinput *map[string]string, pd PodDevices) map[string]string

    // 评分
    ScoreNode(node *corev1.Node, podDevices PodSingleDevice,
              previous []*DeviceUsage, policy string) float32

    // 设备适配
    Fit(devices []*DeviceUsage, request ContainerDeviceRequest,
        pod *corev1.Pod, nodeInfo *NodeInfo, allocated *PodDevices) (bool, map[string]ContainerDevices, string)

    // NUMA 支持
    AssertNuma(annos map[string]string) bool

    // 节点清理
    NodeCleanUp(nn string) error

    // 资源使用量更新
    AddResourceUsage(pod *corev1.Pod, n *DeviceUsage, ctr *ContainerDevice) error

    // 资源名称获取
    GetResourceNames() ResourceNames
}

func (dev *MyDevices) GetNodeDevices(n corev1.Node) ([]*device.DeviceInfo, error) {
    // 1. 从 Node Capacity 中读取设备数量
    cards, ok := n.Status.Capacity.Name(
        corev1.ResourceName(MyResourceCores),
        resource.DecimalSI,
    ).AsInt64()

    // 2. 读取显存总量
    memoryTotal, _ := n.Status.Capacity.Name(
        corev1.ResourceName(MyResourceMemory),
        resource.DecimalSI,
    ).AsInt64()

    // 3. 构建 DeviceInfo 列表
    devices := []*device.DeviceInfo{}
    for i := 0; i < numDevices; i++ {
        devices = append(devices, &device.DeviceInfo{
            Index:        uint(i),
            ID:           fmt.Sprintf("%s-mydevice-%d", n.Name, i),
            Count:        100,
            Devmem:       int32(memoryPerDevice),
            Devcore:      100,
            Type:         MyDeviceType,
            Health:       true,
            DeviceVendor: MyCommonWord,
        })
    }
    return devices, nil
}

func (dev *MyDevices) Fit(
    devices []*device.DeviceUsage,
    request device.ContainerDeviceRequest,
    pod *corev1.Pod,
    nodeInfo *device.NodeInfo,
    allocated *device.PodDevices,
) (bool, map[string]device.ContainerDevices, string) {
    // 1. 遍历候选设备
    // 2. 检查设备类型匹配
    // 3. 检查显存余量
    // 4. 检查算力余量
    // 5. 检查 UUID 白名单/黑名单
    // 6. 返回分配方案
}