Ascend Device Plugin

Introduction

This Ascend device plugin is implemented for HAMi scheduling.

Memory slicing is supported based on virtualization template, lease available template is automatically used. For detailed information, check templeate

Prequisites

ascend-docker-runtime

Compile

make all

Build

docker buildx build -t $IMAGE_NAME .

Deployment

Due to dependencies with HAMi, you need to set

devices.ascend.enabled=true

during HAMi installation. For more details, see 'devices' section in values.yaml.

devices:
  ascend:
    enabled: true
    image: "ascend-device-plugin:master"
    imagePullPolicy: IfNotPresent
    extraArgs: []
    nodeSelector:
      ascend: "on"
    tolerations: []
    resources:
      - huawei.com/Ascend910A
      - huawei.com/Ascend910A-memory
      - huawei.com/Ascend910B
      - huawei.com/Ascend910B-memory
      - huawei.com/Ascend310P
      - huawei.com/Ascend310P-memory

Note that resources here(hawei.com/Ascend910A,huawei.com/Ascend910B,...) is managed in hami-scheduler-device configMap. It defines three different templates(910A,910B,310P).

label your NPU nodes with 'ascend=on'

kubectl label node {ascend-node} ascend=on

Deploy ascend-device-plugin by running

kubectl apply -f ascend-device-plugin.yaml

Usage

You can allocate a slice of NPU by specifying both resource number and resource memory. For more examples, see examples

...
    containers:
    - name: npu_pod
      ...
      resources:
        limits:
          huawei.com/Ascend910B: "1"
          # if you don't specify Asend910B-memory, it will use a whole NPU. 
          huawei.com/Ascend910B-memory: "4096"