Enable Enflame GCU Sharing

Introduction

HAMi now supports Enflame DRS hard partition scheduling, aligned with Enflame native scheduler behavior.

DRS is a hard-slice mode similar to NVIDIA MIG and Ascend VNPU templates.

Prerequisites

• Enflame gcushare-device-plugin >= 2.2.14 (please consult your device provider; only gcushare-device-plugin is needed, do not install gcushare-scheduler-plugin)
• driver version >= 1.8.7
• kubernetes >= 1.24
• enflame-container-toolkit >= 2.0.50

Enable Enflame DRS Scheduling

• Deploy gcushare-device-plugin on Enflame nodes (please consult your device provider to acquire its package and documentation)

caution

Install only gcushare-device-plugin. Do not install the gcushare-scheduler-plugin package.

Follow Online installation to deploy HAMi. Keep devices.enflame.enabled=true (the default). This registers the vendor resources with the bundled scheduler extender; the vendor plugin or toolkit must be installed separately. See the Chart values reference for devices.enflame.*.

note

The default resource names are:

• enflame.com/drs-gcu for direct DRS slice requests
• enflame.com/gcu-memory for memory requests (in MiB)
• enflame.com/gcu-core for core requests (as a percentage)

Customize these names through devices.enflame.resourceNameDRSGCU, resourceNameGCUMemory, and resourceNameGCUCore in Helm values, keeping them consistent with the vendor plugin. Do not manually edit the global ConfigMap; see Global device configuration.

Running DRS Workloads

HAMi supports two request styles:

1. Direct DRS slice request (enflame.com/drs-gcu)
2. Unified memory/core request (enflame.com/gcu-memory + enflame.com/gcu-core); HAMi converts it to a DRS profile internally.

Direct DRS slice request

apiVersion: v1
kind: Pod
metadata:
  name: gcushare-pod-drs
  namespace: kube-system
spec:
  terminationGracePeriodSeconds: 0
  containers:
    - name: pod-gcu-example1
      image: ubuntu:22.04
      imagePullPolicy: IfNotPresent
      command: ["sleep"]
      args: ["100000"]
      resources:
        limits:
          enflame.com/drs-gcu: 3
        requests:
          enflame.com/drs-gcu: 3

Request by memory/core (recommended unified API)

apiVersion: v1
kind: Pod
metadata:
  name: gcushare-pod-by-spec
  namespace: kube-system
spec:
  terminationGracePeriodSeconds: 0
  containers:
    - name: pod-gcu-example1
      image: ubuntu:22.04
      imagePullPolicy: IfNotPresent
      command: ["sleep"]
      args: ["100000"]
      resources:
        limits:
          enflame.com/gcu-memory: 20480 # MiB
          enflame.com/gcu-core: 40 # percent
        requests:
          enflame.com/gcu-memory: 20480
          enflame.com/gcu-core: 40

tip

More examples are available in the examples/enflame folder.

Scheduling Annotations

During scheduling, HAMi writes DRS-compatible annotations such as:

• assigned-containers
• enflame.com/gcu-assigned
• enflame.com/gcu-assigned-index
• enflame.com/gcu-assigned-minor
• enflame.com/gcu-request-size

These annotations are consumed by the Enflame device plugin during the Allocate phase.