Installing Furiosa DRA Driver

Installing Furiosa DRA Driver#

Note

Furiosa DRA Driver is in alpha stage, which means its features are experimental and may change in future releases.

Users should be aware that the driver may not be fully stable or feature-complete, and they should exercise caution when deploying it in production environments.

Furiosa DRA Driver#

The Furiosa DRA Driver implements the Kubernetes DRA interface for FuriosaAI NPU devices, and its features are as follows:

Discovering the FuriosaAI NPU devices and registering to a Kubernetes cluster as DRA resources.
Tracking the health of the devices and reporting to a Kubernetes cluster.
Running AI workload on the top of the Furiosa NPU devices within a Kubernetes cluster.

Prerequisites#

The Furiosa DRA Driver requires the following prerequisites:

Kubernetes 1.34+ is required.
- In this case, CDI must be enabled on each node.
- Starting from containerd 2.0, CDI is enabled by default.

Deploying Furiosa DRA Driver with Helm#

The Furiosa DRA Driver Helm chart is available at furiosa-ai/helm-charts.

To configure deployment as you need, you can modify charts/furiosa-dra-driver/values.yaml.

You can deploy the Furiosa DRA Driver by running the following commands:

helm repo add furiosa https://furiosa-ai.github.io/helm-charts
helm repo update
helm install furiosa-dra-driver furiosa/furiosa-dra-driver -n furiosa-system

When DRA Driver is successfully deployed, you can see the registered FuriosaAI NPU resources by running the following command:

kubectl get resourceslices

- apiVersion: resource.k8s.io/v1
  kind: ResourceSlice
  metadata:
    creationTimestamp: "2026-01-12T02:57:12Z"
    generateName: cntk002-npu.furiosa.ai-
    generation: 1
    name: cntk002-npu.furiosa.ai-stczm
    ownerReferences:
      - apiVersion: v1
        controller: true
        kind: Node
        name: cntk002
        uid: 0a0a5366-cb39-4e8d-9aae-56a885709887
    resourceVersion: "8934931"
    uid: 8909a748-d798-4e98-b738-5c08d771e011
  spec:
    devices:
      - attributes:
          arch:
            string: rngd
          driverVersion:
            version: 2025.4.0-dev2+0e9e56a
          index:
            int: 0
          numaNode:
            int: 0
          resource.kubernetes.io/pciBusID:
            string: "0000:17:00.0"
          resource.kubernetes.io/pcieRoot:
            string: "pci0000:16"
          uuid:
            string: 09512C86-0709-4404-8943-414B404A4F44
        capacity:
          core:
            value: "8"
          memory:
            value: 48Gi
        name: npu0
      - attributes:
          arch:
            string: rngd
          driverVersion:
            version: 2025.4.0-dev2+0e9e56a
          index:
            int: 1
          numaNode:
            int: 0
          resource.kubernetes.io/pciBusID:
            string: "0000:43:00.0"
          resource.kubernetes.io/pcieRoot:
            string: "pci0000:42"
          uuid:
            string: B65D0C77-7333-4D90-AC82-3E22EE5F88EE
        capacity:
          core:
            value: "8"
          memory:
            value: 48Gi
        name: npu1
    driver: npu.furiosa.ai
    nodeName: cntk002
    pool:
      generation: 1
      name: cntk002
      resourceSliceCount: 1

Available attributes#

Below table lists available attributes for FuriosaAI NPU devices.

You need to attach npu.furiosa.ai at the front of each attribute key when you use them in ResourceClaimTemplate.

The only exceptions are resource.kubernetes.io/pcieRoot and resource.kubernetes.io/pciBusID, which are standard attributes defined by Kubernetes.

Attribute	Type	Description
`index`	int	The index of the NPU device
`uuid`	string	The UUID of the NPU device
`arch`	string	The architecture of the NPU device
`driverVersion`	string	The driver version represented in semver
`numaNode`	int	The NUMA node value that NPU device belongs to
`resource.kubernetes.io/pciBusID`	string	The PCIe bus ID of the NPU device
`resource.kubernetes.io/pcieRoot`	string	The pcieRoot value of the NPU device

Example Usage#

Request single NPU#

Below example shows how to request single FuriosaAI NPU in a Pod spec with ResourceClaimTemplate.

First, create a ResourceClaimTemplate that defines a request for a single NPU device. Then, create a Pod that references the template to claim the NPU resource.

apiVersion: resource.k8s.io/v1
kind: ResourceClaimTemplate
metadata:
 name: single-npu
spec:
 spec:
   devices:
     requests:
       - name: npu
         exactly:
           count: 1
           deviceClassName: npu.furiosa.ai

apiVersion: v1
kind: Pod
metadata:
  name: furiosa-llm
spec:
  containers:
    - name: furiosa-llm
      image: furiosaai/furiosa-llm:latest
      command:
        - sleep
        - "86400"
      env:
        - name: HF_TOKEN
          valueFrom:
            secretKeyRef:
              key: token
              name: hf-token-secret
      resources:
        claims:
          - name: npu
  resourceClaims:
    - name: npu
      resourceClaimTemplateName: single-npu
---
apiVersion: v1
kind: Secret
metadata:
  name: hf-token-secret
type: Opaque
data:
  token: <your_base64_encoded_hf_token>

Request quad NPU on same NUMA node#

Similar to the previous example, below example shows how to request four FuriosaAI NPUs that are located on the same NUMA node in a Pod spec.

First, create a ResourceClaimTemplate that defines a request for four FuriosaAI NPUs on the same NUMA node. Then, create a Pod that references the template to claim the NPU resource.

apiVersion: resource.k8s.io/v1
kind: ResourceClaimTemplate
metadata:
  name: quad-npu-with-same-numa-node
spec:
  spec:
    devices:
      constraints:
        - matchAttribute: npu.furiosa.ai/numaNode
      requests:
        - name: npu
          exactly:
            count: 4
            deviceClassName: npu.furiosa.ai

apiVersion: v1
kind: Pod
metadata:
  name: furiosa-llm
spec:
  containers:
    - name: furiosa-llm
      image: furiosaai/furiosa-llm:latest
      command:
        - sleep
        - "86400"
      env:
        - name: HF_TOKEN
          valueFrom:
            secretKeyRef:
              key: token
              name: hf-token-secret
      resources:
        claims:
          - name: npu
  resourceClaims:
    - name: npu
      resourceClaimTemplateName: quad-npu-with-same-numa-node
---
apiVersion: v1
kind: Secret
metadata:
  name: hf-token-secret
type: Opaque
data:
  token: <your_base64_encoded_hf_token>