Notice about component imagesThis container is part of a product or application and is not meant to be installed individually. Instead, it should be installed via an operator or other automated source. See the overview tab for links to any products using this container to learn more about installation methods.Image identifiers
Overview
The image runs a single binary that registers with the kubelet, advertises the FuriosaAI NPUs on the node as allocatable resources, and reports their health. It must run on nodes that have FuriosaAI NPUs and needs access to the host's device plugin socket directory so it can register with the kubelet and detect kubelet restarts. The container starts the binary automatically through its default command.
Configuration
Behavior is controlled by a small configuration structure:
config:
partitioning: none
debugMode: false
- partitioning selects how each NPU card is exposed (none, 2core.12gb, or 4core.24gb), which in turn sets the resource name pods request. The default is none.
- debugMode enables verbose debug logging. The default is false.
Specific devices can be disabled on a per-node basis with disabledDeviceUUIDs. The UUID of a device can be obtained with furiosa-smi info:
config:
partitioning: none
debugMode: false
disabledDeviceUUIDs:
node_a:
- "uuid1"
- "uuid2"
node_b:
- "uuid3"
- "uuid4"
If disabledDeviceUUIDs is not specified, the default is an empty list.
Deploying with Helm
The recommended way to deploy on Kubernetes is with the official Helm chart, published at https://github.com/furiosa-ai/helm-charts. Deployment options can be
helm repo add furiosa https://furiosa-ai.github.io/helm-charts
helm repo update
helm install furiosa-device-plugin furiosa/furiosa-device-plugin -n kube-system
Using the NPUs in a workload
Once the plugin is running, request NPUs from a pod using the resource name that matches the configured partitioning mode. For example, with the default none partitioning:
resources:
limits:
furiosa.ai/rngd: "1"
Verifying
After deployment, confirm that the NPU resources are advertised on the nodes:
kubectl get nodes -o custom-columns=NAME:.metadata.name,RNGD:.status.allocatable.'furiosa\.ai/rngd'
Nodes with FuriosaAI NPUs should report a non-zero count for the corresponding resource, after which pods can be scheduled against it.
