Skip to content

Cloud Native Support

These components let you use LPUs in containers and Kubernetes the way you use GPUs. Containers receive cards through the standard Container Device Interface (CDI), and Kubernetes workloads request cards as hyperaccel.ai/lpu resources.

Environment Installation Documentation
Docker, containerd, CRI-O (without Kubernetes) The hyperaccel-lpu package on the host Container support
Kubernetes hyperaccel-lpu on the nodes and the ha-lpu-operator Helm chart in the cluster Kubernetes support

Components

Component What it does Version
ha-container-toolkit (ha-ctk) Discovers the cards on a node, generates the CDI specification (/etc/cdi/lpu.yaml), and configures the container runtime to use CDI 0.1.8
bertha-device-plugin Advertises cards as the Kubernetes resource hyperaccel.ai/lpu and passes the cards allocated to a Pod through CDI. Excludes cards that the driver reports as faulty from allocation 0.1.8
bertha-feature-discovery Adds node labels for card count, NUMA, and PCIe switch topology 0.1.5
ha-webhook Adds a preference that steers Pods requesting multiple cards toward nodes with enough cards under the same PCIe switch 0.1.5
ha-lpu-exporter Exposes per-card health and usage as Prometheus metrics 0.1.7
ha-lpu-operator Installs and manages these components through a single BerthaCluster resource Chart 0.1.8 (image 0.1.6)

Install the driver (the ha kernel module) separately on the node host, not inside a container.

How LPUs become available to a Pod

A Pod requests the number of LPUs it needs through the hyperaccel.ai/lpu resource rather than specifying device files directly. The following steps explain the flow from preparing a node to connecting the allocated LPUs to the container.

  1. Prepare the host to use LPUs.

    The ha kernel module creates a device file for each LPU, such as /dev/ha0 or /dev/ha1. ha-ctk records the device files, user-mode driver libraries, and other information needed by the container in the CDI specification (/etc/cdi/lpu.yaml). The specification is regenerated when the LPU configuration changes.

  2. Report the node's LPU information to Kubernetes.

    The remaining components start after the validator checks driver health. feature-discovery adds node labels for the LPU count and NUMA and PCIe switch topology, while device-plugin reports the LPU count and each LPU's NUMA node to kubelet. This information is used to select a node for the Pod.

  3. Select a node that meets the Pod's request.

    When a Pod requests hyperaccel.ai/lpu, the scheduler selects a node based on available LPUs and node requirements. For Pods requesting multiple LPUs, the webhook adds a preference for nodes with enough LPUs under the same PCIe switch. This is a preference, so the Pod can still be scheduled on another node if it meets the other requirements.

  4. Connect the allocated LPUs to the container.

    The allocated LPUs become available to the container through device-plugin and CDI. The container can then access the device files and user-mode driver libraries for those LPUs, allowing the application to use them.