Prerequisites

Cluster, node, and tooling requirements before installing the driver.

Cluster, software, and hardware requirements for the DRA Driver for NVIDIA GPUs.

Tip

Most of these prerequisites can be installed and managed for you by the NVIDIA GPU Operator.

Requirement	Version / Notes
Kubernetes	v1.34.2 or later, with at least one node that has one or more NVIDIA GPUs. The use of DRA became GA in Kubernetes v1.34+ and earlier versions required the `DynamicResourceAllocation` feature gate.
Helm	v3.8 or later.
NVIDIA Driver	v565 or later for GPU allocation. v570.158.01 or later if using ComputeDomains.
CDI	Enabled in your container runtime. This is enabled by default in containerd 2.0+ and CRIO v1.27+. The DRA Driver uses CDI to expose GPUs to containers.
Node Feature Discovery (NFD)	Labels GPU nodes in the cluster. The DRA Driver uses these labels to target the GPU kubelet plugin to the correct nodes.

ComputeDomains additional prerequisites

If you plan to use ComputeDomains, you also need:

NVIDIA Driver v570.158.01 or later. The IMEXDaemonsWithDNSNames feature gate is enabled by default and requires this driver version. The ComputeDomain plugin will fail to start on older drivers unless IMEXDaemonsWithDNSNames is explicitly disabled.
Multi-Node NVLink (MNNVL) hardware. Nodes must be connected via NVLink fabric, such as GB200 NVL72 and similar systems.
GPU Feature Discovery (GFD) deployed via the GPU Operator. GFD generates the nvidia.com/gpu.clique node labels required by ComputeDomains.
On all GPU nodes where the nvidia-imex-* packages are installed, the nvidia-imex.service systemd unit must be disabled:

systemctl disable --now nvidia-imex.service && systemctl mask nvidia-imex.service

Host-managed IMEX

By default the driver owns the nvidia-imex daemon lifecycle, per the requirement above. For clusters where the operator already runs nvidia-imex as a host service (for example through systemd), set resources.computeDomains.imex.mode=hostManaged to stop the driver from creating per-ComputeDomain nvidia-imex DaemonSets. This inverts the requirement above: nvidia-imex.service must be installed, configured, and left running on every participating GPU node. Also, the nvidia-imex service should be configured with IMEX_CMD_ENABLED=1 and IMEX_CMD_UNIX_DOMAIN_PATH=/etc/nvidia-imex/imex_ctrl.sock (configurable via Helm).

Host-managed mode requires two Helm values together:

--set featureGates.HostManagedIMEXDaemon=true \
--set resources.computeDomains.imex.mode=hostManaged

featureGates.HostManagedIMEXDaemon is an alpha gate that only unlocks setting imex.mode=hostManaged. It does not by itself change any behavior. Setting imex.mode=hostManaged without this gate enabled is a startup validation error (helm install/helm template fails immediately, and if bypassed, the controller and kubelet plugin pods will fail to start).
imex.isolation selects the IMEX isolation strategy, and applies under both driverManaged and hostManaged mode. It must be set to one of:
- domain (default) — all workloads running in the same IMEX domain share the same channel (0). Under driverManaged mode this is inherent to the model (the driver creates one nvidia-imex daemon per ComputeDomain); under hostManaged mode, multiple ComputeDomains can still run against the same host IMEX domain, all receiving channel 0, with no isolation between them.
- channel — intended to eventually give each workload a unique channel within an IMEX domain. Not implemented yet: setting it is a startup validation error regardless of imex.mode.
- Any other value is a startup validation error.
Changing imex.mode or imex.isolation on a cluster with active ComputeDomain workloads is not supported and is not enforced by the driver, drain and remove existing ComputeDomains first.

Host IMEX readiness check (Unix domain socket required)

Because there is no driver-managed daemon to report readiness under host-managed IMEX, the kubelet plugin validates the host nvidia-imex daemon itself at channel claim prepare time: it runs nvidia-imex-ctl -q --u=<socket path> chrooted into the driver root and requires a READY response before handing out a channel. A claim fails, and is retried, rather than succeeding silently, if the host daemon isn’t reachable or ready.

To enable it:

On every participating GPU node, add to /etc/nvidia-imex/config.cfg:
```
IMEX_CMD_ENABLED=1
IMEX_CMD_UNIX_DOMAIN_PATH=/etc/nvidia-imex/imex_ctrl.sock
```
then restart nvidia-imex.service so it creates the socket file (IMEX config changes require a restart to take effect).
nvidia-imex-ctl must be present under the driver root alongside nvidia-smi (it ships with the nvidia-imex package).
If a non-default socket path is used, set resources.computeDomains.imex.hostSocketPath (Helm value, plumbed to the kubelet plugin as IMEX_HOST_SOCKET_PATH) to match. It defaults to /etc/nvidia-imex/imex_ctrl.sock.

Install prerequisites with NVIDIA GPU Operator

The NVIDIA GPU Operator is a Kubernetes operator that automates the deployment and lifecycle management of all NVIDIA software components needed to provision and monitor GPUs in a cluster.

It can manage the following DRA Driver for NVIDIA GPUs prerequisites for you:

NVIDIA Driver (v565+ for GPU allocation, v570.158.01+ for ComputeDomains). The GPU Operator installs a default driver that meets the DRA Driver’s prerequisites. To use a specific version, see Common chart customization options in the GPU Operator documentation.
CDI enabled through the NVIDIA Container Toolkit.
Node Feature Discovery (NFD).
GPU Feature Discovery (GFD), required for ComputeDomains.

If you choose to install the GPU Operator, follow the DRA Driver for NVIDIA GPUs install guide in the GPU Operator documentation. It covers installing the GPU Operator with the NVIDIA Kubernetes Device Plugin disabled and installing the DRA Driver for NVIDIA GPUs.

Last modified July 2, 2026: Add readiness checks for the host managed IMEX daemon using a unix socket (f914c8f1)