NVIDIA one-click install

Gryvia turns NVIDIA operators on. It does not replace them. None of this path has been proven on real GPUs, NVSwitch or InfiniBand in this repository.

Base (already in install-k3s-gpu.sh)

sudo ./scripts/install-k3s-gpu.sh server --api-key '<secret>'

GPU Operator: driver, toolkit, device plugin, DCGM. Checklist: docs/gpu-validation.md.

Training fabric

helm upgrade gryvia oci://ghcr.io/zyvorai/charts/gryvia -n gryvia-system \
  --set nvidia.enabled=true \
  --set nvidia.driver.rdma.enabled=true \
  --set nvidia.gds.enabled=true \
  --set nvidia.mig.strategy=mixed \
  --set nvidia.migManager.enabled=true \
  --set nvidiaNetwork.enabled=true \
  --set nvidiaNetwork.nicClusterPolicy.enabled=true \
  --set nvidiaPlatform.timeSlicing.enabled=true \
  --set nvidia.devicePlugin.config.name=gryvia-time-slicing \
  --set nvidiaPlatform.validate.enabled=true

With nicClusterPolicy.enabled the chart renders the NicClusterPolicy itself; the Network Operator chart installs or upgrades its CRDs in a pre-install/pre-upgrade hook first. To manage the policy by hand instead, leave the flag off and kubectl apply -f examples/nvidia/nic-cluster-policy.yaml.

Inference (NIM)

helm upgrade gryvia ... \
  --set nvidiaNim.enabled=true \
  --set nvidiaNim.ngcApiKey="$NGC_API_KEY"
kubectl apply -f examples/nvidia/nim-service.yaml

k8s-nim-operator is pinned to 2.0.1 in Chart.yaml (NGC publishes it without a v). The example NIMService is not a licensed catalog and needs an image pull secret ngc-secret (see the comment in the example). The API key passed with --set is stored in the Helm release Secret; create the ngc-api Secret yourself if that is not acceptable.

Flags

Flag	What it actually starts
nvidia.enabled	GPU Operator
nvidia.driver.rdma.enabled	GPUDirect RDMA in the driver container
nvidia.gds.enabled	nvidia-fs (GDS)
nvidia.mig.strategy / migManager.enabled	MIG manager
nvidiaPlatform.timeSlicing	ConfigMap gryvia-time-slicing
nvidiaNetwork.enabled	Network Operator
nvidiaNetwork.nicClusterPolicy.enabled	NicClusterPolicy (DOCA, RDMA shared plugin, Multus)
nvidiaNim.enabled	NIM Operator
nvidiaPlatform.validate	post-install nvidia-smi Job, not NCCL

Sharing policy labels the GPU Operator already watches. The controller that writes them is opt-in: set gpuOperator.gpuSharing=true (flag --enable-gpu-sharing), otherwise a GryviaGPUSharingPolicy does nothing:

• time-slicing → nvidia.com/device-plugin.config=gryvia-time-slicing. That key's replica count is the chart's nvidiaPlatform.timeSlicing.replicas, not the policy's maxPodsPerGPU. Set gpuOperator.devicePluginConfig=true and the policy writes its own key gryvia-ts-<maxPodsPerGPU> into the ConfigMap and labels nodes with it (Ready=False ConfigWriteFailed if the write fails). Whether the device plugin then advertises that many GPUs is not verified without a GPU node.
• mig → nvidia.com/mig.config=<first profile> (must be a mig-parted name)

Fabric Manager

There is no flag for it. On NVSwitch/HGX nodes the GPU Operator driver container runs Fabric Manager; look for it inside the nvidia-driver-daemonset pod rather than for a new DaemonSet.