Commit 7c8a254

@test "container variant: ccManager and vgpuDeviceManager pinned off" {

TMP=$(mktemp -d); trap 'rm -rf "$TMP"' EXIT

render_variant container

grep -A 1 '^ ccManager:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

grep -A 1 '^ vgpuDeviceManager:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

}

@test "container variant: sandboxWorkloads off, defaultWorkload stays upstream 'container'" {

TMP=$(mktemp -d); trap 'rm -rf "$TMP"' EXIT

render_variant container

grep -A 2 '^ sandboxWorkloads:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

grep -A 2 '^ sandboxWorkloads:' "$TMP/rendered.yaml" | grep -q 'defaultWorkload: container'

}

@test "container variant: driver, toolkit, vfioManager off; devicePlugin on" {

TMP=$(mktemp -d); trap 'rm -rf "$TMP"' EXIT

render_variant container

grep -A 1 '^ driver:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

grep -A 1 '^ toolkit:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

grep -A 1 '^ vfioManager:' "$TMP/rendered.yaml" | grep -q 'enabled: false'

grep -A 1 '^ devicePlugin:' "$TMP/rendered.yaml" | grep -q 'enabled: true'

}

- name: container

dependsOn:

- cozystack.networking

components:

- name: gpu-operator

path: system/gpu-operator

valuesFiles:

- values.yaml

- values-container.yaml

install:

privileged: true

namespace: cozy-gpu-operator

releaseName: gpu-operator

The files in this directory are **not** templates. They are reference artifacts that document one working configuration for running GPU workloads directly in pods on a Talos-based Cozystack cluster, together with the DCGM metrics needed by the `gpu/gpu-performance` Grafana dashboard.

For non-Talos Linux hosts that already provide the NVIDIA driver and `nvidia-container-toolkit` via the distro package manager (the typical apt-installed Ubuntu/Debian shape), pick the first-class `container` variant instead — `variant: container` in your Package CR is enough and no `hostPaths.driverInstallDir` override is needed because the operator-validator auto-detects the host driver at `/host/usr/bin/nvidia-smi`. The files here remain the right starting point on Talos, where the extension installs `nvidia-smi` at `/usr/local/bin/` and the auto-detect path misses it.

The out-of-the-box `values-passthrough.yaml` for this package targets the sandbox (VFIO passthrough to KubeVirt VMs) scenario. The files here illustrate an alternative — running CUDA workloads in regular pods with the NVIDIA device plugin — and the workarounds it currently requires on Talos.

gpu-operator:

sandboxWorkloads:

# Default upstream workload on a node without the per-node

# 'nvidia.com/gpu.workload.config' label is 'container' when

# sandboxWorkloads is off — which is exactly what this variant

# wants. No defaultWorkload override needed; the counterpart

# overrides in values-passthrough.yaml / values-vgpu.yaml exist

# only because those variants flip sandboxWorkloads to true.

enabled: false

driver:

# Host already runs the NVIDIA driver (apt-installed). The

# operator's driver DaemonSet would attempt to load its own

# kernel module and clash with the host one.

enabled: false

devicePlugin:

# Publish nvidia.com/gpu to the kubelet so pods can request GPUs.

enabled: true

vfioManager:

# Defensive: vfioManager unbinds the host driver to bind vfio-pci.

# With the host driver present and devicePlugin enabled, vfio

# binding must not happen. Upstream defaults this to off when

# sandboxWorkloads is disabled, but the explicit pin protects

# against an upstream default flip.

enabled: false

toolkit:

# Host already has nvidia-container-toolkit installed (the typical

# apt-installed shape). Running the operator's toolkit DaemonSet

# would overwrite /etc/containerd/config.toml and the CDI hooks

# already wired by the host package, breaking the host install.

# NVIDIA's documentation recommends disabling this when the

# toolkit is preinstalled on the node.

enabled: false