feat: optional /dev/fuse injection via CW_INJECT_FUSE_DEVICE

[wimarkuske]

What

Adds an opt-in flag — CW_INJECT_FUSE_DEVICE=true / chart value injectFuseDevice: true — that injects /dev/fuse (with the device-cgroup rw rule) into device-plugin allocations. This lets rootless apptainer / squashfuse run in non-privileged GPU pods (e.g. SUNK slurmd) without a privileged securityContext or a separate fuse device plugin.

Default behavior is unchanged (flag off).

Why

On cgroup v2 + containerd 2.x, mounting /dev/fuse into a container as a hostPath is not sufficient — open("/dev/fuse") is gated by the per-container device cgroup, which denies char 10:229 even for root. Only the kubelet can populate that allow-list at pod creation, via one of: privileged: true (all devices), a DRA driver, or a device plugin DeviceSpec.

Since slurmd already requests this plugin's sunk.coreweave.com/accelerator resource (and the plugin already returns GPU DeviceSpecs under PASS_DEVICE_SPECS), the cheapest path is to append a /dev/fuse DeviceSpec to that same allocation. The kubelet then creates the node and adds the cgroup rule — non-privileged.

Note: cdi.k8s.io/* pod annotations do not work here — containerd 2.x only consumes CDI from the CDIDevices CRI field, not annotations.

Change

• internal/plugin/server.go — apiDeviceSpecs() appends a /dev/fuse DeviceSpec (HostPath/ContainerPath /dev/fuse, rw) when CW_INJECT_FUSE_DEVICE=true.
• deployments/helm/nvidia-device-plugin/values.yaml — new injectFuseDevice: false.
• deployments/helm/nvidia-device-plugin/templates/daemonset-device-plugin.yml — renders the env on the plugin container when the value is set.

Scope / follow-up

Covers the PassDeviceSpecs (volume-mounts) device-list strategy, which is what SUNK uses. CDI device-list strategies allocate via updateResponseForCDI() and are not covered by this PR — adding a fuse CDI device through cdiHandler is a clean follow-up.

Validation (live, CoreWeave SUNK test cluster cw1690 / wm-usw09b, k8s 1.35, containerd 2.1.4)

A patched build of this image was deployed as the sunk-nvidia-device-plugin with CW_INJECT_FUSE_DEVICE=true. On a slurmd pod requesting only sunk.coreweave.com/accelerator: 8 (no extra resource, no privileged):

• /dev/fuse present as crw-rw-rw- (character device)
• unprivileged (uid 65534) open() on /dev/fuse succeeds (was EPERM before)
• apptainer exec docker://alpine … mounts via squashfuse + user namespaces successfully

Helm render verified: env absent by default, present only with injectFuseDevice=true.

Enablement (downstream)

Via SUNK chart values (subchart passthrough), once the SUNK nvidia-device-plugin dependency is bumped to a release containing this change:

nvidia-device-plugin:
  enabled: true
  injectFuseDevice: true

🤖 Generated with Claude Code