aws-k8s-nvidia variant fails to boot on g7 instance family: ghostdog selects tesla driver but GPU (10de:2c3a) requires open kernel modules

Image I'm using

bottlerocket-aws-k8s-1.36-nvidia-x86_64-v1.62.1 (ami resolved via Karpenter alias bottlerocket@latest, us-east-1)

What I expected to happen

Nodes from the g7 instance family (Blackwell-generation GPU, PCI ID 10de:2c3a) boot, load a working NVIDIA driver, and join the EKS cluster.

What actually happened

Boot fails before kubelet ever starts, so the node never joins the cluster (Karpenter NodeClaim stuck at Registered: Unknown — Node not registered with cluster, then relaunch loop). From the EC2 console log:

ghostdog[2336]: Error: grid is not preferred driver: tesla
ghostdog[2337]: Error: open-gpu is not preferred driver: tesla
...
nvidia-migmanager[2755]: '/usr/bin/nvidia-smi' failed - stderr:
[FAILED] Failed to start NVIDIA MIG manager service.
[DEPEND] Dependency failed for Generate CDI specifications.
[DEPEND] Dependency failed for Bottlerocket initial configuration complete.
[DEPEND] Dependency failed for Activate configured.target.

ghostdog selects the proprietary tesla driver for this GPU. The g7 family's GPU requires the NVIDIA open kernel modules — the proprietary driver rejects it. (Verified on AL2023 with the same 580-series driver: proprietary logs NVRM: The NVIDIA GPU ... (PCI ID: 10de:2c3a) requires use of the NVIDIA open kernel modules and fails RmInitAdapter, while the open kernel module at 580.159.03 drives the GPU fine — nvidia-smi shows "NVIDIA RTX PRO 4500 Blackwell Server Edition", CUDA 13.0.)

Because nvidia-smi fails, nvidia-migmanager fails, and the dependency chain through configured.target collapses — the host never reaches the point of starting kubelet.

How to reproduce

Launch any g7-family instance with the aws-k8s-*-nvidia variant (reproduced on multiple g7 sizes/launches; the failure is deterministic).

Suggested fix

Teach the driver-selection logic (ghostdog / driver preference) that device ID 10de:2c3a (and other g7-family GPU IDs) must use the open-gpu driver, not tesla.

Impact

Bottlerocket nvidia variants are currently unusable on the g7 instance family, with no settings/userdata escape hatch (admin/control host-containers can't intervene before configured.target fails). We had to move our GPU node group to AL2023 as a workaround.