--- # Settings specific to workers with an attached NVIDIA GPU. # This group is a child of `workers` — it inherits workers.yml and only # overrides/adds what differs for a GPU node. # GPU nodes are not Longhorn storage nodes — override the workers.yml default # (which assumes /dev/sdb + /dev/sdc are present) so setup_longhorn.yml does # not try to format disks that don't exist in that layout. longhorn_disks: [] # Applied to the node object in Kubernetes by roles/gpu_device_plugin after # kubeadm join, so ollama/other GPU workloads can target it via nodeSelector. node_labels: gpu: nvidia # nvidia-container-toolkit package version (empty = latest) nvidia_container_toolkit_version: "" # Dedicated local RAID1 storage for Ollama model files (see roles/gpu_model_storage). # Carved out of otherwise-unpartitioned space on the same mirrored disks used # for the OS (sda/sdb have ~3.6TB free beyond the ~30GB OS RAID1 partition) — # model data must NOT go on the root filesystem, it's far too small. gpu_model_storage_devices: - /dev/sda - /dev/sdb gpu_model_storage_partition_size_gb: 1000