gcp-gke-gpu-node-pool

Nullstone module that adds a GPU node pool with time-slicing to an existing GKE cluster.

Designed for serving multiple GPU workloads (e.g. vLLM models) per physical GPU: with the defaults, each g4-standard-48 node (1x NVIDIA RTX PRO 6000, 96GB VRAM) advertises nvidia.com/gpu: 3 allocatable via GKE time-slicing, so 3 pods each requesting 1 GPU slot schedule onto one node.

Pair with the gcp-gke-gpu-cores capability on a gcp-gke-service app to request GPU slots, target this pool, and tolerate the GPU taint.

Requirements

• The cluster must run GKE 1.34+ (set min_master_version = "1.34" on the gcp-gke block). The apply fails with a precondition error otherwise.
• GKE automatically taints GPU nodes with nvidia.com/gpu=present:NoSchedule. System workloads (kube-dns, external-secrets) cannot schedule here — the cluster must keep at least one untainted node pool.
• G4 machines only support Hyperdisk boot disks (disk_type defaults to hyperdisk-balanced).
• GPU drivers are auto-installed by GKE (gpu_driver_version, default LATEST).

Connections

• cluster — cluster/gcp/k8s:gke
  • The GKE cluster to attach the node pool to. Reuses the cluster's node service account when available (gcp-gke 0.7.0+).

Inputs

• num_node_zones: number — default 2
  • The number of zones to allocate GPU nodes. Zones are pulled from the zones available in the cluster's region. NOTE: GPU machine families are not available in every zone; check with gcloud compute accelerator-types list --filter="name=<accelerator_type>".
• machine_type: string — default g4-standard-48
• accelerator_type: string — default nvidia-rtx-pro-6000
• accelerator_count: number — default 1
• max_shared_clients_per_gpu: number — default 3
  • Pods per physical GPU via time-slicing. Set 1 to dedicate GPUs.
• gpu_driver_version: string — default LATEST
• min_node_count / max_node_count: number — defaults 1 / 3 (per zone)
• spot: bool — default false
• disk_type: string — default hyperdisk-balanced
• disk_size: number — default 100

VRAM budgeting with time-slicing

Time-slicing shares GPU compute, not memory — every pod sees the full GPU and there is no memory isolation. Budget VRAM explicitly: the sum of VRAM fractions of pods sharing a GPU should stay at or below 0.90. For example, 3 vLLM models on one 96GB GPU each set --gpu-memory-utilization 0.30.

Outputs

• node_pool_name: string — the name of the GPU node pool.
• node_selector: map(string) — node selector labels targeting this pool (cloud.google.com/gke-nodepool: <pool name>).
• gpu_sharing_strategy: string — TIME_SHARING, or empty when dedicated.
• max_shared_clients_per_gpu: number — pods per physical GPU.
• accelerator_type: string — the GPU accelerator type.
• accelerator_count: number — physical GPUs per node.