soperator

Soperator (Slurm-on-Kubernetes)

When To Use

Use when a customer wants a managed Slurm cluster on Nebius (foundation-model pretraining, large eval sweeps, HPC batch) instead of, or alongside, SkyPilot — and wants npa to drive it. npa soperator deploy wraps the public nebius/nebius-solutions-library soperator Terraform recipe from a compact declarative spec, so customers get a working Slurm cluster without hand-editing the recipe's large tfvars.

Three-tier contract:

• CLI: npa soperator plan|deploy --spec <cluster.yaml>, npa soperator status --name <n>, npa soperator destroy --name <n>.
• SDK: npa.sdk.soperator.plan(spec) / deploy(spec) / destroy(name) with SoperatorSpec / WorkerPoolSpec.
• YAML / agent: apiVersion: npa.soperator/v0.0.1 spec; workflow toolRef: infra.soperator.deploy.
• Fleet: npa.fleet/v0.0.1 targets may select backend: soperator and embed this same spec under soperator:. Standalone, SDK, and fleet call the same backend lifecycle; fleet remains responsible for identity, selection, concurrency, inventory, and shared-network ownership. Soperator names are fleet-wide physical/context identities and therefore must be unique even when targets belong to different projects.

Spec (npa.soperator/v0.0.1)

Multiple worker pools with different presets are first-class; each pool can enable a node-local Docker/Enroot image cache disk (NETWORK_SSD_IO_M3) so large GPU tool images don't thrash the boot disk.

apiVersion: npa.soperator/v0.0.1
name: npasop                 # company_name; kube context = nebius-<name>-slurm
region: us-central1          # or resolved from ~/.npa config
control_plane:
  system: { min_size: 3, max_size: 24 }         # preset omitted: derive XS..XL upstream
  controller: {}                                # preset omitted: derive with same tier
  login: { preset: 16vcpu-64gb }                 # login needs >= 16vcpu (sufficiency)
workers:
  - name: cpu8
    platform: cpu-d3
    preset: 8vcpu-32gb
    docker_cache: true          # node-local IO_M3 image cache
    docker_cache_gib: 930       # divisible by 93
  - name: gpu
    platform: gpu-b200-sxm
    preset: 8gpu-160vcpu-1792gb # GPU workers must be fabric-capable (8-GPU SXM)
    size: 2
    fabric: us-central1-b       # required for GPU presets; 1-GPU can't cluster
    preemptible: true           # on-demand GPU quota is often 0; preemptible works
    # For reserved capacity, set preemptible: false and exactly one runtime
    # selector. Never commit a live ID/name:
    # capacity_block_group: <capacity-block-group-id>
    # capacity_block_group_name: <unique-capacity-block-group-name>
    docker_cache: true
accounting: false
# omitted REST preserves the legacy default (follows accounting); GPU checks still run directly
slurm_operator_version: "4.1.6"
k8s_version: "1.34"
node_group_version: "72"