controller, slurmd-cw-*) and extras images (controller-extras, slurmd-cw-*-extras) built on top of the base:
controller: Primarily for the control plane components, but you can also use it for CPU nodes.controller-extras: Extras for use on login nodes or CPU nodes.slurmd-cw-*: Images for GPU nodes.slurmd-cw-*-extras: Extras images for GPU nodes.
| Image | Based on | Image type |
|---|---|---|
controller (22.04) | ubuntu:jammy-20240911.1 | Base |
controller-extras (22.04) | controller (22.04) | Extra |
slurmd-cw-cu122 (22.04) | nccl-tests:12.2.2-devel-ubuntu22.04-nccl2.27.5-1-d5a135d | Base |
slurmd-cw-cu124 (22.04) | nccl-tests:12.4.1-devel-ubuntu22.04-nccl2.27.5-1-d5a135d | Base |
slurmd-cw-cu128 (22.04) | nccl-tests:12.8.1-devel-ubuntu22.04-nccl2.28.3-1-528fb68 | Base |
slurmd-cw-cu129 (22.04) | nccl-tests:12.9.1-devel-ubuntu22.04-nccl2.28.3-1-528fb68 | Base |
slurmd-cw-cu130 (22.04) | nccl-tests:13.0.1-devel-ubuntu22.04-nccl2.28.3-1-528fb68 | Base |
slurmd-cw-cu122-extras (22.04) | slurmd-cw-cu122 (22.04) | Extra |
slurmd-cw-cu124-extras (22.04) | slurmd-cw-cu124 (22.04) | Extra |
slurmd-cw-cu128-extras (22.04) | slurmd-cw-cu128 (22.04) | Extra |
slurmd-cw-cu129-extras (22.04) | slurmd-cw-cu129 (22.04) | Extra |
slurmd-cw-cu130-extras (22.04) | slurmd-cw-cu130 (22.04) | Extra |
slurm-cw-122(CUDA 12.2) is deprecated as of SUNK v6.5.0 and is scheduled for removal in the next major version release.slurm-cw-124(CUDA 12.4) is deprecated as of SUNK v6.9.0 and is scheduled for removal in the next major version release.
ubuntu-20.04 variants are no longer built after September 3, 2025 (CUDA 12.x nccl2.27.5-1-d5a135d are the last available versions).Other packages in the images that aren’t listed in the following sections might be dependencies for the listed packages and might change.
Select an image
This section explains where to pull SUNK images from and how theimage field resolves a registry and name.
| Image registry | Availability |
|---|---|
ghcr.io/coreweave/slurm-containers | Public |
images.coreweave.com/cluster-images/coreweave/slurm-containers | Internal to CKS |
image fields support a name field to specify the desired image name:
- If you don’t specify a repository, SUNK uses the default registry plus the image
name. - If you specify a repository, it overrides the value used. SUNK uses neither the
namenor the default registry.
controller, as shown in the following example, uses the controller image from the default registry, ghcr.io/coreweave/slurm-containers/controller.
External SUNK users
If you run SUNK outside CKS, you must point the Helm chart at the public registry instead of the internal CKS one. To do this, change theglobal.cks value in the slurm Helm chart as follows:
global.cks.enabled=false
The default value, global.cks.enabled=true, pulls images from the internal CKS registry, which isn’t available externally.
Base images
Base images provide the foundation for SUNK’s control plane and node images. This section describes their purpose and the packages they include.Purpose of base images
Base images serve as the foundation for the Slurm control plane components as well as the login and CPU-only slurmd pods.Main packages with pinned versions
- Slurm: Workload manager.
- s6-overlay: Process supervisor and service manager.
- Pyxis: Plugin for Slurm that enables containerized jobs.
- enroot: Container runtime.
- kubectl: Command-line tool for interacting with Kubernetes clusters.
NVSHMEM and GDRCopy supportIf you need an image with NVSHMEM and GDRCopy support, you can request to use ncore-image v2.10.1.
You must apply a patch to
ibgda in your container and contact Support to get access to this new image.
See the changelog entry for full details.
For implementation instructions, see NVSHMEM and GDRCopy Support.Main packages from repositories
- End-user utilities: htop, ping, traceroute, net-tools, numactl, sudo, wget
- OpenSSH server: Enables secure remote access to the login nodes.
- SSSD (System Security Services Daemon): Provides integration with external authentication providers using LDAP.
- Munge: Authentication service for creating and validating credentials.
- Environment modules: Enables dynamic modification of the user’s environment.
- libnvidia-container: NVIDIA Container Toolkit for GPU support.
Extra images
Extra images extend the base images with additional user-facing tools and utilities. Use them when you want a broader set of end-user tools on login or CPU nodes without building your own image. This section lists the packages added on top of the base images.Additional packages with pinned versions
- Conda: Package, dependency, and environment management for multiple languages.
- Micromamba: Lightweight alternative to Conda.