Skip to content

Alps Extended Images

The Alps infrastructure (specifically the networking stack) requires custom-built libraries and specific environment settings to fully leverage the high-speed network. To reduce the burden on users and ensure best-in-class performance, we provide pre-built Alps Extended Images based on popular container images, starting with those commonly used by the ML/AI community.

Info

See our communication library guide for detailed information about how to build containers with optimised support for the Slingshot network used by Alps.

Note

All extended images are thoroughly tested and validated to ensure correct behavior and optimal performance (see contributing section).

Images

The images are hosted on the CSCS internal artifactory repository and can only be pulled from within the Alps environment.

Repository URL

jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/<image:tag>
Base Image Alps Extended Image URL
nvcr.io/nvidia/pytorch:26.01-py3 ngc-pytorch:26.01-py3-alps3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps3
nvcr.io/nvidia/pytorch:25.12-py3 ngc-pytorch:25.12-py3-alps3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:25.12-py3-alps3
nvcr.io/nvidia/nemo:25.11.01 ngc-nemo:25.11.01-alps3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:25.11.01-alps3

Network Stack: libraries and versions

Library Version Notes
libfabric 2.5.0a1 Built from commit 102872c0280ce290d9d663945dad8a36ceb53c50 + patch (removing dependency on shs-14 API, which is not available on Alps)
NCCL 2.29.3-1* Patched by applying https://github.com/NVIDIA/nccl/pull/1979 to the 2.29.3 release
aws-ofi-plugin git-394ae7b* Built from commit 394ae7b20dd0e6b4e5f63652e15e9da100d5fe83 + patched by applying https://github.com/aws/aws-ofi-nccl/pull/1056
nvshmem 3.4.5-0
OpenMPI 5.0.9
The alps2 images are deprecated
Base Image Alps Extended Image URL
nvcr.io/nvidia/pytorch:26.01-py3 ngc-pytorch:26.01-py3-alps2 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps2
nvcr.io/nvidia/pytorch:25.12-py3 ngc-pytorch:25.12-py3-alps2 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:25.12-py3-alps2
nvcr.io/nvidia/nemo:25.11.01 ngc-nemo:25.11.01-alps2 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:25.11.01-alps2

Network Stack: libraries and versions

Library Version Notes
libfabric 2.5.0a1 Built from commit f8262817c337d615a1acceea6cd4ecb526ce548b + patched by applying https://github.com/ofiwg/libfabric/pull/11684
NCCL 2.29.2-1* Patched by applying https://github.com/NVIDIA/nccl/pull/1979 to the 2.29.2 release
aws-ofi-plugin git-eb9877e* Built from commit eb9877e9cfecf725dba0794a5e0fc06f8fdf7f3f + patched by applying https://github.com/aws/aws-ofi-nccl/pull/1056
nvshmem 3.4.5-0
OpenMPI 5.0.9
The alps1 images are deprecated
Base Image Alps Extended Image
nvcr.io/nvidia/pytorch:26.01-py3 ngc-pytorch:26.01-py3-alps1
nvcr.io/nvidia/pytorch:25.12-py3 ngc-pytorch:25.12-py3-alps1
nvcr.io/nvidia/nemo:25.11.01 ngc-nemo:25.11.01-alps1

Tip

Images are continuously updated to incorporate the latest improvements. We strongly recommend periodically checking whether a newer version of an Alps Extended Image is available.

Usage

To use an image directly on Alps via an EDF environment file, set the image to the repository URL followed by the image name and tag.

Danger

  • Do not use the aws_ofi_nccl hook annotation
  • Explicitly disable the cxi hook
  • Use the --environment flag for srun instead of sbatch (i.e. srun --environment=my_edf.toml ...)
  • Use the --network=disable_rdzv_get flag for srun to disable the rendezvous mechanism for network initialization (i.e. srun --network=disable_rdzv_get ... or setting SLURM_NETWORK=disable_rdzv_get)
  • Launch MPI applications with PMIx (i.e. srun --mpi=pmix or setting SLURM_MPI_TYPE=pmix)
Example EDF file
# (1)!
image = "jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps3"
mounts = [
    "/capstor/",
    "/iopsstor/",
]
writable = true
[env]
PMIX_MCA_psec = "native" # (2)!
[annotations]
com.hooks.cxi.enabled = "false" # (3)!
  1. Images will be pulled directly from CSCS’ jfrog artifactory
  2. Pertinent environment variables for optimal network performance are already set in the container image. PMIX_MCA_psec = "native" is recommended here in order to avoid warnings at initialization.
  3. The CXI hook must be disabled such that the container images network libraries have priority over the host system’s libraries.
Example sbatch file
#!/usr/bin/env bash
#SBATCH --account=my_account
#SBATCH --job-name=example
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=4

srun \
    --mpi=pmix \ # (1)!
    --network=disable_rdzv_get \ #(2)!
    --environment=example.edf.toml \ #(3)!
    python my_script.py
  1. The --mpi=pmix flag is required to ensure that PMIx is used as the MPI launcher - without this flag you may encounter errors during initialization.
  2. The --network=disable_rdzv_get flag is required to disable the rendezvous mechanism for network initialization. Alternatively, you can also set the environment variable SLURM_NETWORK=disable_rdzv_get to achieve the same effect.
  3. The --environment must be used as a flag for srun - passing this flag to sbatch will lead to errors related to missing Slurm plugins.

srun errors related to missing Slurm plugins

If you are submitting a batch job with sbatch and using the --environment (i.e. #SBATCH --environment=my_edf.toml) option, this can lead to errors such as: 

srun: error: plugin_load_from_file: dlopen(/usr/lib64/slurm/switch_hpe_slingshot.so): libjson-c.so.3: cannot open shared object file: No such file or directory
srun: error: Couldn't load specified plugin name for switch/hpe_slingshot: Dlopen of plugin file failed
srun: fatal: Can't find plugin for switch/hpe_slingshot
Make sure you use the option for srun instead of sbatch, i.e.: 
srun --environment=my_edf.toml ...

PMIx/ucx errors during initialization

If you see warnings related to PMIx, for example 

No PMIx server was reachable, but a PMI1/2 was detected
or observe ucx logs indicating that 
Transport endpoint is not connected
this likely indicates that Slurm is not configured to use PMIx for launching MPI applications. To resolve this, ensure that you are launching your application with the --mpi=pmix flag, for example: 
srun --mpi=pmix ...
or set the environment variable SLURM_MPI_TYPE=pmix to make PMIx the default MPI launcher.

Pulling and using images with Podman

Extended images can also be pulled using Podman

Pulling with Podman
podman pull docker://jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps3

and/or be used as base images in your own Containerfiles:

Example Containerfile
FROM jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps3

RUN echo "Hello world!"
For further information, please see the guide to building container images on Alps.

Inspecting image provenance labels

Alps Extended Images include OCI labels with provenance metadata (for example, source repository, commit SHA, and build time). You can inspect these labels with podman.

Pull image and inspect labels
IMAGE="jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps3"

# Pull the image
podman pull "$IMAGE"

# Show all labels (JSON)
podman image inspect "$IMAGE" --format '{{ json .Labels }}'

# Show specific labels
podman image inspect "$IMAGE" --format 'Source Repository: {{ index .Labels "org.opencontainers.image.source" }}'
podman image inspect "$IMAGE" --format 'Source Revision: {{ index .Labels "org.opencontainers.image.revision" }}'
podman image inspect "$IMAGE" --format 'Build Time: {{ index .Labels "org.opencontainers.image.created" }}'

Contributing

The Alps extended images are automatically built via a dedicated CI/CD pipeline hosted on GitHub:

github.com/eth-cscs/alps-swiss-ai

Additional tests can be added to the build and test pipeline

New images can be added to the Alps-Images folder

Note

The repository is currently private. Please open a Service Desk ticket to request access.