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 github through github’s Container Registry (GHCR) and are also mirrored on CSCS’ internal jfrog artifactory for optimal performance when pulling images directly on Alps. The latter is only accessible from within the Alps environment, while the GHCR is publicly accessible and can be pulled from anywhere.

Base Image Jfrog URI GHCR URI
nvcr.io/nvidia/pytorch:26.02-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps6
nvcr.io/nvidia/pytorch:26.01-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.01-py3-alps6
nvcr.io/nvidia/pytorch:25.12-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:25.12-py3-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:12.12-py3-alps6
nvcr.io/nvidia/nemo:26.02 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:26.02-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:26.02-py3-alps6
nvcr.io/nvidia/nemo:25.11.01 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:25.11.01-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:25.11.01-py3-alps6
nvcr.io/nvidia/physicsnemo/physicsnemo:25.11 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-physicsnemo:25.11-alps6 ghcr.io/eth-cscs/alps-extended-images/ngc-physicsnemo:25.11-py3-alps6

Network Stack: libraries and versions

Library Version Notes
libfabric 2.6.0
NCCL 2.30.7-1
aws-ofi-plugin git-423a007 Built from commit 423a0073356529a93fd86191f20e27e6b7d84849
nvshmem 3.6.5-0
OpenMPI 5.0.10
The alps4 images are deprecated
Base Image Jfrog URI GHCR URI
nvcr.io/nvidia/pytorch:26.02-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps5
nvcr.io/nvidia/pytorch:26.01-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.01-py3-alps5
nvcr.io/nvidia/pytorch:25.12-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:25.12-py3-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:12.12-py3-alps5
nvcr.io/nvidia/nemo:26.02 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:26.02-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:26.02-py3-alps5
nvcr.io/nvidia/nemo:25.11.01 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:25.11.01-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:25.11.01-py3-alps5
nvcr.io/nvidia/physicsnemo/physicsnemo:25.11 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-physicsnemo:25.11-alps5 ghcr.io/eth-cscs/alps-extended-images/ngc-physicsnemo:25.11-py3-alps5

Network Stack: libraries and versions

Library Version Notes
libfabric 2.5.1
NCCL 2.30.7-1
aws-ofi-plugin git-423a007 Built from commit 423a0073356529a93fd86191f20e27e6b7d84849
nvshmem 3.6.5-0
OpenMPI 5.0.10
The alps4 images are deprecated
Base Image Jfrog URI GHCR URI
nvcr.io/nvidia/pytorch:26.02-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps4
nvcr.io/nvidia/pytorch:26.01-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.01-py3-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.01-py3-alps4
nvcr.io/nvidia/pytorch:25.12-py3 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:25.12-py3-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:12.12-py3-alps4
nvcr.io/nvidia/nemo:26.02 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:26.02-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:26.02-py3-alps4
nvcr.io/nvidia/nemo:25.11.01 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-nemo:25.11.01-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-nemo:25.11.01-py3-alps4
nvcr.io/nvidia/physicsnemo/physicsnemo:25.11 jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-physicsnemo:25.11-alps4 ghcr.io/eth-cscs/alps-extended-images/ngc-physicsnemo:25.11-py3-alps4

Network Stack: libraries and versions

Library Version Notes
libfabric 2.5.0a1 Built from commit 79fd4514977a8f15844fbe26df9f5930eb96311a + patch (removing dependency on shs-14 API, which is not available on Alps)
NCCL 2.29.7-1* 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 3.6.5-0
OpenMPI 5.0.10
The alps3 images are deprecated
Base Image Alps Extended Image URI
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 URI
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.

# (1)!
image = "jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps6"
mounts = [
    "/capstor/",
    "/iopsstor/",
]
writable = true
entrypoint = true # (2)!
[env]
PMIX_MCA_psec = "native" # (3)!
[annotations]
com.hooks.cxi.enabled = "false" # (4)!
  1. Images will be pulled directly from CSCS’ jfrog artifactory
  2. entrypoint = true is recommended for Alps Extended Images. The images initialize Alps-specific runtime settings through the NVIDIA container entrypoint, which sources the hooks under /opt/nvidia/entrypoint.d. Without this setting, runtimes that do not start containers through a login shell may not initialize the Alps environment.
  3. 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.
  4. The CXI hook must be disabled such that the container images network libraries have priority over the host system’s libraries.
# (1)!
image = "ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps6"
mounts = [
    "/capstor/",
    "/iopsstor/",
]
writable = true
entrypoint = true # (2)!
[env]
PMIX_MCA_psec = "native" # (3)!
[annotations]
com.hooks.cxi.enabled = "false" # (4)!
  1. Images will be pulled directly from CSCS’ jfrog artifactory
  2. entrypoint = true is recommended for Alps Extended Images. The images initialize Alps-specific runtime settings through the NVIDIA container entrypoint, which sources the hooks under /opt/nvidia/entrypoint.d. Without this setting, runtimes that do not start containers through a login shell may not initialize the Alps environment.
  3. 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.
  4. The CXI hook must be disabled such that the container images network libraries have priority over the host system’s libraries.

Danger

  • Set entrypoint = true in the EDF when using Alps Extended Images
  • 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)

Why is entrypoint = true recommended?

Alps Extended Images install runtime environment settings for the optimized networking stack inside the container image. These settings are initialized by the NVIDIA container entrypoint.

Current Enroot/Pyxis-based workflows may also initialize the environment through login-shell profile sourcing, but this is runtime-specific behavior and should not be relied on for portability. The future-proof and recommended method is to set entrypoint = true, to ensure that the image entrypoint is always run.

When launching your application with sbatch and srun, make sure to include the necessary flags for optimal performance and correct behavior, for example:

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

podman pull jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps6

podman pull ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps6

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

FROM jfrog.svc.cscs.ch/docker-group-csstaff/alps-images/ngc-pytorch:26.02-py3-alps6

FROM ghcr.io/eth-cscs/alps-extended-images/ngc-pytorch:26.02-py3-alps6

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:26.02-py3-alps6"

# 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-extended-images

Additional tests can be added to the build and test pipeline.

New images can be added to the Alps-Images folder.