Choosing a Docker Image

Every job on the Slices AI infrastructure runs inside a Docker container. Choosing the right base image saves setup time and avoids hard-to-debug CUDA compatibility problems. This page describes the three main sources of images, ranked by how well they are tested on Slices hardware.

Option 1 — GPU Docker Stacks

The Slices AI infrastructure team maintains a set of images specifically built and tested for the available clusters.

Key advantages:

  • Same environment as JupyterHub. The Docker Stacks images are the same images available on the JupyterHub spawn page. This means you can prototype interactively in JupyterHub and then run the exact same environment as a batch job — no surprises when you switch.

  • Rebuilt weekly with the latest stable releases of PyTorch, TensorFlow, and CUDA.

  • Include Jupyter, common scientific Python packages, and correctly configured CUDA/cuDNN stacks.

Available images:

Framework

Image

CUDA

PyTorch

gitlab.ilabt.imec.be:4567/ilabt/gpu-docker-stacks/pytorch-notebook:cuda12-latest

12 (latest)

PyTorch

gitlab.ilabt.imec.be:4567/ilabt/gpu-docker-stacks/pytorch-notebook:cuda12.6-latest

12.6 (compatible with older GPUs)

PyTorch

gitlab.ilabt.imec.be:4567/ilabt/gpu-docker-stacks/pytorch-notebook:cuda13-latest

13

TensorFlow

gitlab.ilabt.imec.be:4567/ilabt/gpu-docker-stacks/tensorflow-notebook:cuda-latest

Latest stable

The -latest tag always resolves to the most recent build. For a reproducible, pinned environment use a versioned tag, e.g. cuda12-pytorch-2.4.0.

Full archive of all available tags: GPU Docker Stacks — Container Registry.

Note

CUDA 12.x images run on most clusters. Two exceptions:

  • GTX 1080 Ti (Pascal, CC 6.1) — requires PyTorch ≤ 2.7 (CUDA 12 support was removed in 2.8). Use images tagged with cuda12.6.

  • RTX 5090 (Blackwell, CC 12.0) — requires PyTorch ≥ 2.8.

See GPU software compatibility for the full compatibility matrix.

Option 2 — Official framework images

PyTorch and TensorFlow publish their own CUDA-enabled Docker images. These are a good choice if you need a specific framework version not yet available in the GPU Docker Stacks, or if you want a minimal image without the Jupyter stack.

  • PyTorch — official images on Docker Hub: docker.io/pytorch/pytorch. Tags follow the pattern <pytorch_version>-cuda<cuda_version>-cudnn<cudnn_version>-runtime (e.g. 2.11.0-cuda12.8-cudnn9-runtime).

  • TensorFlow — official images on Docker Hub: docker.io/tensorflow/tensorflow. Use GPU-enabled tags ending in -gpu (e.g. 2.21.0-gpu).

Caution

Framework images are not tested on Slices hardware. You are responsible for verifying CUDA/cuDNN compatibility with the GPU you request. Consult GPU software compatibility before using a framework image on an older or newer GPU.

Option 3 — NVIDIA NGC containers

NVIDIA NGC is NVIDIA’s catalog of professionally maintained, performance-optimised containers. NGC images are particularly useful for:

  • Distributed / multi-GPU training — NGC containers include optimised builds of NCCL, cuDNN, and NVIDIA DALI that are often faster than the versions bundled in framework images.

  • Cutting-edge framework versions — NGC typically ships new PyTorch and TensorFlow versions within days of release.

  • Large language models — NGC hosts containers for NeMo, TensorRT-LLM, and Triton Inference Server.

To use an NGC container on the Slices AI infrastructure, use the nvcr.io registry prefix:

"image": "nvcr.io/nvidia/pytorch:24.08-py3"

The tag format is YY.MM-py3 (year and month of the release). Browse available tags on the NGC PyTorch page or the NGC TensorFlow page.

NGC containers are large (often 10–20 GB). The first job using a new tag will take longer to start while the image is pulled.

Caution

NGC containers are large and target recent GPU architectures. Always verify CUDA and compute capability compatibility against GPU software compatibility before using an NGC image on an older GPU.

Summary: which image to choose

Situation

Recommended image source

Getting started / standard PyTorch or TensorFlow workflow

GPU Docker Stacks — tested, matches JupyterHub

Need a specific framework version not in GPU Docker Stacks

Official framework image (Docker Hub)

Multi-GPU / distributed training; need optimised NCCL or cuDNN

NVIDIA NGC

None of the above fit; custom dependencies or private packages

Build a custom image based on one of the above — see Job Definition reference