Docker images

User

Any image from Docker Hub, or from a registry your tenant has configured, will run. There is no AiQu-specific base image and nothing to install in your image to make it work.

Choosing one

Focusing the Docker image field offers suggestions in two groups:

• Previous - images you have used before.
• Base - the images behind the Quick Deploy templates.

Picking a suggestion fills the field, applies the ports that image declares, and selects the registry it belongs to. That is why choosing a suggested Jupyter image gets you a working Jupyter port without thinking about it.

GPU images

A GPU image needs a GPU queue and at least one GPU requested. The common choices:

Image	For
nvidia/pytorch:24.07-py3	PyTorch, CUDA already set up
nvidia/tensorflow:24.07-tf2-py3	TensorFlow 2
nvidia/cuda:12.4.1-devel-ubuntu22.04	Build your own on top

Plain images such as ubuntu:22.04 work anywhere, including CPU-only queues.

Asking for a GPU you do not use still costs

Tokens are charged on resources held, not resources used. The Metrics view shows utilisation over time - if it sits near zero, request fewer GPUs next time.

Registries

Docker registry is set to Auto-detect by default, which is right for Docker Hub and for any image whose registry can be worked out from its name.

The selector lists the private and public registries configured for the tenant, grouped. Choose one explicitly when the image lives in a registry that needs credentials.

Tenant administrators add registries - see Tenant settings. Adding one is not something a user can do.

A private image with no registry selected fails immediately

The pull is attempted without credentials, is denied, and the job goes to failed within seconds of starting. The logs say pull access denied. Select the registry, or clone the job and set it.

Building your own

Why

If you install the same packages at the start of every job, you are paying tokens to run apt-get repeatedly. Bake them into an image once.

Dockerfile

FROM nvidia/pytorch:24.07-py3
RUN pip install --no-cache-dir transformers datasets accelerate
WORKDIR /workspace

Push and use

docker build -t registry.example.com/team/train:v1 .
docker push registry.example.com/team/train:v1

Then name that image in the job and select the matching registry.

Pin a tag. latest means a job you run today and the same job next month are not the same job, which is a bad property for an experiment.