Vertex AI provides Docker container images that you run asprebuiltcontainers for serverless training. These containers,which are organized bymachine learning (ML) framework and framework version, include commondependencies that you might want to use in training code. Often, using aprebuilt container is simpler thancreating your own custom container fortraining.

This document lists the prebuilt containers for training, and it describes howto use them with aPython training application.

Vertex AI supports each framework version based on a schedule tominimize security vulnerabilities. Review theVertex AI framework supportpolicy to understand the implications ofthe end-of-support and end-of-availability dates.

Available container images

Each of the following container images is available in severalArtifact Registry repositories, whichstore data in variouslocations. You can use any ofthe URIs for an image when you perform serverless training; each provides the samecontainer image. When using Google Cloud console to perform serverless training,the Google Cloud console selects the URI that best matches thelocation whereyou're using Vertex AI. This reduceslatency.

Ray on Vertex AI container images are available.The following versions of prebuilt containers are supported:

Train with a prebuilt container

To use a prebuilt container, read theguide to configuring containersettings.

If you're using a container image that supports GPUs, make sure to specify themachineSpec.acceleratorType andmachineSpec.acceleratorCount fields when you createyourCustomJob,HyperparameterTuningJob, orTrainingPipeline resource.

What's next

• Learn how tocreate a Python trainingapplication to use with aprebuilt container.