Docs

Volcengine Kubernetes Engine

Best practices

Using AI container images

AI container images are a series of Docker images provided by Volcano Engine, which are pre-installed with common AI frameworks, tools, and related dependencies. They offer a quick-start and highly consistent AI development environment. This topic describes how to use AI container images.

Background information

In the preliminary stages of AI model training and inference, you typically work with specific combinations of AI frameworks and tools (for example, the PyTorch framework and DeepSpeed library). After performing algorithm tuning, model injection, or script development in a basic environment, you need to submit repackaged images to the job queue of the training or inference platform.

AI container images accelerate the setup of AI development, training, and inference environments. Together with other tools and product capabilities in the AI Suite, they provide services covering the entire lifecycle of AI applications.

Prerequisites

A basic environment for AI training or inference has already been set up with VKE resources such as clusters, VPCs, and GPU nodes.

Limit

VKE AI container images can only be pulled within a VPC and cannot be pulled from the public network. You need to pull images from the VKE cluster or ECS instance in the corresponding region of Volcano Engine.

The following document mainly introduces the practices officially tested and verified. In order to achieve the expected results and comply with the Volcengine Kubernetes Engine (VKE) Limits, you must strictly follow the procedures (or the recommended resources) in this document. If you need to modify the procedures to suit your service requirements, contact the customer manager for consultation.

Procedure

Step 1: Obtain the address of the desired AI container image

Volcano Engine provides a series of pre-built AI container images, enabling rapid deployment and consistent AI development environments. The following table contains the supported versions, support status, and image addresses of AI container images.

tip

You need to replace {regionID} in an image address with the actual region ID. For regions where AI container images are available and their region IDs, see Regions and availability zones.

Pay attention to the following information about pre-built AI container images:

The software and dependencies in pre-built AI container images are limited to versions officially compiled and made generally available by NVIDIA.

Pre-built AI container images are updated quarterly, following the update cycle of official software versions from NVIDIA. The updated images will use the latest generally available software versions released by NVIDIA.

Machine learning framework version	Other software and dependency version	Operating system version	Still supported	Image address
-	CUDA 11.8.0 Python 3.10	Ubuntu 22.04	Yes	ai-containers-{regionID}.cr.volces.com/deeplearning/cuda11.8.0-py3.10-ubuntu22.04:latest
Pytorch 2.0.1	CUDA 11.8.0 Python 3.10	Ubuntu 22.04	Yes	ai-containers-{regionID}.cr.volces.com/deeplearning/torch2.0.1-cuda11.8.0-py3.10-ubuntu22.04:latest
Tensorflow 2.13.0	CUDA 11.8.0 Python 3.10	Ubuntu 22.04	Yes	ai-containers-{regionID}.cr.volces.com/deeplearning/tensorflow2.13.0-cuda11.8.0-py3.10-ubuntu22.04:latest
Pytorch 2.0.1	Megatron 23.05 CUDA 11.8.0 Python 3.10	Ubuntu 22.04	Yes	ai-containers-{regionID}.cr.volces.com/deeplearning/megatron23.05-torch2.0.1-cuda11.8.0-py3.10-ubuntu22.04:latest
Pytorch 2.0.1	Deepspeed 0.10.0 CUDA 11.8.0 Python 3.10	Ubuntu 22.04	Yes	ai-containers-{regionID}.cr.volces.com/deeplearning/deepspeed0.10.0-torch2.0.1-cuda11.8.0-py3.10-ubuntu22.04:latest

Step 2: Pull the AI container image to start a container

Obtain the AI container image, pull it into VKE, and use it to start a container. This AI container image is used to create a basic training environment.

In the left navigation bar, click Cluster. On the page that appears, click the name of the target cluster that is available for AI training or inference. The management page of the cluster opens.

In the left navigation bar, click Workload > Pod to enter the pod management page.

Click Create with Yaml to create a pod that uses a pulled pre-built AI container image. The following sample code shows the YAML content.

Step 3: Prepare the training environment

After the pod enters the Running status, log in to the pod console via VKE to further prepare the environment as needed. For example, you can use development tools to develop algorithms and scripts, or import model files.

Step 4: Package the image and push it to CR

Use the AI container image as the base image to build a custom AI development environment via a Dockerfile. The following Dockerfile provides an example:

Push the derived image to the Volcano Engine CR service. For details, see Pulling images from and pushing images to a standard CR instance.

What to do next

At this point, the derived image packaged based on the AI container image has been pushed to Volcano Engine CR. You can deploy model training or inference jobs in the VKE cluster.

Last updated: 2026.07.24 15:50:47