Lindorm Ray custom image - build push and configure - Lindorm

Prerequisites

Ensure that you meet the following prerequisites:

You have activated Lindorm Compute Engine (Activate Service).
You have a Container Registry (ACR) instance in the same region as your Lindorm instance (Create an Enterprise Edition instance).

Usage notes

Follow these rules to ensure stability and compatibility:

In your Dockerfile, do not modify any files in the following directories or their subdirectories:
- /tmp/ray/
- /opt/
- The Ray installation directory. For example, with Python 3.11, the path is /home/ray/anaconda3/lib/python3.11/site-packages/ray/.
Push the custom image to an ACR repository in the same region as your Lindorm instance. Lindorm cannot pull images from other regions.

Core workflow

Using a custom image involves four main steps:

Prepare and build: Write a Dockerfile based on an official Lindorm Ray public image and build your custom image.
Push the image: Push the built image to your Alibaba Cloud ACR repository.
Authorize access: Grant Lindorm one-time access to your ACR repository.
Configure and use: When creating a Ray resource group, specify your custom image in the Advanced Configurations section.

Step 1: Prepare and build a custom image

Get the Lindorm Ray public image address

Lindorm provides public images with pre-configured Ray environments. Use them as the base for your custom image.

Image pull address format:

docker pull lindorm-compute-public-registry.cn-hangzhou.cr.aliyuncs.com/lindorm-ray/ray:[image-tag]

Available tags:

Image tag	Python version	GPU support
`2.50.1-1.1.1-py310-cpu`	Python 3.10	No
`2.50.1-1.1.1-py311-cpu`	Python 3.11	No
`2.50.1-1.1.1-py310-gpu`	Python 3.10	Yes (CUDA pre-installed)
`2.50.1-1.1.1-py311-gpu`	Python 3.11	Yes (CUDA pre-installed)

Dependency version guidance

Select compatible dependency versions from the official Ray dependency list. Use pip install -c requirements_compiled.txt to apply version constraints.

Write a Dockerfile

Create a Dockerfile that extends a Lindorm Ray public image with your dependencies.

Example Dockerfile:

FROM lindorm-compute-public-registry.cn-hangzhou.cr.aliyuncs.com/lindorm-ray/ray:2.50.1-1.1.1-py311-cpu

RUN pip install --no-cache-dir numpy pandas

Important

Do not modify files in the /tmp/ray/, /opt/, or Ray installation directories, or their subdirectories. For example, in a Lindorm Ray image with Python 3.11, the Ray installation directory is /home/ray/anaconda3/lib/python3.11/site-packages/ray/.

Build the image

Run the following command in the directory containing the Dockerfile to build the image.

docker build -t your-image-name:tag .

Step 2: Push the image to ACR

Push the built image to your ACR repository.

Log in to ACR

docker login --username=<your-username> <your-registry>

Tag and push the image

docker tag your-image-name:tag <your-registry>/<namespace>/<repo>:tag
docker push <your-registry>/<namespace>/<repo>:tag

Step 3: Authorize Lindorm to access ACR

On the Ray resource group creation page, authorize Lindorm to access your ACR repository.

Click "Password-free Access to ACR Custom Images".
In the dialog box, confirm the authorization.

Note

This authorization is one-time. Subsequent resource groups reuse it.

Step 4: Use the custom image

When creating a Ray resource group, add the following JSON in the Advanced Configurations section to specify your custom image.

{ 
"clusterConfig.image.custom.acr.instanceId":"your-acr-instanceId",
"clusterConfig.image":"your-full-image-addr"
}

Configuration parameters:

Parameter	Description	Example
`clusterConfig.image.custom.acr.instanceId`	Your ACR instance ID.	`"cri-123456789"`
`clusterConfig.image`	The full image address, in the format `<registry>/<namespace>/<repo>:<tag>`.	`"your-registry-vpc.cn-hangzhou.cr.aliyuncs.com/my-namespace/my-ray-app:1.0"`

You can now create a Ray resource group with your custom image.