When you use a locally developed model or training script to create a custom image, ensure that the custom image complies with the specifications defined by ModelArts.
Specifications
- A custom image cannot be larger than 30 GB. It is best if the custom image is no larger than 15 GB. An oversized image affects the startup of a training job.
- The uid of the default user of a custom image must be 1000.
- GPU drivers cannot be installed in an custom image. When you use GPUs to run a training job, ModelArts automatically places the GPU driver in /usr/local/nvidia of the training environment.
- ModelArts does not support the download of open source installation packages. Install the dependencies required by the training job in the custom image.
Creating a Custom Image for Training Jobs
- Set up the Docker environment on an ECS or a local host.
- Write a Dockerfile based on your needs and the preceding specifications to build a custom image.
- Upload the created custom image to SWR.