Amazon SageMaker AI now supports SOCI-indexed images to start training jobs in seconds

Amazon SageMaker AI has introduced support for Seekable OCI (SOCI)-indexed container images, enabling training and processing jobs to start faster by reducing container pull times. Previously, SageMaker AI required downloading entire container images, often exceeding 10 GB, which added minutes to job startup and incurred additional costs. With SOCI indexing, only the necessary files are fetched initially, while the rest streams in the background, minimizing idle GPU time.

To implement this feature, users can reference a SOCI-indexed image from a private Amazon ECR repository in their training or processing jobs. Pre-indexed AWS Deep Learning Containers are available by using the -soci image tag variant, or users can add a SOCI index to their own images in Amazon ECR using open-source tools without rebuilding the image. The process requires no changes to job definitions, entry points, or data channels.

SageMaker AI automatically checks for a SOCI index and falls back to standard downloads if none is found, ensuring compatibility with existing workflows. If an index is unavailable, the job continues without interruption, maintaining operational continuity. This feature is designed to streamline job initialization while preserving performance and reliability.

The SOCI-indexed images feature is available today in all AWS Regions where SageMaker AI is supported, accessible through SageMaker AI APIs, CLIs, SDKs, and the Console. Users can refer to the documentation for training jobs and processing jobs or consult the launch blog for additional details on implementation and benefits.