AWS adds SOCI image loading to SageMaker AI jobs

AWS announced the feature on October 27. SOCI indexes locate files inside a container image, the packaged environment for a job. Required files load first; the rest streams in the background. Benefits depend on what startup reads. Jobs that immediately touch most files may gain little. Training and processing require private Amazon ECR registries. Without a usable index, SageMaker uses the standard download.

The Deep Learning Containers documentation identifies ready indexed images by the -soci tag suffix. Its supported families include PyTorch Training, TensorFlow Training, vLLM and SGLang. DLC uses the second index format, which packages the original image and its index together. The documentation distinguishes tags tied to a specific image from rolling tags updated with releases. Its linked gallery lists supported versions, so a framework name alone does not identify the image to select.

For automated index creation, the AWS SOCI Index Builder documents an event-driven route. Pushing an image to ECR triggers an EventBridge rule. A filtering Lambda function checks repository and tag patterns; only a match invokes the generator. That function pulls the image, builds the configured index version and pushes the index artifacts back to ECR. Patterns support wildcards: dev:* selects images pushed to the dev repository. The builder supports both index formats and recommends V2.