SageMaker SDK v3 Lets Teams Change Model Code Without Rebuilding Container Images

SageMaker users can now revise local training or inference code and rerun their workflow without rebuilding the container image. The change comes from the SageMaker Python SDK v3’s SourceCode configuration, which syncs a local directory into a selected container when training or deployment begins.

AWS’s redesign separates application code from the runtime that executes it. Teams still bring an image from Amazon ECR, but it may be a user-built image, an AWS Deep Learning Container or a third-party image. That image can retain the system packages, framework libraries and CUDA dependencies a workload needs.

A source directory becomes part of the job

SourceCode takes a local source directory and either a command for training or an entry script for inference. At launch, SageMaker copies that directory into the container. The directory can include supporting modules, configuration files and shell scripts, allowing those components to change independently of the image.

Training and serving use different launch instructions

ModelTrainer configures and launches a training job using the chosen image and a SourceCode command. ModelBuilder uses the same directory-based pattern for inference, then packages the handler with the model artifact. It can target a managed SageMaker endpoint or run in a local container or the current Python process for testing.

SourceCode(
    source_dir="./train/random_forest",
    command="python random_forest.py",
)

The redesign does not eliminate the surrounding setup

AWS demonstrates the workflow with a scikit-learn Random Forest deployment and a Stable Diffusion 3.5 LoRA fine-tuning job using Hugging Face Accelerate for distributed multi-GPU training. The examples show the same API pattern spanning conventional machine learning and generative-model work.