sagemaker-lightgbm

Offline SageMaker local mode example using LightGBM.

Run it inside the repo's JupyterLab container, where SAGEMAKER_LOCAL_* environment variables point at the moto service.

Two images are used (both built from libs/sagemaker-local/docker assets):

  • sagemaker-lightgbm:train — training (Dockerfile); also serves online endpoints, so it carries the full serving stack plus sagemaker-training.
  • sagemaker-lightgbm:inference — leaner inference-only image (Dockerfile.inference), used by offline batch transform.

Build them with make build-images.

  • notebooks/training.ipynb — fit + deploy + predict on the california_housing regression dataset.
  • notebooks/batch_transform.ipynb — offline batch inference through the inference-only image on the first 20 rows of california_housing.
  • notebooks/pipeline.ipynb — a local TrainingStep pipeline on the iris classification dataset.

Tests: make test runs the fast wiring checks; make test-notebooks executes every notebook end to end (docker + moto + a Jupyter kernel, run inside the offline stack).