AutoGluon Container¶

Every SageMaker job and endpoint launched by AutoGluon-Cloud runs in the official AutoGluon Deep Learning Container (DLC). The image is built and maintained by AWS, security-patched on an Amazon Linux 2023 base, and includes AutoGluon together with all the libraries it builds on.

You can also use the image directly — with your own training scripts and inference handlers — when you need more control than AutoGluon-Cloud provides. This section documents the image for both cases.

Custom Scripts

Train and serve your own AutoGluon scripts on SageMaker with the image.

custom-scripts.html

Reference

Inference handler contract, environment variables, and known limitations.

reference.html

Images¶

A single image runs both SageMaker training jobs and inference endpoints. AutoGluon 1.5 and earlier shipped separate autogluon-training and autogluon-inference images.

Variant	Image
GPU	`763104351884.dkr.ecr.<region>.amazonaws.com/autogluon:1.6-cu133-amzn2023`
CPU	`763104351884.dkr.ecr.<region>.amazonaws.com/autogluon:1.6-cpu-amzn2023`

The 1.6 tags always point to the latest 1.6.x patch release. To pin a patch release, use the full version (for example 1.6.3-cu133-amzn2023). For the account IDs in other regions, see Available Images in the DLC documentation.

What’s included¶

AutoGluon 1.6.3
Tabular classification and regression: scikit-learn, LightGBM, CatBoost, XGBoost, TabM
Tabular foundation models: Mitra, TabICL, TabDPT, Nori
Time series forecasting: StatsForecast (statistical models such as ETS and ARIMA), GluonTS (deep learning models such as DeepAR, TFT, and PatchTST), MLForecast
Time series foundation models: Chronos, Toto 2.0
Inference server implementing the SageMaker /ping and /invocations contract. Your handler defines model_fn and transform_fn — see Reference.

The image is built on the PyTorch DLC: PyTorch 2.13, Python 3.12, CUDA 13.3 (GPU variant), Amazon Linux 2023.

Foundation model weights are not baked into the image. They are downloaded from Hugging Face the first time a model is used, so the training job or endpoint needs internet access, or the weights must be bundled into the model artifact (see cache_model_artifact()).

How AutoGluon-Cloud uses the image¶

AutoGluon-Cloud picks the image for you based on the framework_version argument of fit(), predict() and deploy() (default "1.6"), the AWS region, and whether the instance type has a GPU. It then supplies its own training and inference scripts — the same kind of scripts you would write yourself following Custom Scripts.

To run AutoGluon-Cloud on a different image — for example, one you extended with extra packages — pass custom_image_uri:

from autogluon.cloud import TabularCloudPredictor

cloud_predictor = TabularCloudPredictor()
cloud_predictor.fit(
    train_data="train.csv",
    predictor_init_args={"label": "class"},
    custom_image_uri="<account_id>.dkr.ecr.<region>.amazonaws.com/my-autogluon:latest",
)

Custom images should be built FROM the AutoGluon image, so that the scripts AutoGluon-Cloud ships keep working.