[shchur]

Summary

AutoGluon 1.6 returns to DLC as one unified SageMaker image instead of separate training and inference images.

• The same CPU or GPU image handles both train and serve.
• CPU and GPU builds share one Dockerfile, parameterized by the PyTorch base image and a matching requirements file.
• The images are based on the PyTorch 2.13 Amazon Linux 2023 SageMaker images.
• Training delegates to the inherited SageMaker training toolkit.
• Serving starts a lightweight Flask/Gunicorn adapter and loads the inference handler packaged with the model artifact.
• AutoGluon 1.6.2 is installed from PyPI using exact, version-specific CPU/GPU package deltas on top of the corresponding PyTorch DLC environment.
• Standard SageMaker configuration such as the program, port, worker count, timeout, and model requirements remains supported.

Multi-model endpoints are explicitly unsupported. AutoMM OCR is also unavailable because Amazon Linux 2023 does not provide Tesseract.

Security scan findings

The published PyTorch 2.13 base images still contain vulnerable cryptography, tornado, and CPU-only sqlparse. These are already fixed in DLC source by #6749, so AutoGluon temporarily overrides them until refreshed base images are published. MLflow is upgraded alongside cryptography to satisfy dependency constraints.

Two findings are allowlisted:

• Ray bundles vulnerable httpcore5 inside its JAR and cannot be patched independently. This matches the existing Ray handling (d6e815555).
• The Lightning finding is a scanner false positive: AutoGluon uses 2.6.6, whose release notes (https://github.com/Lightning-AI/pytorch-lightning/releases/tag/2.6.6) explicitly include the RCE fixes.

Validation

The change includes adapter unit tests, image sanity checks, and SageMaker train-to-serve tests for CPU and GPU images. Both image variants were also built locally with compatible package environments, and the CPU image successfully ran a Chronos-2 prediction job through AutoGluon Cloud on SageMaker.