Skip to content

[Draft] Release AutoGluon 1.6 - #6712

Open
AnirudhDagar wants to merge 1 commit into
aws:mainfrom
AnirudhDagar:feat/autogluon-1.6
Open

AnirudhDagar wants to merge 1 commit into
aws:mainfrom
AnirudhDagar:feat/autogluon-1.6

Conversation

@AnirudhDagar

Copy link
Copy Markdown
Contributor

Purpose

Add AutoGluon 1.6.1 training images (SageMaker, CPU and GPU/cu133) layered on the PyTorch 2.13 Amazon Linux 2023 DLC, replacing the Ubuntu 22.04 pytorch-training based AutoGluon 1.5 images.

  • docker/autogluon/training/Dockerfile: installs autogluon into the base venv; restores the base's fsspec afterwards (autogluon.multimodal caps it at 2025.3 while the base's s3fs needs 2026.x; AutoGluon 1.6.2 lifts the cap, after which the uv pip check exemption goes); removes Transformer Engine (unused by AutoGluon, and peft imports it at load time, which fails without an NVIDIA driver); re-templates telemetry and OSS compliance.
  • Image configs, pipeline / PR / dispatch-release / SageMaker test workflows following the huggingface-pytorch-training and tensorflow-training layout. metadata.framework stays autogluon, job_type carries the training distinction.
  • SageMaker SDK v3 tests that fit a TabularPredictor on CPU (ml.m5.2xlarge) and GPU (ml.g4dn.2xlarge) and assert CUDA visibility, a non-empty leaderboard, above-chance roc_auc and predictor reload.
  • Sanity training gate, test-suite registry entry, available-images data.

Known gaps: autogluon inference image for 1.6.

Test Plan

  • Build both images locally with the same --build-arg set the build-image action passes.
  • Run the repo sanity suite (test_sanity_training.py, filesystem, credentials, OSS compliance) inside each image.
  • uv pip check in each image.
  • Run the SageMaker entry script (train_tab.py) inside each image with SM_CHANNEL_* pointed at the test data; on the GPU image with SM_NUM_GPUS=1 on a g4dn.12xlarge.
  • MultiModalPredictor smoke fit on GPU.
  • PR workflow on this PR: build, sanity, security, telemetry, SageMaker tests.

Test Result

Local, on a dev desktop (CPU) and a g4dn.12xlarge (4x T4):

CPU GPU
Build ok (4.17 GB uncompressed) ok (19.49 GB uncompressed, ~10 GB compressed)
Sanity suite (44 tests), filesystem, credentials, OSS compliance pass pass
uv pip check only the known autogluon-multimodal fsspec line same
torch CUDA / matmul on GPU n/a 4 GPUs visible, ok
TabularPredictor fit + evaluate (roc_auc) 0.87, reloads 0.87, reloads
MultiModalPredictor fit on GPU n/a ok

cc @shchur


Toggle if you are merging into main Branch

PR Checklist

  • I ran pre-commit run --all-files locally before creating this PR. (Read DEVELOPMENT.md for details).

SageMaker CPU and GPU (cu133) images layered on the unified pytorch DLC,
replacing the Ubuntu 22.04 pytorch-training based AutoGluon 1.5 images.

- docker/autogluon/training/Dockerfile: installs autogluon into the base
  venv, restores the base's fsspec (autogluon.multimodal caps it at 2025.3,
  the base's s3fs needs 2026.x; the cap is being lifted in AutoGluon 1.6.2),
  removes Transformer Engine (unused, and peft imports it at load time which
  fails without an NVIDIA driver), re-templates telemetry and OSS compliance.
- Image configs, pipeline / PR / dispatch-release / SageMaker test workflows
  following the huggingface-pytorch-training and tensorflow-training layout.
- SageMaker SDK v3 tests fitting a TabularPredictor on CPU and GPU.
- Sanity gate, test-suite registry entry and available-images data.

Known gaps: no tesseract on AL2023 (AutoMM OCR unavailable), no inference image.

Signed-off-by: Anirudh Dagar <anirudhdagar6@gmail.com>
@AnirudhDagar
AnirudhDagar marked this pull request as ready for review September 4, 2026 15:22
@AnirudhDagar AnirudhDagar changed the title Release AutoGluon 1.6 [Draft] Release AutoGluon 1.6 Sep 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant