Legacy ML Modernization

A lot of production ML on AWS was built between 2017 and 2020, and a lot of it is still running on things that no longer exist. Apache MXNet and Gluon were retired to the Apache Attic in 2023. Chainer stopped development in 2019. SageMaker Edge Manager was discontinued in April 2024. Notebook instances still work, but the platform has moved to Studio spaces and the unified SageMaker platform. If that describes your estate, the risk is not that it stops working tomorrow; it is that it stops being patchable, hireable, and upgradeable, quietly.

We migrate the stacks others abandoned.

What we migrate

MXNet, Gluon, Chainer, and TensorFlow 1.x to PyTorch

We port model definitions layer by layer, reproduce the original training run to establish a numerical baseline, and validate the converted model against held-out data and against the old model's predictions before anything is switched over. Where a model is small enough, we sometimes retrain from scratch in PyTorch instead; the baseline tells us which is cheaper. Current Hugging Face and PyTorch deep learning containers replace the frozen MXNet containers, which have not received dependency or CVE updates in years.

Notebook instances to Studio spaces

Notebook-instance estates usually mean lifecycle scripts, local data on EBS volumes, and IAM roles that have grown by accretion. We move teams onto SageMaker Studio JupyterLab and Code Editor spaces (or SageMaker Unified Studio where the data platform justifies it), rebuild the environment as container images or lifecycle configurations, and put shared code into a repository instead of a notebook directory.

Edge Manager to ONNX Runtime and IoT Greengrass V2

With Edge Manager gone, AWS's own recommendation is ONNX for a cross-platform runtime and AWS IoT Greengrass V2 for deployment. We export models to ONNX, validate accuracy parity, package inference as Greengrass components, and replace Edge Manager's fleet monitoring with Greengrass telemetry and CloudWatch. Target hardware today is NVIDIA Jetson Orin, Raspberry Pi 5, and AMD or Arm industrial boards; if you are still on Jetson TX2, we plan the hardware refresh alongside the software migration.

Classic SageMaker to SageMaker AI and the unified platform

The December 2024 rename split "Amazon SageMaker" into SageMaker AI (the build / train / deploy service) and a new unified data-and-AI platform with Unified Studio, Lakehouse, and Catalog. For most teams the immediate change is small; for teams with a data lake and analytics users, the unified platform is an opportunity to consolidate tooling. We assess which applies to you and plan the adoption without disrupting running workloads.

How an engagement runs

1. Inventory (week 1). Every model, container, notebook, endpoint, and edge device, with its framework version, owner, and traffic. Most teams are surprised by what turns up.
2. Risk ranking. Frozen containers with known CVEs and models with no reproducible training run go first.
3. Migration in slices. One model or one team at a time, each with a numerical validation gate, shadow traffic where possible, and a rollback path.
4. Decommission. Old endpoints, notebook instances, and containers are actually turned off, and the savings are reported.

Why NeuralArmada

• Senior, US-based consultants who have worked on SageMaker since the MXNet days and know both sides of the migration.
• Validation-first: no model is switched over without a measured parity check.
• We leave behind modern, maintainable infrastructure as code, not another bespoke stack.

Inherited an ML estate nobody wants to touch? Contact us for a migration inventory.