MXNet estate assessment
Inventory every model, pipeline, and serving path; map CVE and hardware exposure; get a port-vs-contain recommendation per model, with the evidence behind it.
Model porting
Conversion or re-implementation into PyTorch or ONNX, accepted only when a numerical-parity harness says the ported model is the same model.
Pipeline & serving migration
Gluon training code, SageMaker MXNet containers, and MXNet Model Server replaced with maintained equivalents, cut over after a parallel run.
Containment support
For what cannot move yet: pinned reproducible builds, CVE monitoring over a frozen tree, isolation, and an exit plan with dates on it.
Parity, not plausibility
Ported models are validated against the original at documented tolerances, on inputs that cover the distribution edges — not a smoke-test batch.
Fluent in both stacks
MXNet and Gluon idioms where the systems live today; PyTorch and the ONNX ecosystem where they are going. The judgement sits in the gap.
Honest about retraining
When weight conversion will not hold, we say so early and estimate the retraining rather than hiding it inside an optimistic conversion.
Parallel runs before cutover
New pipelines run beside the old on the same traffic — outputs, latency, and failure behavior compared — with rollback kept warm.