MXNet Estate Assessment
Know what you have before you move it.
You can't plan an exit from a framework you haven't mapped. Most teams that adopted MXNet did it years ago — through GluonCV, GluonNLP, or the early SageMaker MXNet containers — and the people who built those systems have often moved on. The assessment rebuilds the picture.
What we inventory
- Models. Every MXNet model in production or on a retraining schedule: architecture, Gluon vs symbolic API, custom operators, weight formats, and who (if anyone) still understands it.
- Training pipelines. Gluon training code, data loaders, augmentation stacks, hyperparameter tooling, and the infrastructure they run on — including SageMaker MXNet estimator jobs.
- Serving paths. Inference services, batch scoring jobs, MXNet Model Server / multi-model endpoints, and edge deployments, with their throughput and latency envelopes.
- Dependency and CVE exposure. The full transitive tree: MXNet's own frozen dependencies, the NumPy/SciPy versions it pins you to, and the CVEs that will never be patched upstream.
- Hardware and driver constraints. Which CUDA, cuDNN, and driver versions each workload requires, and where those conflict with what your platform team wants to run fleet-wide.
What you get
A written assessment with a port-vs-contain recommendation per model:
- Port — the model converts cleanly or is worth re-implementing; we estimate the porting and validation effort, including whether weight conversion will hold or retraining is required.
- Contain — the model can't move yet; we specify the pinning, isolation, and monitoring posture it needs, and a date by which the decision gets revisited.
- Retire — the model no longer earns its keep; the honest recommendation is sometimes to switch it off.
Each recommendation comes with the evidence behind it: parity-risk factors (custom operators, non-standard layers, numerical edge cases), dependency exposure, and the operational cost of leaving it in place.
Why start here
Porting effort varies wildly between models that look similar from the outside. A stock ResNet with standard Gluon blocks converts in days; a model with custom symbolic operators or hybridized control flow can take an order of magnitude longer, and sometimes the right answer is retraining against the original data. The assessment tells you which situation you're in per model — before you commit budget to the wrong plan.
The assessment is a fixed-scope piece of hourly work, sized in the initial conversation based on how many models and pipelines you're running. Start the conversation.