Start an edge model deployment plan by confirming model sources and permitted uses, input and output definitions, the target device and software versions. Preserve the original model, representative samples, label definitions and existing results. Begin with a limited deployment trial and agree on acceptance criteria for function, task quality and resource limits.
Hand over conversion files with their configuration
Record conversion commands, tool versions, operator support and known limitations. For quantization, retain calibration data coverage and preprocessing. Calculate and save the output model file checksum. Link that file to the original model and conversion logs. A successful conversion message does not establish accepted inference on the target device.
Compare outputs on the same samples
Check input tensors first, then compare outputs from the original and deployed models. Check classification results, bounding boxes or other actual task outputs using agreed tolerances and quality metrics. Save inconsistent samples and explanations. List insufficient samples and conditions that were not tested as unverified.
Measure workload and check recovery
Under the agreed concurrent workload, record timing boundaries, samples, resources and exceptions. Check startup, model loading failure, process restart and version rollback separately. Deliver models, configurations, dependencies, comparison results and open issues. Add verification when increasing concurrency or operating duration later.
For conversion and evaluation scope, see ONNX Runtime quantization documentation.