After an inference backend change, earlier results can serve as comparison material. Function, task quality and performance conclusions for the new environment need corresponding evidence. Record changes to model files, conversion configuration, inputs, outputs and software dependencies before defining the retest scope.

What remains after successful conversion?

Confirm actual model execution on the target device. Check input tensors, output fields, exceptions and fallback. Use representative samples to compare classification, detection or other task outputs, retaining differences and tolerances. Conditions that were not tested cannot be reported as verified compatibility.

How should performance be compared again?

Keep the input, batch size, warm-up, timing boundaries and concurrent workload unchanged throughout each comparison. Record model execution and application latency separately. Compare sample distributions and resource use. State changes in numerical precision, threading or data copying location. Comparing two displayed times alone is insufficient.

What records belong in the handover?

Retain both model versions and checksums, environment configurations, samples, output comparisons, performance records and open issues. Confirm startup, failure handling, restart and rollback paths. Conclusions about simultaneous tasks, long-term operation or site performance need evidence covering the actual conditions.

See ONNX Runtime graph optimization documentation.