15.251 14.9.11 Ciclo de vida de modelos
MOD-01. Register source model with digest, license, training dataset and ontology versions. MOD-02. Export a portable interchange artifact such as ONNX when supported. MOD-03. Validate graph, preprocessing, outputs and numerical tolerance against the source implementation. MOD-04. Build optimized runtime artifact on or for the exact hardware/runtime profile. MOD-05. Record precision, dynamic shapes, calibration dataset and builder configuration. MOD-06. Benchmark accuracy before performance; reject engines outside tolerance. MOD-07. Sign model manifest and all referenced blobs. MOD-08. Stage in the edge model cache without activation. MOD-09. Run compatibility and resource preflight. MOD-10. Execute shadow or canary analysis on representative local attempts. MOD-11. Compare outputs, abstention, latency, memory and thermal behavior to the active model. MOD-12. Activate by immutable deployment profile revision. MOD-13. Retain previous-known-good model and engine until rollback window expires. MOD-14. Monitor drift proxies, quality-state changes and correction patterns. MOD-15. Reprocess historical attempts only through explicit jobs with a new result revision. MOD-16. Retire artifacts only after lineage and reproducibility checks.
TensorRT construye un engine serializado optimizado para un target y luego lo ejecuta mediante su runtime. Por ello, el engine se trata como artefacto específico del perfil, no como archivo universal; el manifiesto conserva runtime, hardware, shapes, precisión y builder config [P49-R12].
15.251.1 ModelManifest mínimo
- modelId, semanticVersion y immutable digest.
- trainingDatasetRelease, ontologyVersion y guidelineVersion.
- input/output schemas y unidades.
- preprocessing y postprocessing completos.
- runtime, opset, plugins y compatibility matrix.
- precision profile y calibration data reference.
- quality benchmark report y intended deployment profiles.
- license, SBOM/provenance, signature y signer identity.
- known limitations, excluded classes y fallback model.