A model artifact is a persisted package of files that lets a serving system initialize and execute a trained model. In production, the predictor retrieves artifacts, loads them into memory, caches them, and uses them to answer inference requests. section 7.1.4 section 10.2.5
A deployable artifact commonly separates model execution data from descriptive metadata:
- model.pth stores the trained model state used when initializing the model graph. section 7.1.4
- vocab.pth supplies vocabulary data required to preprocess incoming text. section 7.1.4
- manifest.json records algorithm, framework, versions, model identity, code and model versions, and output-class mappings. section 7.1.4
- A model manager resolves the model identifier, loads the files, initializes the architecture, and caches the ready model. section 7.1.4
Artifact lookup during inference
A frontend sends a model identifier and document to a gRPC predictor. The predictor loads that model's artifact from shared storage, preprocesses the document, executes the cached model, and returns the prediction. In a larger deployment, a model service queries a metadata and artifacts store for the matching inference container and model file. section 7.1.4 section 10.2.5
├── manifest.json #1
├── model.pth #2
└── vocab.pth #3
// A sample manifest.json file
{
"Algorithm": "intent-classification", #4
"Framework": "Pytorch",
"FrameworkVersion": "1.9.0",
"ModelName": "intent",
"CodeVersion": "80bf0da",
"ModelVersion": "1.0",
"classes": { #5
"0": "cancel",
"1": "ingredients_list",
"2": "nutrition_info",
"3": "greeting",
.. .. ..
}
Inspect a concrete artifact directory and manifest, including framework versions, model identity, and class mappings. section 7.1.4
