seldondev

Add support for /v2/models/{model name}/config endpoint in Executor

12

This is part of Triton extensions https://github.com/triton-inference-server/server/blob/main/docs/protocol/extension_model_configuration.md

In the current design of SC we have the assumption that a model has one `ModelUri` however in the case of white box explainers we can have upto 2 artifacts...

Config-driven detectors based onhttps://docs.google.com/presentation/d/1pMgmF5xmMhc_t6SCN2jnGKWc1f7UV3AccOv63rui0TY/edit?usp=sharing

Discussion points fromGraduation of Alibi-detect runtime in MLServer

The are inference cases e.g. tf models trained with mlflow (`TypeError: Only dict and DataFrame input types are supported`) that dont consume directlynp.ndarray In these cases the alibi runtime will...

- [ ] Add example on serving explainers through Alibi Explain runtime - [ ] Add docs around extra config required to serve explainers.

So far we have tested in mlserver Anchors and IntegratedGradients explainers. We should at some point start checking if they other (e.g. SHAP and counterfactual) techniques work well in the...