XGBoost Server

If you have a trained XGBoost model saved you can deploy it simply using Seldon's prepackaged XGBoost server.

Prerequisites

Seldon expects that your model has been saved as model.bst, using XGBoost's bst.save_model() method. Note that this is the recommended approach to serialise models.

To maximise compatibility between the serialised model and the serving runtime, it's recommended to use the same toolkit versions at both training and inference time. The expected dependency versions in the latest XGBoost pre-packaged server are as follows:

Package	Version
xgboost	1.4.2

Usage

To use the pre-packaged XGBoost server, it's enough to declare XGBOOST_SERVER as the implementation for your model. For example, for a saved Iris model, you could consider the following config:

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: xgboost
spec:
  name: iris
  predictors:
  - graph:
      children: []
      implementation: XGBOOST_SERVER
      modelUri: gs://seldon-models/xgboost/iris
      name: classifier
    name: default
    replicas: 1

You can try out a worked notebook with a similar example.

Open Inference Protocol (or V2 protocol)

The XGBoost server can also be used to expose an API compatible with the Open Inference Protocol. Note that, under the hood, it will use the Seldon MLServer runtime.

In order to enable support for the Open Inference Protocol, it's enough to specify the protocol of the SeldonDeployment to use v2. For example,

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: xgboost
spec:
  name: iris
  protocol: v2 # Activate the Open Inference Protocol
  predictors:
  - graph:
      children: []
      implementation: XGBOOST_SERVER
      modelUri: gs://seldon-models/xgboost/iris
      name: classifier
    name: default
    replicas: 1

You can try a similar example in this worked notebook.