Skip to content

AUTOML_DEPLOYMENT.md

Latest commit

 

History

History
207 lines (136 loc) · 13 KB

AUTOML_DEPLOYMENT.md

File metadata and controls

207 lines (136 loc) · 13 KB

AutoML Experiment, Deployment, and MCP Server

This guide walks through: (1) creating an AutoML experiment on Red Hat OpenShift AI, (2) deploying the best model, and (3) using that deployment with this MCP server so tools (e.g. invoke_churn) call your live model.

It is based on the Red Hat AI examples – Predict Customer Churn tutorial. You can follow the same pattern for other datasets (e.g. credit risk) by swapping the pipeline inputs and schema.

Overview

Phase What you do
1. AutoML experiment Create a project, S3 connections, run the AutoML pipeline (e.g. Telco Churn), view the leaderboard.
2. Deploy best model Register the chosen model, set up the AutoGluon ServingRuntime (KServe), deploy the model, get the inference URL and token.
3. Use with MCP server Set DEPLOYMENT_URL and DEPLOYMENT_TOKEN in .env; run the MCP server; tools will POST to your deployment.

Prerequisites

  • Access to Red Hat OpenShift AI (self-managed or cloud).
  • Permissions to create projects, connections, pipelines, workbenches, and deployments.
  • This repo with the MCP AutoML server (see main README for Python/setup).

Phase 1: Create and run the AutoML experiment

Follow the churn prediction tutorial up to and including View the leaderboard. Summary:

1.1 Create a project

  • In OpenShift AI, go to Projects and create a new project (e.g. customer-churn-ml or credit-risk-ml).

1.2 Create S3 connections

Create two S3-compatible connections in the project:

  1. Results storage – for pipeline artifacts and leaderboard (e.g. automl-results-s3). You will use this when configuring the Pipeline Server.
  2. Training data – for the dataset (e.g. customer-churn-data-s3). Note the connection name; you will use it as train_data_secret_name in the pipeline run.

See: Create the S3 connections.

1.3 Configure the Pipeline Server

  • In your project, open Pipelines (or project details) and Configure pipeline server.
  • Set the Object storage connection to the same bucket/credentials as your results S3 connection so runs and artifacts (leaderboard, models) are stored there.
  • Choose Default database or External MySQL as needed. Create/Save and wait until the Pipeline Server is ready.

See: Configure the Pipeline Server.

1.4 Create a workbench with connections

  • In Workbenches, create a workbench and attach both the results and training-data S3 connections so you can access artifacts and data without a restart.

See: Create workbench with connections attached.

1.5 Upload the dataset to S3

  • Download the dataset (e.g. WA_FnUseC_TelcoCustomerChurn.csv for churn).
  • Upload it to the bucket used by the training data connection. Note the bucket name and object key (path) for the pipeline run.

See: Upload the training dataset to S3.

1.6 Add the AutoML pipeline and run it

  • Get the compiled AutoML pipeline from the repo: autogluon_tabular_training_pipeline (branch rhoai_automl), e.g. pipeline.yaml.
  • In OpenShift AI Pipelines, add it as a new Pipeline Definition (upload/create from YAML).
  • Create a pipeline run with at least:
    • train_data_secret_name – name of the training-data S3 connection
    • train_data_bucket_name – bucket name
    • train_data_file_key – object key of the CSV (e.g. data/WA_FnUseC_TelcoCustomerChurn.csv)
    • label_column – e.g. Churn
    • task_type – e.g. binary
    • top_n – e.g. 3
  • Start the run and wait for completion.

See: Add the AutoML pipeline, Run AutoML with the required inputs.

1.7 View the leaderboard

  • Open the run Artifacts and locate the leaderboard (e.g. HTML). Download or open it and pick the best model (e.g. top-ranked) to deploy.

See: View the leaderboard.

Phase 2: Deploy the best model

Follow the tutorial from Model Registry through Deployment Scoring.

2.1 Register the model (optional but recommended)

  • Create a Model registry (one-time, if not already done) under Settings → Model resources and operations → AI registry settings.
  • In Registry → Model registry, Register model:
    • Model location: Object storage (S3), using the same artifact store as your pipeline.
    • Path: root folder of one refitted predictor (e.g. under .../autogluon-models-full-refit/<task_id>/model_artifact/<ModelName>_FULL/).
    • Set Model name, Version, and Source model format (e.g. custom / AutoGluon), then Register.

See: Model Registry.

2.2 Prepare the AutoGluon ServingRuntime (KServe)

  • Build the serving image on the cluster (ImageStream + BuildConfig) using the tutorial’s YAML (Git source, Dockerfile, output to ImageStream).
  • Create the ServingRuntime from the tutorial’s ServingRuntime YAML:
    • Set metadata.namespace to your project.
    • Set spec.containers[0].image to the built image (e.g. image-registry.openshift-image-registry.svc:5000/<namespace>/autogluonkserveimagev1:latest).
  • In OpenShift AI: Settings → Serving runtimes → Add serving runtime → upload the YAML, select REST and Predictive model, then Create.

See: Prepare the ServingRuntime for AutoGluon with KServe.

2.3 Deploy the model

  • Projects → your project → Deployments → Deploy model.
  • Model location: S3; use the path to the refitted model (same as registry or from run artifacts).
  • Model type: Predictive model.
  • Model framework: e.g. autogluon - 1.
  • Serving runtime: AutoGluon ServingRuntime for KServe.
  • In Advanced settings:
    • Require token authentication – enable if you want to use a Bearer token (recommended for the MCP server).
    • Make model deployment available through an external route – enable so you can call the endpoint from your machine (for the MCP server).
  • Deploy model and wait until the deployment is running.

See: Model Deployment.

2.4 Get the inference URL and token

  • Open the deployment details. Under Inference endpoint, copy the external URL (only if you enabled the external route).
  • Deployment URL for the MCP server is the predict endpoint, for example:
    • <EXTERNAL_BASE_URL>/v1/models/<MODEL_NAME>:predict
      where <MODEL_NAME> is the deployment’s Resource name (lowercase, no spaces). Example: https://my-model-myproject.apps.example.com/v1/models/my-churn-model:predict
  • Token (if you enabled token auth): Projects → your project → Deployments → expand the deployment → use the Token secret value as DEPLOYMENT_TOKEN.

See: Deployment Scoring for the exact request format (e.g. instances with per-field arrays). The MCP server sends payloads in that same shape.

Phase 3: Use the deployment with the MCP server

3.1 Match the tool schema to the deployment

The MCP server sends a JSON body like:

{
  "instances": [
    { "feature1": [value1], "feature2": [value2], ... }
  ]
}

Your deployment (e.g. AutoGluon churn) expects one object per instance with the same feature names as in training. This repo provides churn_schema.json and the tool invoke_churn in tools_config.yaml. The schema properties match the churn model’s inputs (e.g. gender, tenure, Contract, Churn, etc. as in the tutorial’s curl example).

Set DEPLOYMENT_URL (and DEPLOYMENT_TOKEN if used) to your churn predict endpoint. To add another tool, add a new JSON Schema and entry in tools_config.yaml with its own schema_path and the same deployment_url_env / deployment_token_env if needed.

3.2 Set environment variables

In the MCP server directory, ensure .env exists and contains the deployment values from Phase 2:

mv template.env .env

Edit .env:

  • DEPLOYMENT_URL = full predict URL, e.g. https://my-model-myproject.apps.example.com/v1/models/my-churn-model:predict
  • DEPLOYMENT_TOKEN = token from the deployment’s Token secret (leave empty or omit if you did not enable token auth)

Keep LLAMA_STACK_CLIENT_* (or other LLM vars) if you use the demo client; see main README.

3.3 Run the MCP server and call the tool

  1. Start the server:

    python mcp_automl/mcp_server.py

  2. Use the deployment via MCP:

    • Demo client: run python mcp_automl/interact_with_mcp.py and ask a question that triggers the tool (e.g. churn).
    • Cursor: add the MCP server with URL http://127.0.0.1:8000/sse (see README – Attaching to Cursor and Ollama).
    • Ollama: use an MCP-capable client with the same URL, or use the demo client with Ollama as the LLM.

The tool will POST to DEPLOYMENT_URL with Authorization: Bearer <DEPLOYMENT_TOKEN> and the validated input as instances; the response (e.g. predictions) is returned to the caller.

Quick reference

Step Where What to set / do
AutoML run OpenShift AI Pipelines Pipeline definition, run params (train_data_*, label_column, task_type, top_n)
Leaderboard Run Artifacts Pick best model for deployment
Deploy model Project → Deployments S3 path, AutoGluon runtime, external route, token auth
Inference URL Deployment details <base>/v1/models/<resource-name>:predict
Token Deployment → Token secret DEPLOYMENT_TOKEN in .env
MCP server .env DEPLOYMENT_URL, DEPLOYMENT_TOKEN
Tool schema tools_config.yaml + churn_schema.json Match deployment input features

References