Deploy & inference

ModelBuilder

Central V3 abstraction for creating a deployable model configuration.

Inputs

trained model/resource/package + serving schema/image

Outputs

SageMaker Model resource and deployable Endpoint

Core explanation

I think ModelTrainer on the training side and ModelBuilder on the deployment side.

Mental model

What this means before worrying about syntax.

ModelBuilder can take a framework model, inference specification or registered model package, build serving artifacts/configuration and deploy it.

ModelBuilder is the V3 serving-side abstraction that turns a model/inference specification into deployable serving resources.
Training artifact and inference contract meet here.
Request/response schema and preprocessing are part of production design.

Lifecycle position

trained or registered model -> ModelBuilder.build() -> deployable model -> deploy() -> endpoint

Implementation anatomy

The sequence you would actually build or configure.

1
Identify model source.
2
Define schema/serialization as needed.
3
Select serving image and compute.
4
Build serving configuration.
5
Deploy.
6
Smoke-test representative requests and capture endpoint/version identity.

Objects / terms worth remembering

sagemaker.serve.ModelBuilderSchemaBuilderbuild()deploy()

Concrete example

A real model-development or production flow.

Input JSON contains complaint text

↓

Inference loads vectorizer + classifier

↓

Transforms text then predict_proba

↓

Response returns category, probability, model_version

↓

Application applies 0.75 routing threshold

Implementation lab

Concrete code, data manipulation and configuration for this lifecycle step.

Build and deploy from a trained object/resource

python

Current V3 ModelBuilder returns core Model/Endpoint resources rather than the old Predictor abstraction.

from sagemaker.serve.model_builder import ModelBuilder

builder = ModelBuilder(
    model=trained_model_or_model_package,
    role_arn=SAGEMAKER_ROLE_ARN,
    sagemaker_session=session,
    instance_type="ml.m5.xlarge",
)

model_resource = builder.build(
    model_name="complaint-classifier-v17"
)

endpoint = builder.deploy(
    endpoint_name="complaint-classifier-prod",
    initial_instance_count=2,
)

Invoke the returned Endpoint resource

python

payload = {
    "text": "I do not recognize this debit card purchase"
}

result = endpoint.invoke(data=payload)
print(result)

Design decisions that matter

Serving mode and compute.
Explicit request/output contract.
Dependency packaging.
Where feature retrieval/business rules live.

Common failure modes

Local load success does not guarantee serving-image dependency compatibility.
Opaque feature retrieval inside model code creates coupling.
A model package still needs a valid serving contract.

Production readiness check

Items to close before treating this step as production-ready.

✓ model source✓ serving image✓ schema✓ dependencies✓ instance type✓ smoke test✓ version traceability