feat(sagemaker): add support for serverless inference endpoints

[pymia]

Implements SageMaker Serverless Inference endpoints as requested in issue #23148.

• Add ServerlessProductionVariantProps interface with maxConcurrency, memorySizeInMB, and provisionedConcurrency
• Extend EndpointConfig to support serverless variants alongside existing instance variants
• Add comprehensive validation for serverless configuration parameters
• Enforce mutual exclusivity between instance and serverless variants
• Add CloudFormation template generation for ServerlessConfig properties
• Include extensive test coverage for validation scenarios and error cases

Issue # 23148

Closes #23148.

Reason for this change

AWS SageMaker Serverless Inference is not supported in the CDK SageMaker L2 constructs. Users can only configure instance-based endpoints, missing the serverless option for intermittent/unpredictable traffic patterns that could benefit from cost-effective serverless inference.

This feature was explicitly planned in the original SageMaker Endpoint L2 construct RFC with Instance-prefixed classes designed to make room for Serverless-prefixed analogs.

Description of changes

Implements AWS SageMaker Serverless Inference support in CDK SageMaker L2 constructs, enabling cost-effective serverless endpoints for intermittent workloads:

• New ServerlessProductionVariantProps interface extending ProductionVariantProps with AWS-compliant serverless properties:
  • maxConcurrency: 1-200 range (required)
  • memorySizeInMB: 1024-6144MB in 1GB increments (required)
  • provisionedConcurrency: 1-200 range, optional, must be ≤ maxConcurrency
• New addServerlessProductionVariant() method with comprehensive input validation
• Extended EndpointConfigProps with optional serverlessProductionVariant property
• Mutual exclusivity enforcement between instance and serverless variants per AWS constraints
• Single serverless variant limit per endpoint configuration (AWS limitation)
• Comprehensive synthesis-time validation with clear, actionable error messages
• CloudFormation integration leveraging existing L1 construct ServerlessConfig support

Usage Example:

import * as sagemaker from '@aws-cdk/aws-sagemaker-alpha';

declare const model: sagemaker.IModel;

// Create serverless endpoint configuration
const endpointConfig = new sagemaker.EndpointConfig(this, 'ServerlessEndpointConfig', {
  serverlessProductionVariant: {
    model: model,
    variantName: 'serverlessVariant',
    maxConcurrency: 10,
    memorySizeInMB: 2048,
    provisionedConcurrency: 5, // optional
  },
});

Describe any new or updated permissions being added

N/A - No new IAM permissions required. Leverages existing SageMaker model and endpoint permissions.

Description of how you validated changes

• Unit tests: Added 12 comprehensive serverless variant tests covering all validation scenarios:

  • Memory size validation (1024-6144MB in 1GB increments)
  • Concurrency range validation (1-200 for both max and provisioned)
  • Mutual exclusivity enforcement between instance and serverless variants
  • Single serverless variant limit per AWS constraints
  • Cross-environment model compatibility validation
  • Error condition testing with clear error messages
  • CloudFormation template generation verification

• Integration tests: Extended existing integration test with serverless endpoint configuration, verified CloudFormation template generation with correct ServerlessConfig properties:

  ServerlessEndpointConfig:
    Type: AWS::SageMaker::EndpointConfig
    Properties:
      ProductionVariants:
        - ServerlessConfig:
            MaxConcurrency: 10
            MemorySizeInMB: 2048
            ProvisionedConcurrency: 5
          VariantName: serverlessVariant

• Comprehensive testing results: 63/63 unit tests pass (100% success rate), 4/4 integration tests pass, no regressions detected across 16,024+ CDK tests

Checklist

• My code adheres to the CONTRIBUTING GUIDE and DESIGN GUIDELINES

---

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache-2.0 license

[aws-cdk-automation]

(This review is outdated)

[pahud]

taking a look.

[pahud]

❌ Features must contain a change to a README file.
❌ Features must contain a change to an integration test file and the resulting snapshot.

As this is a new feat we need

1. update README with very focusd and minimal description.
2. add new intet test or refresh existing relevant integ tests and update snapshots

✅ Updated pull request passes all PRLinter validations. Dismissing previous PRLinter review.

[abidhasan-aws]

Hi @pymia,
Thanks for your contribution. I have left some comments :)

[abidhasan-aws]

I wasn't able to find any documentation that says instanceProductVariant and serverlessProductVariant cannot be used simultaneously for a single endpoint. Could you please provide the source that refers to this restriction?

[pymia]

The instance based deployment and serverless deployment should not exist at the same time.
Reference: Amazon SageMaker Deploy Model, and AWS::SageMaker::EndpointConfig ProductionVariant.

[abidhasan-aws]

CDK has one README per service. SageMaker-alpha already has a README, so we don't need to create a new one.
We can add the documentation related to this PR in packages/@aws-cdk/aws-sagemaker-alpha/README.md.

[abidhasan-aws]

We can keep the integration test in one file. We have another integration test file integ.endpoint-config. We can put all the necessary integration-test related code in that file and remove this one.