ModelBuilder.deploy() should expose DataCacheConfig and other CreateInferenceComponent API parameters

[manuwaik]

Describe the feature you'd like

Expose additional CreateInferenceComponent API parameters through ModelBuilder.deploy() when deploying Inference Components. Currently, ModelBuilder._deploy_core_endpoint() builds a minimal InferenceComponentSpecification and hardcodes several values. The following API-supported configuration is not surfaced:

• Specification.DataCacheConfig.EnableCaching — Cache model artifacts and container images on instances for faster auto-scaling cold starts
• Specification.BaseInferenceComponentName — Adapter component deployment (e.g. LoRA adapters attached to a base model)
• Specification.Container (Image, ArtifactUrl, Environment) — Custom container images, artifact URLs, and environment variables at the IC level
• VariantName (top-level) — Currently hardcoded to "AllTraffic", not configurable for multi-variant endpoints

How would this feature be used? Please describe.

Our immediate need is DataCacheConfig.EnableCaching. When deploying large model artifacts (700MB+ Triton ensembles) on auto-scaling endpoints, new instances must re-download all artifacts during scale-out. Enabling caching eliminates this overhead.

Ideally these would be optional parameters on deploy() or on ResourceRequirements:

builder.deploy(
    endpoint_name="my-endpoint",
    inference_component_name="my-ic",
    instance_type="ml.g5.2xlarge",
    initial_instance_count=1,
    inference_config=ResourceRequirements(
        requests={"memory": 8192, "num_accelerators": 1, "num_cpus": 2, "copies": 1}
    ),
    data_cache_config={"enable_caching": True},  # new
)

Describe alternatives you've considered

The current workaround is to use ModelBuilder.build() for model creation, then call EndpointConfig.create(), Endpoint.create(), and InferenceComponent.create() directly via sagemaker-core, which supports the full InferenceComponentSpecification:

from sagemaker.core.resources import EndpointConfig, Endpoint, InferenceComponent
from sagemaker.core.shapes import (
    ProductionVariant,
    InferenceComponentSpecification,
    InferenceComponentComputeResourceRequirements,
    InferenceComponentRuntimeConfig,
    InferenceComponentDataCacheConfig,
)

builder.build(model_name=sm_model_name)

EndpointConfig.create(
    endpoint_config_name=epc_name,
    production_variants=[
        ProductionVariant(
            variant_name="AllTraffic",
            instance_type="ml.g5.2xlarge",
            initial_instance_count=1,
        )
    ],
    execution_role_arn=role,
)

Endpoint.create(
    endpoint_name=endpoint_name,
    endpoint_config_name=epc_name,
)

InferenceComponent.create(
    inference_component_name=ic_name,
    endpoint_name=endpoint_name,
    variant_name="AllTraffic",
    specification=InferenceComponentSpecification(
        model_name=sm_model_name,
        compute_resource_requirements=InferenceComponentComputeResourceRequirements(
            min_memory_required_in_mb=8192,
            number_of_accelerator_devices_required=1,
            number_of_cpu_cores_required=2,
        ),
        data_cache_config=InferenceComponentDataCacheConfig(enable_caching=True),
    ),
    runtime_config=InferenceComponentRuntimeConfig(copy_count=1),
)

This works but defeats the purpose of ModelBuilder.deploy() as a high-level abstraction. Customers shouldn't need to drop down to sagemaker-core or boto3 for commonly used API parameters.

Additional context

• API reference: CreateInferenceComponent
• The shape classes (InferenceComponentDataCacheConfig, InferenceComponentContainerSpecification, etc.) already exist in sagemaker.core.shapes — they just need to be wired into ModelBuilder._deploy_core_endpoint().
• SageMaker SDK version: 3.x