ModelBuilder silently drops source_code when model data is a PipelineVariable #6450

PySDK Version

• PySDK V2 (2.x)
• PySDK V3 (3.x)

Describe the bug
ModelBuilder.build() silently drops source_code when s3_model_data_url is a pipeline variable.
This combination is required when a pipeline creates a model from the output of a training step:

• s3_model_data_url=<PipelineVariable>
• source_code=SourceCode(...)
• sagemaker_session=<PipelineSession>
  The generated CreateModel request contains the pipeline-variable model artifact, but its container environment is empty.
  ModelStep also contains no runtime repack step.
  As a result, script-mode inference containers cannot find the entry script.
  The same configuration works when s3_model_data_url is a concrete S3 URI.

To reproduce

1. Configure AWS credentials and a SageMaker execution role.
2. Set the role ARN:

export SAGEMAKER_ROLE_ARN="arn:aws:iam::<account-id>:role/<sagemaker-role>"
export AWS_REGION="us-east-1"

3. Save the following script as reproduce.py.
4. Run python reproduce.py.

import os
import tempfile
from pathlib import Path

import boto3

from sagemaker.core import image_uris
from sagemaker.core.training.configs import SourceCode
from sagemaker.core.workflow.parameters import ParameterString
from sagemaker.core.workflow.pipeline_context import PipelineSession
from sagemaker.mlops.workflow.model_step import ModelStep
from sagemaker.serve import ModelBuilder

region = os.environ.get("AWS_REGION", "us-east-1")
role_arn = os.environ["SAGEMAKER_ROLE_ARN"]

boto_session = boto3.Session(region_name=region)
pipeline_session = PipelineSession(boto_session=boto_session)

image_uri = image_uris.retrieve(
    framework="sklearn",
    region=region,
    version="1.4-2-py312",
    py_version="py3",
    instance_type="ml.m5.large",
    image_scope="inference",
)

model_data = ParameterString(
    name="ModelDataUrl",
    default_value="s3://example-bucket/model.tar.gz",
)

with tempfile.TemporaryDirectory() as source_dir:
    entry_script = Path(source_dir) / "inference.py"
    entry_script.write_text(
        """
def model_fn(model_dir):
    return None

def input_fn(request_body, content_type):
    return request_body

def predict_fn(data, model):
    return data

def output_fn(prediction, accept):
    return prediction
""".strip()
    )

    builder = ModelBuilder(
        image_uri=image_uri,
        s3_model_data_url=model_data,
        source_code=SourceCode(
            source_dir=source_dir,
            entry_script="inference.py",
        ),
        role_arn=role_arn,
        instance_type="ml.m5.large",
        sagemaker_session=pipeline_session,
    )

    model_step = ModelStep(
        name="CreateModel",
        step_args=builder.build(model_name="pipeline-source-code-repro"),
    )

    primary_container = model_step.arguments["PrimaryContainer"]
    environment = primary_container.get("Environment", {})

    print("PrimaryContainer environment:", environment)
    print("Runtime repack step count:", len(model_step.steps))

    assert environment == {}, environment
    assert len(model_step.steps) == 0

    print("Bug reproduced: source_code was silently dropped.")

Actual output:

PrimaryContainer environment: {}
Runtime repack step count: 0
Bug reproduced: source_code was silently dropped.

No SageMaker model or pipeline execution is created by this reproduction.

Expected behavior
ModelBuilder must preserve source_code when s3_model_data_url is a pipeline variable.
The generated ModelStep must add a runtime repack step that:

1. Reads the model artifact after the training step completes.
2. Adds inference.py under code/.
3. Creates a self-contained model.tar.gz.
4. Sets SAGEMAKER_PROGRAM.
5. Sets SAGEMAKER_SUBMIT_DIRECTORY to /opt/ml/model/code.

If pipeline-variable repacking is not supported, ModelBuilder.build() must raise a clear error. It must not silently remove the source code.

Screenshots or logs
The generated primary container has no script-mode environment:

PrimaryContainer environment: {}
Runtime repack step count: 0

The resulting CreateModel request contains the model-data pipeline expression but no inference-code configuration.
System information

• SageMaker Python SDK version: sagemaker==3.24.0
• Component versions:
  • sagemaker-core==2.24.0
  • sagemaker-serve==1.24.0
  • sagemaker-mlops==1.24.0
• Framework name (eg. PyTorch) or algorithm (eg. KMeans): Scikit-learn
• Framework version: 1.4-2-py312
• Python version: 3.12
• CPU or GPU: CPU
• Custom Docker image (Y/N): N
• Operating system: macOS, Darwin 25.6.0
  Additional context
  The concrete-S3 version of this problem was fixed by: fix(serve): repack source_code for image_uri/ModelTrainer builds #6112
  The v3.24 passthrough implementation recognizes a model artifact only when it is a string that starts with s3://. A PipelineVariable does not satisfy that check.
  The passthrough branch then clears:

self.source_dir = None
self.entry_point = None
self.dependencies = None
self.git_config = None

The SDK already contains pipeline-variable runtime-repack logic in _upload_code(), and ModelStep supports _RepackModelStep.
That logic is not reached from this passthrough path.
The existing integration test covers a concrete S3 model URI, but it does not cover a model-data pipeline variable: https://github.com/aws/sagemaker-python-sdk/blob/v3.24.0/sagemaker-serve/tests/integ/test_passthrough_source_code_repack_integration.py

Current workaround:

1. Upload the inference source as a separate S3 archive.
2. Pass SAGEMAKER_PROGRAM and SAGEMAKER_SUBMIT_DIRECTORY through env_vars.
   This workaround is acceptable for a transient batch-transform model. It does not produce the self-contained artifact required for
   later model-package registration.