workflow._RepackModelStep can fail if source_dir contains requirements.txt

[akreuzer]

Describe the bug
_RepackModelStep is a part of the sagemaker.workflow.model_step.ModelStep.
Its purpose is to attach a source_dir to the plain model.tar.gz.
As I understand, this is done using a training step that does not do any real training but just repacks the model, see

sagemaker-python-sdk/src/sagemaker/workflow/_utils.py

Lines 155 to 158 in 284ddbe

	repacker = SKLearn(
	framework_version=FRAMEWORK_VERSION,
	instance_type=INSTANCE_TYPE,
	entry_point=REPACK_SCRIPT,

If the source_dir contains a requirements.txt, this step can fail. This is due to the fact in the step above the requirements will be installed even though they might not be compatible. See attached logs.

To reproduce
An example of a source_dir where this is the case is Jumpstart LightGBM Inference source_dir available at s3://jumpstart-cache-prod-us-east-2/source-directory-tarballs/lightgbm/inference/regression/v1.1.0/sourcedir.tar.gz
Its requirements.txt looks like this.

/opt/ml/model/code/lib/lightgbm/tenacity-8.0.1-py3-none-any.whl
/opt/ml/model/code/lib/lightgbm/plotly-5.1.0-py2.py3-none-any.whl
/opt/ml/model/code/lib/lightgbm/graphviz-0.17-py3-none-any.whl
...

Adding this source_dir to a model using the following code results in pipeline failure.

model = Model(
    image_uri=deploy_image_uri,
    source_dir=deploy_source_uri_cache,
    entry_point="inference.py",
    model_data=step_train.properties.ModelArtifacts.S3ModelArtifacts,
    sagemaker_session=pipeline_session,
    role=role,
)

step_model = ModelStep("RegisterModel", 
    step_args=model.register(
        content_types=["text/csv"],
        response_types=["application/json"],
        inference_instances=["ml.t2.medium"],
        transform_instances=["ml.m5.xlarge"],
        model_package_group_name="lgbm-test",
        
    ))

(Note that in the above code, deploy_source_uri_cache points to a copy of s3://jumpstart-cache-prod-us-east-2/source-directory-tarballs/lightgbm/inference/regression/v1.1.0/sourcedir.tar.gz as it is modified by _RepackModelStep)

Expected behavior
_RepackModelStep should work regardless of the content of source_dir.

Screenshots or logs

System information
A description of your system. Please provide:

• SageMaker Python SDK version: 2.103.0
• Framework name (eg. PyTorch) or algorithm (eg. KMeans): JumpStart LightGBM
• Framework version: N/A
• Python version: 3.8
• CPU or GPU: CPU
• Custom Docker image (Y/N): N

Additional context
Add any other context about the problem here.

[qidewenwhen]

Hi @akreuzer, thanks for using SageMaker!

This issue seems to be duplicate with #3143, which has been resolved.
Could you please upgrade your sagemaker SDK and see if it works for you?

[akreuzer]

Great 👍 Works