This is my code:

from sagemaker.sklearn import SKLearnModel


role = sagemaker.get_execution_role()

model = SKLearnModel(
    model_data= f"s3://{default_bucket}/{prefix}/model.tar.gz",  
    role=role,
    entry_point="inference.py",    
    framework_version="1.2-1",      
    py_version="py3",
)

predictor = model.deploy(
    instance_type="ml.c5.large",
    initial_instance_count=1,
    container_startup_health_check_timeout=180
)

s3://{default_bucket}/{prefix}/model.tar.gz contains:

Contents of the tarball: (I also tried putting requirements.txt in code/ as advised on documentation for pytorch models on tarball structure)
?rw-r--r-- sagemaker-user/users    4349839 2024-11-29 19:22:21 model.pkl 
?rw-r--r-- sagemaker-user/users         24 2024-12-02 14:43:26 inference.py 
?rw-r--r-- sagemaker-user/users      44212 2024-11-29 19:23:17 explainer 
?rw-r--r-- sagemaker-user/users         24 2024-12-02 14:43:26 requirements.txt 

requirements.txt contains:

dill
pandas
joblib

(and I even know pandas is installed by default, checking the aws code)

When I try to deploy I get an error because inference.py has on the first line import dill saying the module is not found and in Cloudwatch I see only inference 1.0.0 is installed (which I assume is my script).

I know I could probably create a subprocess within inference.py and call pip there, but I want to do this properly.