Initialization of default JumpStart SageMaker Session is slow

[CallumAtCarter]

Through the following import chain, the DEFAULT_JUMPSTART_SAGEMAKER_SESSION constant is initialized:

<module> (.venv\Lib\site-packages\sagemaker\jumpstart\constants.py:237)
<module> (.venv\Lib\site-packages\sagemaker\image_uris.py:24)
<module> (.venv\Lib\site-packages\sagemaker\estimator.py:30)
<module> (.venv\Lib\site-packages\sagemaker\__init__.py:18)

This involves creating many boto clients (at least 5 by my count), each time trying to find AWS credentials using botocore's CredentialResolver.load_credentials which (because I don't have credentials in environment variables, I'm using STS internally) ends up making slow network requests and is generally inefficient (searches through 11 credential providers). In total it is adding 13 seconds to my local startup time.

I'm using my own SageMaker Session whenever creating a new Predictor, and I don't believe I'm ever using the DEFAULT_JUMPSTART_SAGEMAKER_SESSION (if I modify the SDK source to skip its initialization everything still works). Could it be lazily loaded internally?

System information

• SageMaker Python SDK version: 2.210.0
• Python version: 3.11

[mohanasudhan]

@evakravi

[AndreiVoinovTR]

Hi All, is there any timeline for an investigation and/or fix? We have been experiencing the same effect..