Failed to import code copied into the /opt/ml/model/code/lib directory

[anotinelg]

System Information

Python Version: 3.6
sagemaker-python-sdk Version: master (1.42.1.dev0)
sagemaker-mxnet-serving-container: 1.1.4.dev0) (using sagemkaer-infernece 1.0.1)
Are you using a custom image: no

Describe the problem

I am trying to deploy a model with a script depending on a library. The dependency to the library is configured with the dependenciesparameter of the MXNetmodel.
The model is successfully created, but When the model is started, the lib is not found by the code. It seems that that the sagemaker-mxnet-serving-container does not include in sys.path the /opt/ml/code/lib directory.
I am not sure if the integration bug must be solved into that repository or the sagemaker-mxnet-serving-container. As far as I understand, there is a fundamental changes in that repository which breaks the integration between both.

My source directory is:

- src
---- mypackage
--------__init__.py
--------myscript.py
--------utils.py

I create the model as:

MXNetModel( entry_point='../src/mypackage/myscript.py',
                         dependencies= ['../src/mypackage']
                         ...       
                         py_version='py3')

The output.tar.gz is the following:

- model/
---- my model.params
---- ...
---- code /
----------myscript.py (the script is copied at root)
----------lib/ 
--------------mypackage (the dependent lib is copied into the lib/ directory)
-------------------__init__.py
-------------------myscript.py
-------------------utils.py

The code in question is

sagemaker-python-sdk/src/sagemaker/utils.py

Line 471 in d38ed76

lib_dir = os.path.join(code_dir, "lib")

Nota: I have found a solution adding the following into the myscript.py:

sys.path.append('/opt/ml/model/code/lib')

But I guess this should not be the place to do it.

[mvsusp]

Hi @anotinelg,

Like you mentioned, the MXNet serving container should be able to install Python modules from the /opt/ml/code/lib folder. We will add this issue to the roadmap and let you know here when it will fixed.

Thanks for using SageMaker.

[akrishna1995]

Hi @anotinelg - just double checking if this issue still persists in the latest sagemaker ? seems like the fixes have been merged already. Feel free to reopen if there are any pending issues. Thank you so much