deploy endpoint in complete local mode

[anotinelg]

Is it possible to deploy locally an endpoint without any connection with s3?

With local code, and local instance, i can not get it work.
I look at the following code, and it seems that even if i have local_code, if the repack value is True, it will try to upload to s3 the repacked artefacts.

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

Line 1117 in c4d71f5

if repack:

[ChoiByungWook]

Hey @anotinelg,

Thanks for bringing this issue to our attention.

This looks like a gap in our Python SDK for supporting true local mode when using our Framework Estimator/Model.

There doesn't seem to be a way to disable repack_model through the Framework Estimator/Model constructors.

I think one workaround that may work is to use the Estimator/Model class and point the image_uri to the corresponding image that is used in the Framework Estimator you were attempting to run local mode with, as the base Model class shouldn't call repack_model.

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

Line 282 in c4d71f5

return sagemaker.container_def(self.image_uri, self.model_data, self.env)

I will note this in backlog, which is prioritized based on customer feedback.

[anotinelg]

Actually i am using the MxnetModel which will use the Framework Model and hence the method prepare_container_def which calls the upload_code. I am not sure i understand what you propose. Instead of using MXNetModel, i should directly use the Model class?

[ChoiByungWook]

Actually i am using the MxnetModel which will use the Framework Model and hence the method prepare_container_def which calls the upload_code. I am not sure i understand what you propose. Instead of using MXNetModel, i should directly use the Model class?

Yeah, the MXNetModel class ends up extending the FrameworkModel class, which then extends the Model class.

I believe the base Model class doesn't call repack_model, however you would lose out on a few things that MXNetModel class does for you such as the image_uri, repacking of the model with your inference script. These things you would have to supply to the base Model class.