Make torch an optional dependency / extra

[firegate22]

Modern packages use optional dependecies / extras for specific functionality. I assume this is very well possible for the sagemaker sdk.

A similar request was posted in discussions: #5441

I couldn't find any actual "issue" and hope this gets more visibility.

The main and most urgent topic is the dependency of pytorch which requires 354/620mb size in a clean venv. I would assume that most use cases don't need pytorch at all.

[kassick]

The packages sagemaker-{train,serve,mlops} are mentioned in the optional dependencies extras, but also added as non-optional dependencies in pyproject.toml

This change was introduced in 869f3ee

I'm not sure if the code was designed to handle optional dependencies or not, but pulling torch, cuda and the kitchen sink to use the SageMaker API is not ideal.

[edit]: Depending on sagemaker-mlops instead of sagemaker-sdk also does not help much, since -mlops depends on -train, which then brings torch, cuda, etc.

[matteociccozzi-ah]

Surprising this issue has been open for so long with no attention. Looking at the code doesn't seem like torch it actually needed in sagemaker-core.

[aW3st]

This is blocking us from upgrading to Sagemaker 3. Would love an update from the maintainers.

[JoseJuan98]

This seems to be fixed now in the latest version of the sagemaker-core package. But if you install sagemaker or sagemaker-{mlops,serve} the same issue arises as is a core dependency in the sagemaker-serve package, could this one also be optional?

(Duplicated from issue 5457, but I wasn't sure in which one to comment)

[goelakash]

Opened a PR to address this #6166