Edit on GitHub


MLEM can be extended to support more model types, data types, servers, builders and basically everything listed here. Most of the builtin implementations are also extensions located in mlem.contrib package. It allows MLEM to not load their code if it is not used, which is especially cool because it means their requirements are optional.

Implementing MlemABC

You can start extending MLEM by subclassing any of the MlemAbc subclass that you need.

But no one tried it so far ;)

Your subclass should implement all the abstract methods of the base class.

Also, it needs to define type: ClassVar[str] class field, which will be used as an alias for your implementation.

By default, type will have <module>.<class name> value, but that's not very handy to type in cli, e.g. you'll need to run mlem serve model my_awesome_package.submodule_of_my_awesome_package.abstract.bean.factory.MyAwesomeServerImplementation instead of mlem serve model ъуъ if you don't set type: ClassVar = "ъуъ" for your class

Entry points

For MLEM to know about your implementations, you need to register them via entry points in your setup.py.

You should list all of them in the form {abs_name}.{type} = {module_path}:{class_name} under mlem.contrib entry point key, where

  • abs_name is MlemABC.abs_name of the interface you are implementing
  • type is a value of type field name of your class
  • module path is full path to python module
  • class name is the name of your class

You can see examples in MLEM's setup.py

Extension dynamic loading

By default, when you import MLEM or run MLEM cli commands, MLEM will not load any extensions to minimize overhead. But that would mean that users will have to import them manually, and we don't want that. MLEM can load extensions dynamically, depending on what is imported in user's environment. For example, sklearn extension will be loaded in one of the following cases:

  1. When user imported mlem, sklearn module was already imported
  2. After importing mlem, user imported sklearn
  3. User loaded any object that uses any of sklearn extension implementation.

Some of the fields in MLEM Objects are lazy, which means that they will be loaded only if users accesses them.

Subclassing MlemConfig

As part of your extension, you also can have some configuration options. For that you can subclass MlemConfig class and list your options there just like any pydantic BaseSettings class. In the inner Config class you should set section option, and after that values for your configuration will be loaded from .mlem/config.yaml from corresponding section. See PandasConfig for example


🐛 Found an issue? Let us know! Or fix it:

Edit on GitHub

Have a question? Join our chat, we will help you:

Discord Chat