A configurable module is the fundamental unit of Atria: a class whose behavior is fully described by an attached Pydantic config object, so it can be constructed from a config alone.

ModuleConfig

ModuleConfig is the base class for every configuration in Atria. It is a frozen Pydantic BaseModel with a few extras:

• _target_ injection — the JSON schema automatically includes a _target_ field pointing to the config's class path. This lets Hydra's instantiate reconstruct any config from a plain dict.
• hash — a deterministic hash of the config's fields, used by the registry to detect duplicates and cache results.
• build(**kwargs) — resolves __module_path__ to the owning ConfigurableModule class and constructs it with config=self.
• to_dict() / from_dict() — round-trip to/from Hydra-instantiable dicts for YAML/JSON serialization.

class MyTransformConfig(ModuleConfig):
    kernel_size: int = 3
    sigma: float = 1.0

ConfigurableModule

ConfigurableModule is the base class every registered component inherits from. It is generic over its config type:

class MyTransform(ConfigurableModule[MyTransformConfig]):
    __config__ = MyTransformConfig

    def __call__(self, x):
        ...

Contract enforced at class definition time: - Every non-abstract subclass must declare __config__. - __config__ must be a ModuleConfig subclass. - __config__.__module_path__ is automatically set to the fully-qualified class path of the owning module.

This means the config always knows which class to build, and the class always knows its config type — no separate registration step required.

PydanticConfigurableModule

For components that are themselves Pydantic models (rather than holding a config object), PydanticConfigurableModule provides the same hash, to_dict, and from_dict API without the ConfigurableModule base class overhead. Used internally for data pipeline and storage configurations.

Lifecycle

config dict (YAML/JSON)
    ↓  ModuleConfig.from_dict()
ModuleConfig instance
    ↓  config.build()  OR  RegistryGroup.load_module_config()
ConfigurableModule instance

This lifecycle is the same for datasets, models, transforms, explainers, optimizers — every component in the framework.