Function get_tiny_config

utils/create_dummy_models.py:591–716

Retrieve a tiny configuration from `config_class` using each model's `ModelTester`. Args: config_class: Subclass of `PreTrainedConfig`. Returns: An instance of `config_class` with tiny hyperparameters

(config_class, model_class=None, **model_tester_kwargs)

# TODO: Sam2Video will fail here
def get_tiny_config(config_class, model_class=None, **model_tester_kwargs):
    """Retrieve a tiny configuration from `config_class` using each model's `ModelTester`.

    Args:
        config_class: Subclass of `PreTrainedConfig`.

    Returns:
        An instance of `config_class` with tiny hyperparameters
    """
    model_type = config_class.model_type

    # For model type like `data2vec-vision` and `donut-swin`, we can't get the config/model file name directly via
    # `model_type` as it would be sth. like `configuration_data2vec_vision.py`.
    # A simple way is to use `inspect.getsourcefile(config_class)`.
    config_source_file = inspect.getsourcefile(config_class)
    # The modeling file name without prefix (`modeling_`) and postfix (`.py`)
    modeling_name = config_source_file.split(os.path.sep)[-1].replace("configuration_", "").replace(".py", "")
    # TODO: remark: several configuration classes might be defined in the same modeling directory.
    #   The test directory is still the same, so we are good here.

    try:
        print("Importing", model_type_to_module_name(model_type))
        module_name = model_type_to_module_name(model_type)
        if not modeling_name.startswith(module_name):
            raise ValueError(f"{modeling_name} doesn't start with {module_name}!")
        test_file = os.path.join("tests", "models", module_name, f"test_modeling_{modeling_name}.py")
        models_to_model_testers = get_model_to_tester_mapping(test_file)
        # Find the model tester class
        model_tester_class = None
        tester_classes = []
        if model_class is not None:
            tester_classes = get_tester_classes_for_model(test_file, model_class)
        else:
            for _tester_classes in models_to_model_testers.values():
                tester_classes.extend(_tester_classes)
        if len(tester_classes) > 0:
            # sort with the length of the class names first, then the alphabetical order
            # This is to avoid `T5EncoderOnlyModelTest` is used instead of `T5ModelTest`, which has
            # `is_encoder_decoder=False` and causes some pipeline tests failing (also failures in `Optimum` CI).
            # TODO: More fine grained control of the desired tester class.
            model_tester_class = min(tester_classes, key=lambda x: (len(x.__name__), x.__name__))

            # TODO: SpeechT5ForSpeechToText needs a particular tester to get the working config
            # TODO: this is hacky however, as all model classes share the same config class but having different tester
            # TODO: make this more flexible and roubst
            if config_class.__name__ == "SpeechT5Config":
                for x in tester_classes:
                    if x.__name__ == "SpeechT5ForSpeechToTextTester":
                        model_tester_class = x
                        break

    except ModuleNotFoundError:
        error = f"Tiny config not created for {model_type} - cannot find the testing module from the model name."
        raise ValueError(error)

    if model_tester_class is None:
        error = f"Tiny config not created for {model_type} - no model tester is found in the testing module."
        raise ValueError(error)