llm.config

llm/config.py

HParamT module-attribute

HParamT = Union[bool, float, int, str, None]

Supported Hyperparameter types (i.e., JSON types).

Config

Config(
    mapping: (
        Mapping[str, Any] | Iterable[tuple[str, Any]] | None
    ) = None,
    /,
    **kwargs: Any,
)

Bases: dict[str, Any]

Dict-like configuration class with attribute access.

Example

>>> from llm.config import Config
>>> config = Config({'a': 1, 'b': 2})
>>> config.a
1
>>> config = Config(a=1, b={'c': 2})
>>> config.b.c
2

Parameters:

• mapping (Mapping[str, Any] | Iterable[tuple[str, Any]] | None, default: None ) –

  Initial mapping or iterable of tuples of key-value pairs.

• kwargs (Any, default: {} ) –

  Keywords arguments to add a key-value pairs to the config.

Source code in llm/config.py

def __init__(
    self,
    mapping: Mapping[str, Any] | Iterable[tuple[str, Any]] | None = None,
    /,
    **kwargs: Any,
):
    if mapping is not None:
        if isinstance(mapping, Mapping):
            mapping = mapping.items()
        for key, value in mapping:
            self.__setattr__(key, value)

    for key, value in kwargs.items():
        self.__setattr__(key, value)

flattened_config

flattened_config(
    config: dict[str, Any] | Config | None = None
) -> dict[str, HParamT]

Convert a config to a flat JSONable dictionary.

Note

If torch.distributed.is_initialized(), the world_size will be added to the config.

Parameters:

• config (dict[str, Any] | Config | None, default: None ) –

  Optional starting config.

Returns:

• dict[str, HParamT] –

  Flat dictionary containing only bool, float, int, str, or

• dict[str, HParamT] –

  None values.

Source code in llm/config.py

def flattened_config(
    config: dict[str, Any] | Config | None = None,
) -> dict[str, HParamT]:
    """Convert a config to a flat JSONable dictionary.

    Note:
        If
        [`torch.distributed.is_initialized()`][torch.distributed.is_initialized],
        the `world_size` will be added to the config.

    Args:
        config: Optional starting config.

    Returns:
        Flat dictionary containing only `bool`, `float`, `int`, `str`, or
        `None` values.
    """
    if config is None:
        config = {}

    if dist.is_initialized():
        config['world_size'] = dist.get_world_size()

    config = flatten_mapping(config)
    for key in list(config.keys()):
        if not isinstance(config[key], (bool, float, int, str, type(None))):
            del config[key]

    return config

flatten_mapping

flatten_mapping(
    d: Mapping[str, Any],
    parent: str | None = None,
    sep: str = "_",
) -> dict[str, Any]

Flatten mapping into dict by joining nested keys via a separator.

Warning

This function does not check for key collisions. E.g.,

>>> flatten_mapping({'a': {'b_c': 1}, 'a_b': {'c': 2}})
{'a_b_c': 2}

Parameters:

• d (Mapping[str, Any]) –

  Input mapping. All keys and nested keys must by strings.

• parent (str | None, default: None ) –

  Parent key to prepend to top-level keys in d.

• sep (str, default: '_' ) –

  Separator between keys.

Returns:

• dict[str, Any] –

  Flattened dictionary.

Source code in llm/config.py

def flatten_mapping(
    d: Mapping[str, Any],
    parent: str | None = None,
    sep: str = '_',
) -> dict[str, Any]:
    """Flatten mapping into dict by joining nested keys via a separator.

    Warning:
        This function does not check for key collisions. E.g.,
        ```python
        >>> flatten_mapping({'a': {'b_c': 1}, 'a_b': {'c': 2}})
        {'a_b_c': 2}
        ```

    Args:
        d: Input mapping. All keys and nested keys must by strings.
        parent: Parent key to prepend to top-level keys in `d`.
        sep: Separator between keys.

    Returns:
        Flattened dictionary.
    """
    # https://stackoverflow.com/questions/6027558
    items: list[tuple[str, Any]] = []

    for key, value in d.items():
        new_key = f'{parent}{sep}{key}' if parent is not None else key
        if isinstance(value, collections.abc.Mapping):
            items.extend(flatten_mapping(value, new_key, sep).items())
        else:
            items.append((new_key, value))

    return dict(items)

load_config

load_config(filepath: Path | str) -> Config

Load Python file as a Config.

Note

Attributes starting with _, modules, classes, functions, and builtins will not be loaded from the Python file.

Parameters:

• filepath (Path | str) –

  Python file to load.

Returns:

• Config –

  Configuration attributes loaded from the Python file.

Source code in llm/config.py

def load_config(filepath: pathlib.Path | str) -> Config:
    """Load Python file as a [`Config`][llm.config.Config].

    Note:
        Attributes starting with `_`, modules, classes, functions, and
        builtins will not be loaded from the Python file.

    Args:
        filepath: Python file to load.

    Returns:
        Configuration attributes loaded from the Python file.
    """
    filepath = pathlib.Path(filepath).absolute()
    if not filepath.exists():
        raise OSError(f'{filepath} does not exist.')
    elif filepath.suffix != '.py':
        raise ValueError(
            f'{filepath} is not a Python file. Only .py files are supported.',
        )

    loader = SourceFileLoader(fullname=filepath.stem, path=str(filepath))
    module = types.ModuleType(loader.name)
    loader.exec_module(module)

    attrs: dict[str, Any] = {}
    for key, value in module.__dict__.items():
        if not (
            key.startswith('_')
            or inspect.ismodule(value)
            or inspect.isclass(value)
            or inspect.isfunction(value)
            or inspect.isbuiltin(value)
        ):
            attrs[key] = value

    return Config(**attrs)