omniconfig

Python package for parsing configurations from YAML and TOML files with the command-line interface.

Installation

Pip

You can install omniconfig using pip:

pip install omniconfig

Build from source

1. Clone this repository and navigate to lmquant folder

git clone https://github.com/synxlin/omniconfig
cd omniconfig

2. Install Package

conda env create -f environment.yml
poetry install

Usage

Decorator configclass in omniconfig package can help build argument parser from dataclass in dataclasses. Here is an example:

from dataclasses import dataclass, field
from typing import Any, Callable

import omniconfig


@omniconfig.configclass
@dataclass
class EvalConfig:
    """Evaluation Config class.

    Attributes:
        num_gpus (int): the number of GPUs. Defaults to ``8``.
    """

    num_gpus: int = 8


@omniconfig.configclass
@dataclass
class QuantConfig:
    """Quantization config class.

    Attributes:
        dtype (str): Quantization data type. Defaults to ``"torch.float16"``.
        group_shape (list[int]): Quantization group shape. Defaults to ``(1, -1)``.
    """

    dtype: str = "torch.float16"
    group_shape: list[int] = (1, -1)

    @classmethod
    def update_get_arguments(
        cls: type["QuantConfig"],
        *,
        overwrites: dict[str, Callable | None] | None = None,
        defaults: dict[str, Any] | None = None,
    ) -> tuple[dict[str, Callable | None], dict[str, Any]]:
        """Get the arguments for the quantization configuration."""
        overwrites = overwrites or {}
        defaults = defaults or {}
        overwrites.setdefault(
            "group_shape",
            lambda parser: parser.add_argument(
                "--group-shape",
                nargs="+",
                type=int,
                default=defaults.get("group_shape", (1, -1)),
                help="Quantization group shape",
            ),
        )
        return overwrites, defaults


@omniconfig.configclass
@dataclass
class ModelConfig:
    """Model Config class.

    Attributes:
        name (str): the model name.
        group (str): the model family.
    """

    name: str
    family: str = field(init=False, default="")

    def __post_init__(self):
        self.family = self.name.split("-")[0]


@omniconfig.configclass
@dataclass
class QuantizedModelConfig(ModelConfig):
    """Quantized Model Config class.

    Attributes:
        name (str): the model name.
        group (str): the model group.
        quant (QuantConfig): quantization configuration. Defaults to ``None``.
    """

    quant: QuantConfig | None = None


@omniconfig.configclass
@dataclass
class Config:
    """Config class.

    Attributes:
        model (QuantizedModelConfig): the quantized model config.
        eval (EvalConfig): the evaluation config.
    """

    model: QuantizedModelConfig
    eval: EvalConfig

    @staticmethod
    def parse_args(args: Any = None) -> tuple["Config", dict[str, dict], list[str]]:
        """Parse arguments.

        Args:
            args (list[str], optional): Arguments to parse. Defaults to ``None``.

        Returns:
            tuple[Config, dict[str, dict], list[str]]: Configs from the parsed arguments,
                                                       parsed yaml configs, and unknown arguments.
        """
        parser = omniconfig.ConfigParser("Evaluate Quantized Model")
        parser.add_config(Config)
        config, parsed_args, unknown_args = parser.parse_known_args(args)
        assert isinstance(config, Config)
        return config, parsed_args, unknown_args

    @staticmethod
    def dump_default(path: str = "default.yaml") -> None:
        """Dump default configuration to a yaml file.

        Args:
            path (str, optional): The path to save the default configuration. Defaults to ``"default.yaml"``.
        """
        parser = omniconfig.ConfigParser("Evaluate Quantized Model")
        parser.add_config(Config)
        parser.dump_default(path)

def main(args: Any = None) -> None:  # noqa: C901
    """Evaluate Quantized Model with the given arguments.

    Args:
        args (list[str], optional): Arguments to parse. Defaults to ``None``.
    """
    config, parsed_args, unknown_args = Config.parse_args(args)
    ...

In the example above, configclass will automatically generate the following flags for command-line argument parser:

usage: Evaluate Quantized Model [-h] [--model-name MODEL_NAME] [--model-enable-quant] [--model-quant-dtype MODEL_QUANT_DTYPE] [--model-quant-group-shape MODEL_QUANT_GROUP_SHAPE [MODEL_QUANT_GROUP_SHAPE ...]]
                                [--eval-num-gpus EVAL_NUM_GPUS]
                                [cfgs ...]

positional arguments:
  cfgs                  config file(s)

options:
  -h, --help            show this help message and exit
  --model-name MODEL_NAME
                        the model name for model.
  --model-enable-quant  Enable quant for model. Default: False.
  --model-quant-dtype MODEL_QUANT_DTYPE
                        Quantization data type for model_quant. Default: torch.float16.
  --model-quant-group-shape MODEL_QUANT_GROUP_SHAPE [MODEL_QUANT_GROUP_SHAPE ...]
                        Quantization group shape for model_quant. Default: (1, -1).
  --eval-num-gpus EVAL_NUM_GPUS
                        the number of GPUs for eval. Default: 8.

Note that

configclass will automatically extract help message from the docstring of the class.

configclass will always set the first positional argument as paths to config files. Current supported config file type is YAML and TOML. An example YAML config should be like:

model:
  name: llama2-7b
  enable_quant: true
  quant:
    dtype: sint8
    group_shape:
    - 1
    - -1
eval:
  num_gpus: 8

in QuantConfig, we overwrite the command-line parser for field group_shape by overriding the classmethod update_get_arguments. We can also add any other field inside the classmethod update_get_arguments and override the classmethod update_from_dict to handle these extra fields.

in ModelConfig, the field family is set with init=False, and thus configclass will not generate parser for family field.

class QuantizedModelConfig inherited from ModelConfig, thus it will also inherited all fields from ModelConfig.

in QuantizedModelConfig, the field quant is a type of Optional[QuantConfig], and thus, the parser will automatically add a --model-enable-quant to help set field quant to None.