lightrft.utils.cli_args

Command-line Argument Parser Configuration Module

This module provides functionality to configure command-line arguments for training and inference with various distributed training frameworks like vLLM, SGL, DeepSpeed, and FSDP. It includes arguments for model parallelism, memory utilization, sequence packing, gradient clipping, and logging options.

add_arguments

lightrft.utils.cli_args.add_arguments(parser: ArgumentParser) → None

Add training and inference related arguments to an ArgumentParser.

This function configures an argument parser with options for: - Inference Engine settings (vLLM/SGLang) - Training parameters - FSDP (Fully Sharded Data Parallelism) configuration - Logging and visualization options

Parameters:

parser (argparse.ArgumentParser) – The argument parser to which arguments will be added

Returns:

None

Example:

>>> import argparse
>>> parser = argparse.ArgumentParser()
>>> add_arguments(parser)
>>> args = parser.parse_args()