Implementation:Hiyouga LLaMA Factory V1 Arg Utils

Knowledge Sources	Hiyouga_LLaMA_Factory
Domains	Configuration, Plugin System, Enums
Last Updated	2026-02-06 19:00 GMT

Overview

This module defines configuration enums and plugin configuration utilities used across the V1 argument dataclasses.

Description

The module provides four key components: PluginConfig (a dict subclass with a name property for plugin identification), ModelClass (a StrEnum with LLM, CLS, OTHER values for auto-class selection), SampleBackend (a StrEnum with HF, VLLM values for sampling backend selection), and BatchingStrategy (a StrEnum with NORMAL, PADDING_FREE, DYNAMIC_BATCHING, DYNAMIC_PADDING_FREE values for batch assembly strategy). The get_plugin_config function parses plugin configurations from JSON strings or dictionaries, performing automatic type conversion of string values to booleans, integers, and floats via _convert_str_dict.

Usage

Use these enums and utilities when defining or consuming V1 configuration arguments. PluginConfig and get_plugin_config are used to parse user-provided plugin configurations in model, data, trainer, and sampler argument dataclasses. The enum classes standardize configuration values across the V1 training system.

Code Reference

Source Location

• Repository: Hiyouga_LLaMA_Factory
• File: src/llamafactory/v1/config/arg_utils.py
• Lines: 1-103

Signature

class PluginConfig(dict):
    @property
    def name(self) -> str

PluginArgument = PluginConfig | dict | str | None

class ModelClass(StrEnum):
    LLM = "llm"
    CLS = "cls"
    OTHER = "other"

class SampleBackend(StrEnum):
    HF = "hf"
    VLLM = "vllm"

class BatchingStrategy(StrEnum):
    NORMAL = "normal"
    PADDING_FREE = "padding_free"
    DYNAMIC_BATCHING = "dynamic_batching"
    DYNAMIC_PADDING_FREE = "dynamic_padding_free"

def get_plugin_config(config: PluginArgument) -> PluginConfig | None

Import

from llamafactory.v1.config.arg_utils import (
    PluginConfig, PluginArgument, ModelClass, SampleBackend,
    BatchingStrategy, get_plugin_config,
)

I/O Contract

Inputs

Name	Type	Required	Description
config (get_plugin_config)	PluginArgument	Yes	Plugin configuration as a JSON string (starting with "{"), a dict, a PluginConfig, or None

Outputs

Name	Type	Description
get_plugin_config result	PluginConfig or None	Parsed plugin configuration dict with name property; None if input is None
PluginConfig.name	str	The "name" field from the plugin configuration dictionary

Usage Examples

from llamafactory.v1.config.arg_utils import (
    get_plugin_config, ModelClass, SampleBackend, BatchingStrategy,
)

# Parse plugin config from JSON string
config = get_plugin_config('{"name": "fsdp2", "sharding_strategy": "FULL_SHARD"}')
print(config.name)          # "fsdp2"
print(config["sharding_strategy"])  # "FULL_SHARD"

# Parse plugin config from dict
config = get_plugin_config({"name": "deepspeed", "stage": "3"})
print(config["stage"])      # 3 (auto-converted from string to int)

# Use enum values
model_cls = ModelClass.LLM           # "llm"
backend = SampleBackend.VLLM         # "vllm"
strategy = BatchingStrategy.PADDING_FREE  # "padding_free"

# None returns None
config = get_plugin_config(None)  # None

Related Pages

• Hiyouga_LLaMA_Factory_V1_Config_Model_Args - Model arguments that use ModelClass and PluginConfig
• Hiyouga_LLaMA_Factory_V1_Config_Training_Args - Training arguments that use BatchingStrategy and PluginConfig
• Hiyouga_LLaMA_Factory_V1_Config_Sample_Args - Sample arguments that use SampleBackend
• Hiyouga_LLaMA_Factory_V1_Config_Arg_Parser - Argument parser that processes these configuration types