Implementation:Volcengine Verl BaseConfig

Knowledge Sources	Volcengine_Verl
Domains	Configuration, Data_Structures, API_Surface
Last Updated	2026-02-07 18:00 GMT

Overview

Concrete tool for creating frozen, dict-like dataclass configurations provided by the verl library.

Description

The BaseConfig class is a Python dataclass that implements collections.abc.Mapping, providing a dictionary-like interface for configuration objects. Key features:

• Immutability by default — All fields are frozen after initialization. Attempting to modify a field raises FrozenInstanceError, unless the field name is listed in the class-level _mutable_fields set.
• Dict-like access — Supports bracket notation (config["key"]), .get() with defaults, iteration over field names, and len().
• Hydra compatibility — Includes a _target_ field for Hydra structured config instantiation.

All verl configuration dataclasses (RolloutConfig, ActorConfig, etc.) inherit from BaseConfig to get these safety and convenience features.

Usage

Use this class as the base when defining new configuration dataclasses in verl. Subclass BaseConfig and define fields with type annotations and defaults.

Code Reference

Source Location

• Repository: Volcengine_Verl
• File: verl/base_config.py
• Lines: 1-86

Signature

@dataclass
class BaseConfig(collections.abc.Mapping):
    """The BaseConfig provides dict-like interface for a dataclass config.

    By default all fields in the config is not mutable, unless specified in
    "_mutable_fields". The BaseConfig class implements the Mapping Abstract Base Class.
    """

    _mutable_fields = set()
    _target_: str = ""

    def __setattr__(self, name: str, value): ...
    def get(self, key: str, default: Any = None) -> Any: ...
    def __getitem__(self, key: str): ...
    def __iter__(self): ...
    def __len__(self) -> int: ...

Import

from verl.base_config import BaseConfig

I/O Contract

Inputs

Name	Type	Required	Description
_mutable_fields	set	No	Class-level set of field names that can be modified after init
_target_	str	No	Hydra _target_ for structured config instantiation

Outputs

Name	Type	Description
instance	BaseConfig	A frozen, dict-like dataclass instance
__getitem__	Any	Value of the requested field
__iter__	Iterator[str]	Field names of the dataclass
__len__	int	Number of fields in the dataclass

Usage Examples

Defining a Custom Config

from dataclasses import dataclass
from verl.base_config import BaseConfig

@dataclass
class MyTrainingConfig(BaseConfig):
    learning_rate: float = 1e-4
    batch_size: int = 32
    num_epochs: int = 3
    _mutable_fields = {"num_epochs"}  # Only num_epochs can be changed

config = MyTrainingConfig(learning_rate=2e-4, batch_size=64)

# Dict-like access
print(config["learning_rate"])  # 2e-4
print(config.get("missing_key", "default"))  # "default"
print(len(config))  # 4 (including _target_)

# Immutability
config.num_epochs = 5  # OK — in _mutable_fields
# config.learning_rate = 3e-4  # Raises FrozenInstanceError

Iterating Over Config

config = MyTrainingConfig()

for key in config:
    print(f"{key} = {config[key]}")
# _mutable_fields = set()
# _target_ =
# learning_rate = 0.0001
# batch_size = 32
# num_epochs = 3

Related Pages

• Environment:Volcengine_Verl_Python_Core_Dependencies