Implementation:Alibaba ROLL RewardFLConfig

Knowledge Sources	Alibaba ROLL
Domains	Diffusion_Models, Configuration
Last Updated	2026-02-07 20:00 GMT

Overview

Concrete reward flow configuration dataclass for video diffusion training provided by the Alibaba ROLL library.

Description

The RewardFLConfig class extends BaseConfig with settings for diffusion model training including batch size, gradient norm, and actor_train WorkerConfig with strategy_name="diffusion_deepspeed_train".

Usage

Loaded from YAML via Hydra for reward flow pipelines.

Code Reference

Source Location

• Repository: Alibaba ROLL
• File: roll/pipeline/diffusion/reward_fl/reward_fl_config.py
• Lines: L12-48

Signature

@dataclass
class RewardFLConfig(BaseConfig):
    """
    Configuration for reward flow diffusion training.

    Attributes:
        train_batch_size: int = 8 - batch size
        max_grad_norm: float = 1.0 - gradient clipping
        actor_train: WorkerConfig - worker config with diffusion_deepspeed_train strategy
    """

Import

from roll.pipeline.diffusion.reward_fl.reward_fl_config import RewardFLConfig

I/O Contract

Inputs

Name	Type	Required	Description
YAML config file	str	Yes	Hydra-managed YAML with diffusion model paths

Outputs

Name	Type	Description
RewardFLConfig	RewardFLConfig	Config with actor_train WorkerConfig

Usage Examples

from hydra import compose, initialize
import dacite
from omegaconf import OmegaConf

initialize(config_path="examples/wan2.2-14B-reward_fl_ds")
cfg = compose(config_name="reward_fl_config")
config = dacite.from_dict(data_class=RewardFLConfig, data=OmegaConf.to_container(cfg, resolve=True))

Related Pages

Implements Principle

• Principle:Alibaba_ROLL_Reward_Flow_Configuration

Environment Dependencies

This implementation requires the following environment constraints:

• Environment:Alibaba_ROLL_Python_Runtime_Environment
• Environment:Alibaba_ROLL_Diffusion_Video_Environment

Heuristics Applied

No specific heuristics apply to this implementation.