Implementation:Alibaba ROLL Diffusion DeepSpeed Cluster

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

Overview

Concrete diffusion-specific DeepSpeed cluster initialization provided by the Alibaba ROLL library.

Description

The reward flow pipeline creates a single actor_train cluster with strategy_name="diffusion_deepspeed_train". The diffusion DeepSpeed strategy wraps the WanTrainingModule with DeepSpeed engine management.

Usage

Called during RewardFLPipeline initialization.

Code Reference

Source Location

• Repository: Alibaba ROLL
• File: roll/distributed/strategy/diffusion_strategy.py
• Lines: L19-84

Signature

# Diffusion-specific DeepSpeed train strategy
class DeepSpeedTrainStrategy:
    """Diffusion-specific DeepSpeed wrapper."""

    def train_step(self, batch: DataProto, loss_func: Callable) -> Dict:
        """Execute diffusion training step through DeepSpeed engine."""

    def save_checkpoint(self, save_dir, global_step, ckpt_id, **kwargs) -> None:
        """Save only trainable (LoRA) parameters."""

Import

from roll.distributed.strategy.diffusion_strategy import DeepSpeedTrainStrategy

I/O Contract

Inputs

Name	Type	Required	Description
config	RewardFLConfig	Yes	Config with actor_train WorkerConfig

Outputs

Name	Type	Description
actor_train	Cluster	Single cluster with WanTrainingModule wrapped in DeepSpeed

Usage Examples

pipeline = RewardFLPipeline(pipeline_config=reward_fl_config)
# actor_train cluster with diffusion_deepspeed_train strategy

Related Pages

Implements Principle

• Principle:Alibaba_ROLL_Diffusion_Worker_Initialization

Environment Dependencies

This implementation requires the following environment constraints:

• Environment:Alibaba_ROLL_CUDA_GPU_Environment
• Environment:Alibaba_ROLL_DeepSpeed_Training_Environment
• Environment:Alibaba_ROLL_Diffusion_Video_Environment

Heuristics Applied

This implementation uses the following heuristics:

• Heuristic:Alibaba_ROLL_GPU_Memory_Offload_Strategy