Optimizing RLHF Training for Large Language Models with Stage Fusion

TL;DR

RLHFuse introduces stage fusion techniques to optimize GPU utilization and significantly boost training throughput in RLHF for large language models by overlapping tasks and reducing pipeline inefficiencies.

Contribution

The paper proposes a novel stage fusion approach for RLHF training that enhances GPU efficiency by splitting tasks into finer subtasks and overlapping execution stages.

Findings

Training throughput increased by up to 3.7 times.

Effective mitigation of generation bottlenecks and pipeline bubbles.

Improved GPU utilization in RLHF workflows.

Abstract

We present RLHFuse, an efficient training system with stage fusion for Reinforcement Learning from Human Feedback (RLHF). Due to the intrinsic nature of RLHF training, i.e., the data skewness in the generation stage and the pipeline bubbles in the training stage, existing RLHF systems suffer from low GPU utilization. RLHFuse breaks the traditional view of RLHF workflow as a composition of individual tasks, splitting each task into finer-grained subtasks, and performing stage fusion to improve GPU utilization. RLHFuse contains two key ideas. First, for generation and inference tasks, RLHFuse splits them into sample-level subtasks, enabling efficient inter-stage fusion to overlap the execution of generation and inference stages, thus mitigating the original generation bottleneck dominated by long-tailed samples. Second, for training tasks, RLHFuse breaks them into subtasks of micro-batches and performs intra-stage fusion to concurrently execute these subtasks in the training stage with a fused pipeline schedule, effectively mitigating the pipeline bubbles. The experiments show that RLHFuse increases the training throughput by up to $3.7\times$, compared to existing systems.