ReaL: Efficient RLHF Training of Large Language Models with Parameter Reallocation
Project description
| Documentation | Paper |
ReaL: Efficient RLHF Training for LLMs
with Parameter Reallocation
ReaL (short for ReaLlocation) is a distributed system designed for efficient RLHF training with LLMs.
ReaL introduces a novel approach called parameter reallocation, which dynamically redistributes LLM parameters across the cluster and adapts parallelization strategies during training. By optimizing allocations and parallelism for each computation workload, ReaL minimizes redundant communication while maximizing GPU utilization.
ReaL achieves significantly higher PPO training throughput compared to state-of-the-art open-source systems.
(In the following figure, as the number of GPUs increases, the model size scales up from LLaMA 7B, LLaMA 13B, and CodeLLaMA 34B, to the largest LLaMA 70B.)
Highlights
Efficiency
- Achieves state-of-the-art training throughput for RLHF using parameter reallocation.
- Supports large-scale training with 3D parallelism, ZeRO optimization, and sequence parallelism.
- Enables memory-efficient training with parameter and optimizer offloading.
Ease of Use
- Seamlessly integrates with HuggingFace checkpoints and inference frameworks like vLLM.
- Allows launching local or distributed experiments with a single command.
Check out our tutorial to reproduce the full RLHF procedure (SFT/RW/PPO) with 4×LLaMA-7B in just 30 minutes.
Flexibility
- Offers versatile configuration customization with Hydra structured config.
- Supports many commonly used RLHF algorithms, including DPO, PPO, RAFT, and more.
- Allows the addition of custom algorithms with fewer than 100 lines of code.
Refer to our customization guide for hands-on examples.
Getting Started
We provide pre-built Docker images and PyPI packages.
pip3 install realhf --no-build-isolation
For detailed information, please visit our documentation site.
Acknowledgement
We would like to thank the authors of our paper and the following individuals for their contributions: Shusheng Xu and Jiaxuan Gao from Tsinghua University, and Weilin Liu, Wenjie Ye, and Chuyi He from OpenPsi Inc, for thoroughly testing and using ReaL in their research, and for providing valuable suggestions that greatly improved the system.
Citation
If you find our system useful for your research or production, please cite our paper.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.