A flexible and efficient training framework for large-scale alignment

 English  | 中文  

--- *Latest News* 🔥 - [2024/8] We officially released ChatLearn! Check out our [documentation](docs/en/chatlearn.md). - [ongoing] We are continuously hiring and welcome you to contact us or submit your resume to wanglin.zj@alibaba-inc.com. --- ChatLearn is a large-scale alignment training framework developed by the Alibaba Cloud PAI platform.  Chatlearn has the following advantages: 1. **User-friendly programming interface**: Users can focus on programming individual models by wrapping a few functions, while the system takes care of resource scheduling, data and control flow transmission, and distributed execution. 2. **Highly Scalable Training Methodology**: ChatLearn offers alignment training such as RLHF, DPO, OnlineDPO and GRPO, while also supporting user-defined execution flows for models, enabling a highly convenient and customizable training process. 3. **Diverse Distributed Acceleration Engines**: Users can leverage various computational backends for model construction, such as Megatron-LM, DeepSpeed, vLLM, and others. For instance, we can use Megatron-LM for training and vLLM to expedite inference. 4. **Flexible Parallel Strategies and Resource Allocation**: ChatLearn supports different parallel strategies for various model configurations, enabling the formulation of distinct parallel approaches tailored to each model's computational, memory, and communication characteristics. Additionally, ChatLearn features a flexible resource scheduling mechanism that accommodates exclusive or shared use of resources across models. Through its system scheduling policies, it facilitates efficient serial/parallel execution and optimized GPU memory sharing, enhancing overall performance and efficiency. 5. **High performance**: Compared to current state-of-the-art (SOTA) systems, ChatLearn achieves a 52% performance improvement at the 7B+7B(Policy+Reward) scale and a 137% improvement at the 70B+70B scale. Meanwhile, ChatLearn supports larger-scale alignment training, such as 300B+300B. By providing a comprehensive and efficient framework, ChatLearn empowers researchers and practitioners to train large-scale alignment models with ease, scalability, and improved performance. # Quick Start Please refer to the [documentation](https://chatlearn.readthedocs.io/en/latest/) for a quick start. 1. [Environment and Code Setup](docs/en/installation.md) 2. [End-to-End Training Tutorial with LLaMA/LLaMA2 Model](docs/en/tutorial/tutorial_llama2.md) # Performance We compared the RLHF training throughput of models with different parameter scales, adopting an N+N model configuration where both the Policy model and the Reward model have the same number of parameters. We benchmarked against DeepSpeed-Chat and OpenRLHF with 7B and 70B model configurations. For the 8 GPU setup with a 7B+7B scale, we achieved a 115% speedup; for the 32 GPU setup with a 70B+70B scale, the speedup was 208%. The larger the scale, the more pronounced the acceleration effect becomes. Additionally, ChatLearn can support even larger-scale alignment training, such as at a 300B+300B scale.  Note: The performance of DeepSpeed-Chat and OpenRLHF has already been optimized. # Feature List - Supports RLHF, DPO, OnlineDPO, GRPO, and user-defined Alignment training methods. - Supports Megatron-LM as the backend for training or inference, and vLLM as the backend for inference. - Supports independent configuration of parallel strategies for different models, and efficient parameter synchronization between models. - Supports EMS (Efficient Memory Sharing) functionality, enabling efficient memory sharing between models. - Supports resource types for models: GPU, CPU, such as defining a pure CPU-based Math Reward model. - Support models with Megatron-Core format. # Roadmap The upcoming features for ChatLearn include: - [ ] Support the alignment training for MoE (Mixture of Experts) models - [ ] Integration with DeepSpeed as a training backend - [ ] Support for more models - [ ] Performance Optimization - [ ] Support for more alignment algorithms

We welcome community partners to collaborate and contribute to the development, and welcome to join the DingTalk group: 98090003312 to participate in the discussion.