hiyouga/EasyR1

Leverages HybridEngine architecture and vLLM's SPMD mode for efficient distributed training across multiple algorithms (GRPO, DAPO, Reinforce++, ReMax, RLOO, GSPO, CISPO). Supports both language models (Llama3, Qwen2/2.5/3) and vision-language models (Qwen2-VL/2.5-VL/3-VL, DeepSeek-R1 distill) with padding-free training, LoRA fine-tuning, and multi-node Ray-based orchestration. Integrates with vLLM for inference and standard ML experiment trackers (Weights & Biases, SwanLab, MLflow, TensorBoard).

4,721 stars. Actively maintained with 4 commits in the last 30 days.