ms-swift

Alibaba's ModelScope community maintains ms-swift as its official training stack, covering fine-tuning, alignment, and deployment for more than 600 text LLMs and 300 multimodal models, with first-class support for the Qwen family alongside DeepSeek, GLM, InternLM, and Llama. Techniques span the full post-training menu: supervised fine-tuning, LoRA and QLoRA, GRPO, PPO, DPO, KTO, CPO, SimPO, ORPO, reward modeling, GKD distillation, and embedding or reranker training, with a Megatron backend for large dense and MoE models and Ray integration for distributed jobs. A pip install ms-swift provides the CLI, and a Gradio web UI covers training and inference without writing code. Effective use requires NVIDIA GPUs, with DeepSpeed and quantized modes stretching what fits on a single card. The project is Apache-2.0 licensed with documentation at swift.readthedocs.io. Roughly 15,000 GitHub stars and rapid releases tracking each new Qwen and multimodal model generation make it the default choice for tuning open models from the Chinese ecosystem.