modelscope/ms-swift

ms-swift is a fine-tuning and deployment framework for large language models and multimodal models that supports parameter-efficient and full-parameter training across hundreds of model architectures.

The tool addresses the challenge of adapting pre-trained models to specific tasks and domains. It provides a unified interface for training, inference, evaluation, quantization, and deployment workflows. The framework integrates multiple training methodologies including supervised fine-tuning, direct preference optimization, reinforcement learning approaches like GRPO and its variants, and specialized tasks such as embedding and reranking. It incorporates distributed training techniques including tensor parallelism, pipeline parallelism, and context parallelism to accelerate training on large models. For inference acceleration, it integrates vLLM, SGLang, and LMDeploy, and supports quantization through GPTQ, AWQ, BNB, and FP8 methods.

Teams should adopt this tool if they need to fine-tune or deploy models from the supported ecosystem of text-based and multimodal architectures. It suits projects requiring flexibility across multiple training paradigms, from standard supervised fine-tuning to advanced reinforcement learning approaches. The framework is particularly valuable for teams working with recent model releases, as it maintains support for newly released architectures across both language and vision-language domains.

The maintainers respond to issues and pull requests within a day. Work in the issue tracker centers on bug fixes, feature enhancements, and resolved items, indicating active maintenance focused on stability and incremental capability expansion.