OSS Alternative - Discover Top Open Source Alternatives to Popular Software

modelscope/ms-swift

A comprehensive framework from ModelScope for efficiently fine-tuning, evaluating, and deploying over 1000 large language models and multimodal large models using advanced techniques.

Core Features

Extensive Model Support: Supports 600+ text-only LLMs and 400+ multimodal LLMs, with Day-0 support for new models.

Advanced Training Algorithms: Integrates PEFT methods (LoRA, QLoRA), preference learning (DPO, KTO), and reinforcement learning (GRPO, DAPO) for enhanced model intelligence.

Full-Pipeline Capabilities: Provides end-to-end support for training, inference, evaluation, quantization, and deployment of large models.

Optimized Performance: Features memory optimization (Flash-Attention, Ulysses, Ring-Attention) and distributed training (DeepSpeed, FSDP, Megatron) for scalability.

Broad Hardware Compatibility: Supports various GPUs (A10/A100/H100, RTX, T4/V100), CPUs, MPS, and domestic Ascend NPUs.

Detailed Introduction

ms-swift is a robust and scalable framework developed by the ModelScope community, designed to streamline the entire lifecycle of large language models (LLMs) and multimodal large models (MLLMs). It offers unparalleled support for a vast array of models and integrates cutting-edge training methodologies, including efficient fine-tuning, preference learning, and reinforcement learning algorithms. With comprehensive features for inference, evaluation, quantization, and deployment, ms-swift empowers developers to build, optimize, and deploy advanced AI models across diverse hardware environments efficiently and effectively.