ModelScope Swift: a lightweight infrastructure for efficiently fine-tuning and deploying large models.

ModelScope Swift (MS-Swift for short) is an efficient lightweight infrastructure designed for fine-tuning, reasoning, evaluation and deployment of Large LLMs (LLMs) and Multimodal Large Models (MLLMs). The framework supports more than 400 LLMs and 100+ MLLMs, providing a complete workflow from model training, evaluation to application.MS-Swift not only supports PEFT (Parameter Efficient Fine-Tuning) technology, but also provides a rich library of adapters to support the latest training techniques, such as NEFTune, LoRA+, LLaMA-PRO, and so on. For users unfamiliar with deep learning, MS-Swift also provides a Gradio-based web interface for easy control of training and inference.

Function List

• Supports training, inference, evaluation and deployment of 350+ LLMs and 100+ MLLMs
• Provides adapter libraries for the latest training technologies such as PEFT, LoRA+, LLaMA-PRO and more!
• Gradio-based web interface for easy control of training and inference
• Supports multi-GPU training and deployment
• Provides detailed documentation and deep learning courses
• Supports a wide range of hardware environments, including CPUs, RTX series graphics cards, A10/A100 and other computing cards
• Supports a variety of training methods, such as full-parameter fine-tuning, LoRA fine-tuning, quantization training, etc.
• Provide support for multiple datasets and models for different training tasks

Using Help

Installation process

MS-Swift can be installed in the following three ways:

1. Use the pip command to install:

   # 安装所有功能
   pip install 'ms-swift[all]' -U
   # 仅安装LLM相关功能
   pip install 'ms-swift[llm]' -U
   # 仅安装AIGC相关功能
   pip install 'ms-swift[aigc]' -U
   # 仅安装适配器相关功能
   pip install ms-swift -U
   

2. Installation via source code:

   git clone https://github.com/modelscope/swift.git
   cd swift
   pip install -e '.[llm]'
   

3. Install using a Docker image.

Using the Web Interface

MS-Swift provides a Gradio-based web interface that users can launch with the following command:

SWIFT_UI_LANG=en swift web-ui

The web interface supports multi-GPU training and deployment, and users can easily control the training and inference process.

Training and reasoning

MS-Swift supports a variety of training and inference methods, here are some sample commands:

• Single GPU training:

  CUDA_VISIBLE_DEVICES=0 swift sft --model_type qwen1half-7b-chat --dataset blossom-math-zh --num_train_epochs 5 --sft_type lora --output_dir output --eval_steps 200
  

• Multi-GPU training:

  NPROC_PER_NODE=4 CUDA_VISIBLE_DEVICES=0,1,2,3 swift sft --model_type qwen1half-7b-chat --dataset blossom-math-zh --num_train_epochs 5 --sft_type lora --output_dir output
  

• Reasoning:

  CUDA_VISIBLE_DEVICES=0 swift infer --model_type qwen1half-7b-chat
  

Detailed Documentation

MS-Swift provides extensive documentation and deep learning courses, and users can visit the following links for more information:

• English document
• Chinese document