SWIFT (Scalable lightWeight Infrastructure for Fine-Tuning) is a comprehensive framework for fine-tuning and deploying over 450 large language models (LLMs) and 150+ multi-modal large models (MLLMs). It targets researchers and developers working with LLMs and MLLMs, offering a unified pipeline from pre-training and fine-tuning to inference, evaluation, and deployment, significantly streamlining the MLOps lifecycle for these complex models.

How It Works

SWIFT supports a wide array of training techniques, including parameter-efficient fine-tuning (PEFT) methods like LoRA, QLoRA, DoRA, and GaLore, alongside full-parameter fine-tuning. It also integrates advanced human alignment methods such as DPO, GRPO, KTO, and ORPO. For inference and deployment, SWIFT leverages acceleration engines like vLLM and LMDeploy, and supports quantization via GPTQ and AWQ. The framework is designed for flexibility, allowing custom model and dataset integration, as well as component customization (loss, metrics, optimizers).

Quick Start & Requirements

• Installation: pip install ms-swift -U or from source.
• Python: >=3.9 (3.10 recommended).
• Dependencies: PyTorch >=2.0, Transformers >=4.33, ModelScope >=1.19. Specific versions for DeepSpeed, vLLM, LMDeploy, and EvalScope are recommended for optimal performance. CUDA 12 is recommended for GPU acceleration.
• Documentation: English Documentation

Highlighted Details

• Supports over 150 multi-modal models and datasets, covering vision, audio, and video modalities.
• Integrates 10+ human alignment algorithms (DPO, GRPO, KTO, etc.) for both LLMs and MLLMs.
• Offers a Gradio-based Web UI for zero-threshold training and deployment.
• Provides acceleration for inference, evaluation, and deployment using vLLM and LMDeploy.

Maintenance & Community

• Active development with frequent updates and new feature additions (e.g., GRPO, Megatron support).
• Community support available via Discord.
• Paper accepted at AAAI 2025.

Licensing & Compatibility

• Licensed under the Apache License (Version 2.0). Model and dataset licenses are subject to their original sources.
• Compatible with commercial use, provided original model/dataset licenses are respected.

Limitations & Caveats

• While supporting many hardware types, optimal performance may require specific GPU configurations (e.g., A100/H100 for large-scale training).
• The extensive feature set implies a complex dependency graph; users should carefully manage installations for specific use cases.