Introduction to Plugins
Overview
Transformer Lab's architecture is built around a plugin system that powers all core functionality. Plugins enable:
- Model training
- Inference
- Evaluation
- etc.
Plugins allow anyone to extend the ability of Transformer Lab in a consistent way.
Plugin Structure
At its core, a Transformer Lab plugin is a directory containing specific files that define its functionality, parameters, and dependencies.
Example: Apple MLX LoRA Trainer
Let's examine a real-world example: the MLX LoRA Trainer plugin which enables efficient fine-tuning on Apple Silicon.
transformerlab/plugins/mlx_lora_trainer/
├── index.json
├── info.md
├── main.py
└── setup.sh
Key Files
Each plugin consists of four essential files:
| File | Purpose |
|---|---|
| index.json | Defines plugin metadata including name, version, and required parameters |
| info.md | Provides user documentation explaining the plugin's functionality and usage |
| main.py | Contains the core implementation that executes when the plugin runs |
| setup.sh | Installs dependencies and performs one-time setup operations |
Getting Started
The plugin system makes extending Transformer Lab straightforward. In the following sections, we'll explore how to:
- Creating a new plugin
- Formatting the parameters that a plugin can use
- Using the Plugin SDK to convert an existing script to work inside Transformer Lab