lightning-tune 2026.5.1

A production-ready library for automated, advanced multi-modal LLM finetuning and deployment.

⚡ Lightning Tune: A Production-Ready Finetuning Library

Lightning Tune is a powerful, production-grade Python library for finetuning and deploying both text-only and multi-modal Large Language Models. Built entirely on the Lightning AI ecosystem (PyTorch Lightning, LitGPT, LitServe), it provides a seamless, automated, and efficient workflow from raw data to a deployed API.

Features

• Smart Configuration: Automatically analyze your dataset's schema and size to determine a strong baseline configuration for finetuning.
• Smart Hyperparameter Tuning: Infers sensible starting points for Learning Rate, Number of Epochs, and LoRA Rank based on your dataset size.
• Flexible Multi-Modal Architecture: Define multi-modal towers explicitly in your code or let the smart factory detect them for you.
• Advanced PEFT Methods: Natively supports QLoRA, LoRA, and DoRA for text-only finetuning via Hugging Face's SFTTrainer.
• End-to-End Workflow: A unified API takes you from training to a production-ready inference server.
• Interactive UI: A Gradio-based web interface for a no-code, point-and-click finetuning experience.

Installation

# Clone the repository (or create the project from the provided code)
cd lightning-tune

# Create and activate a virtual environment
uv venv
source .venv/bin/activate

# Install the library and its testing dependencies in editable mode
uv pip install -e ".[test]"

How to Use

1. The Smart Factory (Recommended)

Let the library analyze your dataset and choose the best starting configuration.

# examples/run_smart_multimodal.py
import lightning_tune as lt
from pathlib import Path

# The model repo_id is a mandatory argument
config = lt.PipelineConfig.from_dataset(
    model_repo_id="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    file_path=Path("examples/sample_data_multimodal.csv"),
    image_root_path=Path("examples/")
)

# You can still override any smart setting
config.trainer.max_epochs = 2

trained_model_path = lt.run_finetuning(config)
if trained_model_path.exists():
    lt.launch_server(config, trained_model_path)

2. Manual Configuration (Full Control)

Explicitly define every aspect of your pipeline, including multi-modal towers.

# examples/run_manual_multimodal.py
import lightning_tune as lt
from pathlib import Path

config = lt.PipelineConfig(
    model=lt.ModelConfig(repo_id="TinyLlama/TinyLlama-1.1B-Chat-v1.0"),
    data=lt.DataConfig(
        file_path=Path("examples/sample_data_multimodal.csv"),
        image_root_path=Path("examples/"),
        text_columns=["product_description"],
        output_column="customer_review",
        vision_config=lt.VisionConfig(
            image_column="image_path",
            projection_dim=2048 # Manually set to match TinyLlama's hidden size
        ),
        tabular_config=lt.TabularConfig(
            numerical_columns=["price"],
            categorical_columns=["category"],
            projection_dim=2048
        )
    ),
    trainer=lt.TrainerConfig(max_epochs=2),
    train=lt.TrainConfig(batch_size=1)
)
lt.run_finetuning(config)

3. Using the Gradio Web UI

For a seamless, no-code experience, launch the interactive Gradio application.

python app/main.py

4. Running Tests

pytest