LitGPT

LitGPT is a Python toolkit for developers and researchers who want to train and deploy language models on their own hardware. Its directly editable model implementations suit custom applications and research experiments.

• Training and fine-tuning: Train models on your own datasets or adapt existing models with LoRA, QLoRA and adapters, using supplied training recipes.
• Model choice: Work with Llama, Gemma, Phi 4 and Qwen2.5, including coding models such as Code Llama and Qwen2.5 Coder.
• Memory and scaling: Quantization, lower-precision computation and optional CPU offloading reduce GPU memory demands, while distributed training supports multiple GPUs and TPUs.
• Evaluation and deployment: Evaluate models, deploy them for inference and export weights to other model formats.

LitGPT is open source under Apache 2.0. It supports workloads on a single GPU as well as GPU clusters. Lightning Cloud separately provides hosted GPU workspaces, training clusters and inference APIs; those workloads run on cloud infrastructure.