LitGPT
Lightning AILitGPT is a library of 20+ from-scratch implementations of open LLMs, including Llama, Mistral, Gemma, Phi, and Qwen, with recipes for pretraining, fine-tuning (full, LoRA, QLoRA, Adapter/Adapter v2), and deployment. Each model is a single file with no abstraction layer, extending nanoGPT and Lit-LLaMA. Lightning AI builds it on Lightning Fabric, with Flash Attention, FSDP across 1 to 1,000+ GPUs or TPUs, and 4/8-bit quantization.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public
- maintainer
- Lightning-AI
- core features withheld
- no
LitGPT is released under Apache-2.0 with the full source public. Its README describes the library as unrestricted for enterprise use and documents no paid tier or feature held back from it.
- https://raw.githubusercontent.com/Lightning-AI/litgpt/main/LICENSE.md recorded 2026-09-28
The file is the full Apache License, Version 2.0 with the appendix completed as "Copyright [2023] Lightning AI" and "Licensed under the Apache License, Version 2.0 (the "License");". No other license appears in it.
- https://raw.githubusercontent.com/Lightning-AI/litgpt/main/README.md recorded 2026-09-28
The README gives the public clone command "git clone https://github.com/Lightning-AI/litgpt" and states "LitGPT is released under the [Apache 2.0]". It separately advertises Lightning AI's hosted platform ("Over 340,000 developers use [Lightning Cloud]" with "[GPUs]" priced "from $0.19") and optional "Run on Studios" buttons, but every install, finetune, pretrain and serve command runs from the open package; a search for license-key, edition, premium and subscription language found none, so no library feature is gated.
Adoption
2 medium confidenceAdoption is measured by PyPI downloads of litgpt, the project's only published package.
- https://pypistats.org/api/packages/litgpt/recent recorded 2026-08-13
15,680 downloads in the trailing 30 days for litgpt
Capability
3 medium confidenceLitGPT offers pretraining and fine-tuning recipes, including full fine-tuning, LoRA, QLoRA and Adapter methods, over readable from-scratch model implementations, with FSDP and Flash Attention for multi-GPU training. It documents no preference or reinforcement-learning methods such as DPO, PPO or GRPO, which leaves it behind post-training libraries such as TRL and Axolotl.
- https://raw.githubusercontent.com/Lightning-AI/litgpt/main/README.md recorded 2026-09-28
The README lists "Parameter-efficient finetuning: [LoRA](tutorials/finetune_lora.md), [QLoRA](tutorials/finetune_lora.md), [Adapter](tutorials/finetune_adapter.md), and [Adapter v2](tutorials/finetune_adapter.md)." plus "Flash Attention v2, multi-GPU support via fully-sharded data parallelism" and the banner "1-1000+ GPUs/TPUs". Searched the body for RLHF, DPO, PPO, GRPO and preference: only substring hits inside unrelated words (support), no preference post-training method.
- https://raw.githubusercontent.com/Lightning-AI/litgpt/main/tutorials/finetune_lora.md recorded 2026-09-28
The LoRA tutorial is titled "Finetuning with LoRA / QLoRA" and describes 4-bit QLoRA-style quantization enabled with a flag.
- https://raw.githubusercontent.com/Lightning-AI/litgpt/main/tutorials/pretrain.md recorded 2026-09-28
The pretraining tutorial states "This document explains how to pretrain LLMs using LitGPT." Searched the body for RLHF, DPO, PPO and GRPO: only substring hits inside unrelated words.
Verified 2026-09-24