LitGPT

license

Apache-2.0(OSI)

source

public

maintainer

Lightning-AI

core-gated

ungated

Apache-2.0, full source public. The library's own README describes itself as unrestricted for enterprise use; Lightning AI's separate Studios platform is hosted compute, not a gated core.

• https://github.com/Lightning-AI/litgpt recorded 2026-08-09

  The repository's top-level file listing (16 items) is .devcontainer, .github, .lightning/workflows, config_hub, extensions, litgpt, tests, tutorials plus .gitignore, .pre-commit-config.yaml, CITATION.cff, CODE_OF_CONDUCT.md, CONTRIBUTING.md, LICENSE.md, README.md, and pyproject.toml, with no enterprise, commercial, or paid-tier directory. The repo's own classifier reads spdxId "Apache-2.0". The rendered README's feature checklist states "Enterprise ready - Apache 2.0 for unlimited enterprise use." and "Developer friendly - Easy debugging with no abstraction layers and single file implementations."

• https://raw.githubusercontent.com/Lightning-AI/litgpt/main/LICENSE.md recorded 2026-08-09

  The fetched body is the full Apache License, Version 2.0 (the standard TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION text through the appendix), with the boilerplate appendix completed as "Copyright [2023] Lightning AI" and "Licensed under the Apache License, Version 2.0 (the "License")". No other license or third-party carve-out appears in the file.

• https://lightning.ai/pricing recorded 2026-08-09

  The fetched HTML is a client-rendered SPA shell (a boot splash and a bare div#root) with no server-rendered pricing tiers. The page's meta description reads "The all-in-one platform for AI development. Code together. Prototype. Train. Scale. Serve. From your browser - with zero setup.", describing Lightning AI's hosted compute platform rather than any LitGPT-specific feature. Corroborating only; it does not itself settle core-gated.

15,680 PyPI downloads in the trailing 30 days for litgpt, the project's only published package, which puts reach in the 10K-100K band.

• https://pypistats.org/api/packages/litgpt/recent recorded 2026-08-13

  15,680 downloads in the trailing 30 days for litgpt

Solid PEFT and pretraining recipes with a clean from-scratch implementation, but no RLHF/preference post-training methods (no PPO/GRPO/DPO) and oriented to readable single-model recipes rather than frontier-scale demonstration. Mid-tier within the category on demonstrated method breadth.

• https://github.com/Lightning-AI/litgpt recorded 2026-08-09

  The 'State-of-the-art features' checklist states "State-of-the-art optimizations: Flash Attention v2, multi-GPU support via fully-sharded data parallelism, optional CPU offloading, and TPU and XLA support.", "Reduce compute requirements with low-precision settings: FP16, BF16, and FP16/FP32 mixed.", "Lower memory requirements with quantization: 4-bit floats, 8-bit integers, and double quantization.", and "Parameter-efficient finetuning: LoRA, QLoRA, Adapter, and Adapter v2." The ASCII banner separately reads "Reduce GPU memory (fp4/8/16/32)" against "1-1000+ GPUs/TPUs". A case-insensitive search of the fetched body for PPO, DPO, and GRPO returns no genuine hits (the only 'PPO' match is the substring inside the word 'SUPPORT'), confirming no RLHF/preference post-training methods are documented.

• https://github.com/Lightning-AI/litgpt/tree/main/tutorials recorded 2026-08-09

  The tutorials directory listing includes finetune_full.md, finetune_lora.md, finetune_adapter.md, pretrain.md, pretrain_tinyllama.md, deploy.md, inference.md, quantize.md, and evaluation.md, so full, LoRA, QLoRA, and Adapter finetuning, plus pretraining and deployment, are each backed by a dedicated tutorial rather than only the README checklist's claim.