---
id: torchtune
version: "0.6.1"
license: BSD 3-Clause License Copyright 2024 Meta Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1.… (full text in the JSON record)
license_treatment: permissive
maintenance: active
---

# torchtune — A native-PyTorch library for LLM fine-tuning

License: permissive · Maintenance: active · Downloads: 329.0K/mo

## What it is and what it does

torchtune is a native-PyTorch library that provides end-to-end support for fine-tuning and post-training large language models. It bundles PyTorch implementations of popular models (Llama, Gemma, Mistral, Phi, Qwen) with YAML-configurable training recipes covering the full post-training lifecycle: supervised fine-tuning, knowledge distillation, reinforcement learning methods (DPO, PPO, GRPO), and quantization-aware training. It supports single-device, multi-device, and multi-node training with various weight-update strategies including full fine-tuning and LoRA/QLoRA.

The library targets researchers and practitioners who need to adapt pre-trained models to specific tasks or domains. It abstracts away low-level PyTorch complexity while exposing configuration through YAML, allowing users to experiment with different training methods and model sizes without rewriting training loops. Dependencies include data-loading libraries (torchdata, datasets, huggingface_hub), tokenizers (sentencepiece, tiktoken, tokenizers), and utility packages (omegaconf, psutil, Pillow) to support model loading, preprocessing, and monitoring.

Use it for:

- Fine-tune models on custom datasets using LoRA to reduce memory requirements on a single GPU.
- Perform knowledge distillation to compress a large teacher model into a smaller student model using LoRA on distributed hardware.
- Apply DPO to align model outputs with human preferences using multi-device training.
- Run quantization-aware training on distributed nodes to produce efficient inference-ready models.
- Experiment with different post-training methods via config swaps without rewriting training code.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

torchtune is a PyTorch library for fine-tuning, post-training, and experimenting with large language models using recipes for SFT, knowledge distillation, DPO, PPO, GRPO, and quantization-aware training.

Yes. torchtune is actively maintained (latest release April 2025), has no known vulnerabilities, low install friction, and a permissive BSD license. It is purpose-built for LLM fine-tuning with broad recipe coverage and model support. Install if you need to fine-tune or post-train language models; skip if you only need inference or work exclusively with models outside its supported set.

## Install

pip install torchtune
uv add torchtune
poetry add torchtune

## Installing torchtune

Before you install: Low install friction; pure Python wheel with 14 runtime dependencies including PyTorch ecosystem packages and tokenization libraries. Active maintenance with latest release in April 2025 and 5797 GitHub stars.

License in practice: BSD 3-Clause permissive license allows commercial and private use with attribution and liability disclaimer; no restrictions on derivative works or redistribution.

Quickstart:

pip install torchtune

from torchtune.models.llama3_2 import llama3_2

# Configure and run a training recipe via CLI:
# tune run lora_finetune_single_device --config llama3_2/3B_lora_single_device

Requires Python >=3.9 and PyTorch with CUDA support for GPU training; model weights must be downloaded from huggingface_hub or kagglehub.

Verify before relying:

- Whether all 14 runtime dependencies are required for basic usage or if subsets suffice for specific recipes.
- Specific hardware requirements or CUDA version constraints not documented in the fact sheet.
- Production-readiness and performance characteristics of multi-node training beyond the February 2025 announcement.

## Package facts

- License: BSD 3-Clause License Copyright 2024 Meta Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1.… (full text in the JSON record) (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 329.0K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

llm fine-tuning pytorch, language model post-training, lora qat dpo training, supervised finetuning recipes, rlhf dpo ppo training, quantization aware training, model adaptation tuning, llm-training, model-finetuning, pytorch-native

[View on SkillFed](https://skillfed.io/packages/torchtune) · [View on PyPI](https://pypi.org/project/torchtune/)