torchtune

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 on this page — 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

torchtune on PyPI

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

Evidence: torchtune-0.6.1-py3-none-any.whl

Keywords: pytorch, finetuning, llm

Tags

More Artificial Intelligence packages