torchtune

Overview

torchtune is a PyTorch-native library for post-training tasks such as fine-tuning and adaptation. It provides composable recipes and modular components built directly on PyTorch, focusing on memory efficiency and ease of experimentation.

Best for

Best for
Developers already working in PyTorch who need a lightweight, modular library for fine-tuning and adapting large models.

Use cases

• Fine-tuning large language models with memory-efficient techniques like LoRA and QLoRA
• Adapting pretrained generative models for custom instruction-following or domain-specific tasks
• Reproducing and modifying standardized post-training recipes from the open-source community

Notes

torchtune is a PyTorch-native library for post-training tasks such as fine-tuning and adaptation. It provides composable recipes and modular components built directly on PyTorch, focusing on memory efficiency and ease of experimentation.

5,763 stars on GitHub. Last updated 2026-06-01. Licensed BSD-3-Clause.

Use cases

• Fine-tuning large language models with memory-efficient techniques like LoRA and QLoRA
• Adapting pretrained generative models for custom instruction-following or domain-specific tasks
• Reproducing and modifying standardized post-training recipes from the open-source community

Pros

• Native PyTorch integration with no extra dependencies, easing debugging and customisation
• Actively maintained under the PyTorch ecosystem with a growing set of community recipes
• Optimised for memory-constrained environments, enabling fine-tuning on consumer hardware

Cons

• Limited to post-training workflows, not a full training or inference framework
• Smaller ecosystem and fewer pre-built recipes compared to more established libraries like Hugging Face Transformers
• Requires familiarity with PyTorch and recent fine-tuning techniques to use effectively

Indexed from awesome-llm and enriched against its public facts.