{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/6"}],"enrichment":{"capability":"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.","skillfed_tags":["llm-training","model-finetuning","pytorch-native"],"use_cases":["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."],"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.\n\nThe 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.","worth_installing":"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."},"id":"torchtune","links":{"html":"https://skillfed.io/packages/torchtune","md":"https://skillfed.io/packages/torchtune.md","pypi":"https://pypi.org/project/torchtune/"},"maintenance":{"status":"active"},"meta":{"latest_release":"2025-04-07","license_spdx":null,"license_treatment":"permissive","name":"torchtune","python_support":"supports_current","summary":"A native-PyTorch library for LLM fine-tuning"},"popularity":{"monthly_downloads":329013,"position":7548,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.6.1"}