meta-pytorch/torchtune

Health v2 · maintenance only

Adoption not part of the score

5812 stars · 757 forks observed · 2026-09-21

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Torchtune is a PyTorch-native library for authoring, post-training, and experimenting with large language models. It provides hackable training recipes (SFT, DPO, PPO, GRPO, distillation, QAT), simple implementations of popular LLMs, and YAML-based configuration for training, evaluation, quantization, and inference.

Use cases

• fine-tune llama on my own dataset
• run LoRA or QLoRA finetuning of an open LLM
• train a model with DPO or PPO preference optimization
• distill a large language model into a smaller one
• experiment with LLM training recipes in pure PyTorch
• finetune a multimodal vision-language model
• quantization-aware training for LLMs

When to choose

• you want hackable, PyTorch-native fine-tuning code rather than a black-box trainer
• you need memory-efficient single- or multi-node finetuning of Llama, Qwen, Gemma, Mistral, or Phi models
• you want YAML-configured recipes for SFT, DPO, PPO, GRPO, or distillation

When to avoid

• you need a library under active development - torchtune wound down in 2025 and is no longer actively maintained
• you need point-and-click GUI-based fine-tuning
• you need inference-only serving rather than post-training

Facets

library · maturity maintenance

llm-training machine-learning deep-learning cli configuration-management large-language-models machine-learning deep-learning artificial-intelligence python cli cross-platform fine-tuning lora qlora sft dpo ppo grpo knowledge-distillation pytorch llama qwen gemma mistral multimodal yaml-configs wound-down gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="meta-pytorch/torchtune")