torchtune
PyTorch-native post-training library for LLMs, no longer actively maintained since 2025; code still installs, the team points to torchforge
Overview
torchtune is the PyTorch team's "PyTorch native post-training library" for LLMs: hackable recipes for supervised fine-tuning, knowledge distillation, DPO, PPO, GRPO and quantization-aware training, with LoRA and QLoRA, from one GPU up to multi-node clusters, for Llama 3.x and 4, Gemma, Mistral, Phi and Qwen models. It is no longer under active development. Its README now opens with "Torchtune is no longer actively maintained: torchtune development wound down in 2025," and the maintainers' July 15, 2025 post "The future of torchtune" said they were "stopping active development on torchtune, effective immediately," with only critical fixes through 2025. The last release is v0.6.1 (April 7, 2025). The repository is not archived, so the BSD-3-Clause code still installs and runs, but it gets no new models or features. In the same thread a maintainer pointed users to torchforge, a newer, experimental PyTorch library for agentic RL post-training. See also: PyTorch.
The Verdict
Who Should Use torchtune?
Best For
- Existing users with working recipes on supported models
- Learning how post-training recipes are built in plain PyTorch
Not Ideal For
- New projects: pick an actively maintained toolkit
- Models released after early 2025 (no new model support)
- Anyone who needs security fixes after 2025
What's Great
- Pure PyTorch, few abstractions, easy to read and modify
- Broad recipe set: SFT, distillation, DPO, PPO, GRPO, QAT
- Scales from one device to multi-node
Watch Out For
- Active development stopped in July 2025
- No release since v0.6.1 (April 2025)
- The suggested torchforge is experimental and RL-focused, not a drop-in swap
Pricing
View all features & details
Key Features
- SFT, knowledge distillation, DPO, PPO, GRPO and QAT recipes
- LoRA and QLoRA
- Single-device, multi-device and multi-node training
- Models: Llama 3.x and 4, Gemma and Gemma 2, Mistral, Phi 3 and 4, Qwen 2, 2.5 and 3
Timeline
- v0.6.0 (Mar 2025): tensor parallel, multi-node SLURM, distributed DPO
- v0.6.1 (Apr 7, 2025): last release
- Jul 15, 2025: active development stopped
- Oct 2025: maintainers point to torchforge
Real-World Usage
Status
- Repository open, not archived; issues and Discord stay open for support
- No new features since mid-2025
Where Users Went
- torchforge for RL post-training (experimental)
- Actively maintained toolkits such as TRL, Axolotl and Unsloth
How It Compares
| Feature | torchtune | TRL | Axolotl |
|---|---|---|---|
| Status | Unmaintained since 2025 | Active, v1.17 (Oct 2026) | Active, v0.21 (Oct 2026) |
| Style | Plain PyTorch recipes | Python trainers on Transformers | YAML config |
| RL methods | DPO, PPO, GRPO | GRPO, RLOO, DPO, KTO | DPO, KTO, ORPO, GRPO |
| License | BSD-3-Clause | Apache-2.0 | Apache-2.0 |
Sources: GitHub · The future of torchtune (Jul 2025) · Releases · Introducing torchforge
Is this your tool? Add a badge