TorchTune
PyTorch library of post-training recipes for LLMs, including SFT, DPO, PPO and GRPO. No longer maintained.
How the 6.4 was reached
- Quality of output6/10
5,810 stars, but development wound down in 2025 and the library is no longer maintained; 467 open issues. github.com
- Ease of use4/10
Library is no longer actively maintained, and no docs page was found. github.com
- Pricing value10/10
Free under BSD-3-Clause with no paid dependency. github.com
- API and integration quality8/10
Integrates with Hugging Face Hub, Weights & Biases, LM Eval Harness, ExecuTorch and torchao; documented overview. meta-pytorch.org
- Solves the problem it claims to solve4/10
Delivers post-training recipes, but the project is wound down and no longer maintained, which undercuts the ongoing authoring and deployment claim. github.com
6.4/10 is the mean of the 5 criteria that apply. Scored on the five-criteria rubric v1, 29 September 2026. How scoring works
Pricing
- BSD-3-Clause license
- No paid dependency
Key Features
- Recipes for SFT, knowledge distillation, DPO, PPO, GRPO and quantization-aware training
- Supports Llama, Gemma, Mistral, Phi and Qwen models
- Multi-node training supported as of February 2025
- Activation offloading and multimodal QLoRA in v0.4.0
- Integrates with Hugging Face Hub, Weights & Biases, LM Eval Harness, ExecuTorch and torchao
Pros & Cons
Pros
- Free under BSD-3-Clause
- Wide set of post-training recipes
- Integrates with common ML tools
- 5,810 GitHub stars show real interest
Cons
- Development wound down in 2025 and it is no longer maintained
- 467 open issues
- No docs page was found
- New model support is unlikely to arrive
Useful as a readable reference for post-training recipes, or for pinned legacy work. For new projects, pick a library that is still maintained.
Try TorchTune →Added to scored.tools on
Competitors to TorchTune
Other tools in the coding category worth comparing.
