TorchTune

Information

TorchTune is a PyTorch-native solution for fine-tuning and experimenting with modern LLMs, especially in ecosystems close to Meta model workflows. It supports methods such as DoRA and PPO for RLHF-related training and includes optimization paths such as PyTorch 2.5 compilation, which can provide roughly 20-24% speed improvements in the right scenarios.

Common use cases

• custom PyTorch-based fine-tuning,
• Meta-model oriented experimentation,
• RLHF and adapter-training workflows,
• and teams that prefer writing or extending their own training code instead of relying only on higher-level wrappers.

Practical note

TorchTune is a good fit for PyTorch developers who want more direct control over training code while still using a purpose-built LLM training toolkit.

See also

• LLM
• Hugging Face PEFT / TRL
• LitGPT