Introducing Tunix: A JAX-Native Library for LLM Post-Training- Google Developers Blog

Introduces Tunix, a new open-source, JAX-native library for LLM post-training. Tunix provides trainers for Supervised Fine-Tuning (SFT) with LoRA/QLoRA support, Direct Preference Optimization (DPO), Proximal Policy Optimization (PPO), Group Relative Policy Optimization (GRPO), Group Sequence Policy Optimization (GSPO-token), and knowledge distillation (logit-based and attention transfer). It is designed for performance on TPUs, integrates with MaxText, and offers a customizable 'white-box' design. Quantitative results demonstrate ~12% relative improvement in pass@1 accuracy on GSM8K using Gemma 2 2B-IT with GRPO.