MaxText Expands Post-Training Capabilities: Introducing SFT and RL on Single-Host TPUs- Google Developers Blog

MaxText now supports Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on single-host TPU configurations (e.g., v5p-8, v6e-8). SFT allows fine-tuning on labeled datasets using Hugging Face datasets and checkpoints. RL support includes GRPO (Group Relative Policy Optimization) and GSPO (Group Sequence Policy Optimization), leveraging vLLM for inference. Installation instructions and example commands for running SFT and RL are provided.