MaxText expands post-training capabilities: SFT and RL on single-host TPUs

AI Impact Summary

MaxText has expanded its capabilities by introducing SFT and RL support directly on single-host TPUs, leveraging JAX and Tunix. This allows developers to efficiently adapt pre-trained models for specialized tasks and complex reasoning, particularly utilizing GRPO and GSPO algorithms. The streamlined workflow and scalability options open up opportunities for rapid experimentation and model refinement, especially for developers working with models like Gemma 3.