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

In the rapidly evolving landscape of large language models (LLMs), pre-training is only the first step. To transform a base model into a specialized assistant or a high-performing reasoning engine, post-training is essential. Today, we are excited to announce new features in MaxText that streamline this process: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) now available on single-host TPU configurations (such as v5p-8 and v6e-8).

By leveraging the power of JAX and the efficiency of the Tunix library, MaxText provides a high-performance, scalable path for developers to refine their models using the latest post-training techniques. You can explore the full documentation for SFT and RL to start your post-training journey on TPUs today.

Supervised Fine-Tuning (SFT): Precision Tuning Made Simple

Supervised Fine-Tuning is the primary method for adapting a pre-trained model to follow specific instructions or excel at niche tasks. With the new single-host SFT support, users can now take an existing MaxText or Hugging Face checkpoint and fine-tune it on labeled datasets with minimal setup.

Key Highlights:

• Seamless Integration: Native support for Hugging Face datasets (e.g., ultrachat_200k).
• Flexible Checkpoints: Use existing MaxText checkpoints or convert Hugging Face models (like Gemma 3) directly within the ecosystem.
• Optimized Execution: Powered by Tunix, a JAX-based library specifically designed for post-training efficiency.

Reinforcement Learning (RL): Advancing Reasoning Capabilities

For tasks requiring complex logic and reasoning—such as math or coding—Reinforcement Learning is a game-changer. MaxText now supports several state-of-the-art RL algorithms on single-host TPUs, utilizing vLLM for high-throughput inference during the training loop. For example,

1. Group Relative Policy Optimization (GRPO) GRPO is a memory-efficient variant of PPO (Proximal Policy Optimization). It eliminates the need for a separate value function model, instead generating multiple responses per prompt and calculating relative advantages within the group. This significantly reduces the hardware footprint, making advanced RL accessible on a single TPU host.
2. Group Sequence Policy Optimization (GSPO) GSPO focuses on sequence-level importance ratios and clipping. It improves training stability and efficiency by rewarding model behavior at the sequence level, making it particularly effective for enhancing performance on benchmarks like GSM8K.

Getting Started

To begin using these new features, ensure you have the latest post-training dependencies installed:

Running SFT:

You can launch an SFT run using the train_sft module, specifying your model, dataset, and output directory:

Running RL (GRPO/GSPO):

For RL, the train_rl module handles the loading of policy and reference models, executes the training, and provides automated evaluation on reasoning benchmarks:

What’s Next?

While single-host support provides a powerful entry point for many developers, MaxText is built for scale. These same workflows are designed to transition seamlessly to multi-host configurations for those training larger models and utilizing massive datasets. Please stay tuned for more updates in this direction from us in the future.