Training on a Google Cloud TPU instance

Welcome to the 🤗 Optimum-TPU training guide! This section covers how to fine-tune models using Google Cloud TPUs.

Supported Models

See Supported Models.

Getting Started

Prerequisites

Before starting the training process, ensure you have:

  1. A configured Google Cloud TPU instance (see Deployment Guide)
  2. Optimum-TPU installed with PyTorch/XLA support:
pip install optimum-tpu -f https://storage.googleapis.com/libtpu-releases/index.html
export PJRT_DEVICE=TPU

Example Training Scripts

You can now follow one of our several example scripts to get started:

  1. Gemma Fine-tuning:

  2. LLaMA Fine-tuning:

Tips

We recommend fine-tuning using bf16 on TPU, as those operations are often extremely fast while keeping the training stable and precise enough.