Tensor Processing Units (TPU) are AI accelerator made by Google to optimize performance and cost from AI training to inference.

This repository exposes an interface similar to what Hugging Face transformers library provides to interact with a magnitude of models developed by research labs, institutions and the community.

We aim at providing our user the best possible performances targeting Google Cloud TPUs for both training and inference working closely with Google and Google Cloud to make this a reality.

We currently support a few LLM models targeting text generation scenarios:

• 💎 Gemma (2b, 7b)
• 🦙 Llama2 (7b) and Llama3 (8b). On Text Generation Inference with Jetstream Pytorch, also Llama3.1, Llama3.2 and Llama3.3 (text-only models) are supported, up to 70B parameters.
• 💨 Mistral (7b)

optimum-tpu comes with an handy PyPi released package compatible with your classical python dependency management tool.

pip install optimum-tpu -f https://storage.googleapis.com/libtpu-releases/index.html

export PJRT_DEVICE=TPU

optimum-tpu provides a set of dedicated tools and integrations in order to leverage Cloud TPUs for inference, especially on the latest TPU version v5e and v6e.

Other TPU versions will be supported along the way.

As part of the integration, we do support a text-generation-inference (TGI) backend allowing to deploy and serve incoming HTTP requests and execute them on Cloud TPUs.

Please see the TGI specific documentation on how to get started.

optimum-tpu provides an optional support of JetStream Pytorch engine inside of TGI. This support can be installed using the dedicated CLI command:

optimum-tpu install-jetstream-pytorch

To enable the support, export the environment variable JETSTREAM_PT=1.

Fine-tuning is supported and tested on the TPU v5e. We have tested so far:

• 🦙 Llama-2 7B, Llama-3 8B and newer;
• 💎 Gemma 2B and 7B.

You can check the examples:

• Fine-Tune Gemma on Google TPU
• The Llama fine-tuning script