What is libtpu?

libtpu is the core library that enables machine learning frameworks like JAX, PyTorch, and TensorFlow to execute models on Google Cloud TPUs. It provides core functionality for compilation, inter-chip communication (ICI), and runtime execution.

libtpu also includes a set of SDK primitives for direct TPU interaction and deployment.

Learn more about Cloud TPUs at Google Cloud TPUs.

Version 0.0.46

Compatibility

• Python Compatibility: We add the support for 3.15 and 3.15ft. The libtpu SDK is now compatible with the Python versions 3.11, 3.12, 3.13, 3.14, 3.14-ft, 3.15 and 3.15ft where ft corresponds to free threaded python variant.

Version 0.0.45

Compatibility

• Python Compatibility: The libtpu SDK is now compatible with the Python versions 3.11, 3.12, 3.13, 3.14 and 3.14-ft where ft corresponds to free threaded python variant. We have re-enabled support Python 3.11 to keep it compatible with TorchTPU.

Version 0.0.43

Compatibility

• Python Compatibility: The libtpu SDK is now compatible with the Python versions 3.12, 3.13, 3.14 and 3.14-ft where ft corresponds to free threaded python variant. We have dropped support for Python 3.11 and 3.13-ft following Jax.

Version 0.0.21.1

Compatibility

• JAX Compatibility: libtpu supports JAX 0.7.1 or newer.

• Python Compatibility: The libtpu SDK is now compatible with the Python versions 3.11, 3.12, 3.13, 3.13-ft, 3.14 and 3.14-ft where ft corresponds to free threaded python variant.