Peregrine: A Transport Library for ML Workloads on TPUs

Our goal is to build an easy-to-use, light-weight, high-performance, and general transport library to serve machine learning workloads on TPU hosts.

Project Status

• Peregrine is still in experimental mode.
• You must contact us to get our support.

Releases

• Via Bazel Central Registry
• Via GitHub

Code Organization

• peregrine/src/: the library and all the unit tests
• peregrine/test/: the integration tests and benchmarks
• ci/: continuous integration in GitHub

Quick Start

• run.sh builds and tests all the code.

• To start the integration test, run

$ bazelisk run //peregrine/test/integration:main