Pax is a Jax-based machine learning framework for training large scale models. Pax allows for advanced and fully configurable experimentation and parallelization, and has demonstrated industry leading model flop utilization rates.
This PR updates the Dockerfile in paxml/contrib/gpu to be compatible with upstream, fixes the path in paxml/contrib/gpu/docker/build.sh, and adjusts the 126M Lambada config to report strict accuracy by default.
This PR updates the Dockerfile in
paxml/contrib/gputo be compatible with upstream, fixes the path inpaxml/contrib/gpu/docker/build.sh, and adjusts the 126M Lambada config to report strict accuracy by default.