Pytorch 1.11 is out! This release includes more than 3,300 commits since version 1.10 and 434 contributors- including PyTorch developers who worked tirelessly to make this happen. Alongside the new features are beta versions of TorchData and functorch.
The TorchData library provides easy-to-use, high performance data loading primitives so you can build your own flexible and efficient pipelines.
functorch is a new library that brings compositionally useful function transforms to Pytorch. It’s currently in beta and is heavily inspired by Google JAX.
Distributed Data Parrallel (DDP) static graph optimizations are available in a stable release.