PyTorch 1.12 Released

What's new in PyTorch 1.12?

TorchArrow: TorchArrow is a new library released in beta with this update. The library is built to speed up preprocessing over batch data through an easy-to-use API.
Module: A new beta feature for Module computation is the functional API. This new functional_call() API lets you take complete control over parameters used in a Module's computation.
Complex Numbers: This update introduces two new features for complex numbers. The first is new functionality for complex convolutions, and the second is experimental support for the Complex32 datatype.
TorchData: DataPipe has improved compatibility with DataLoader. PyTorch now supports AWSSDK-based DataPipes. DataLoader2 has been introduced as a way to manage interactions between DataPipes and other APIs and backends.
functorch 0.2.0: PyTorch's own JAX-like feature set sees significant coverage improvement for functorch.jvp and APIs that rely on it. The new functionalize() method allows a function to be passed in and returns a non-mutating version of it.
nvFuser: nvFuser is the new, faster default fuser for compiling to CUDA devices.
Matrix Multiplication Precision: By default, matrix multiplication on float32 datatypes will now work in full precision mode, which is slower but will result in more consistent outcomes.
Channels Last Memory Format: New support for Channels Last memory format on computer vision models brings big performance gains on vision model inference.
Bfloat16: Less precise datatypes provide much faster computation times, so new improvements to the Bfloat16 datatype have been made in 1.12.
Accelerated Training On Mac: Support for taking advantage of Apple silicon GPUs in PyTorch has been added in a prototype state. Apple's Metal Performance Shaders are used as the backend for operations done on these GPUs.
BetterTransformer: Certain Transformer Encoder modules get a performance boost with new fastpath implementations.
FSDP API: Released in version 1.11 as a prototype, the FSDP API hits a beta release with version 1.12's release, adding a number of it's own improvement.

Find out more