PyTorch Just Got a Massive Update

Manpreet Singh
2 min readMar 14, 2022

Welcome back! PyTorch is a very popular machine learning framework that was developed by Facebook, if you’re new to PyTorch, check out their website below:

PyTorch ended up getting a pretty big update the other day, this specific update now brings some awesome features to PyTorch, if you want to see their release log, here is a link to it:

So, let’s talk about some of my favorite features within this new version PyTorch!

FuncTorch

This new version of PyTorch now includes functorch as well, this was a feature that was heavily inspired by Google JAX, this library adds composable function transforms, providing more capability within PyTorch! Here are some examples of this new library:

TorchData

This new version of PyTorch now includes the beta release of TorchData, this is a library that makes creating data pipelines a bit more flexible, there are also some new commands that go along with this new feature as well:

Distributed Training

There have also been some updates to distributed training as well, this specific feature includes the ability to enable static graphs! Here is an example command of this in the DDP API:

ddp_model = DistributedDataParallel(model, static_graph=True)

You can also check out all of the release notes within this update below:

--

--