What’s New in PyTorch 1.8? Key Features, APIs, and Performance Boosts

The PyTorch team announced PyTorch 1.8, integrating more than 3,000 commits since the 1.7 release and delivering major updates in compilation, code optimization, and scientific‑computing front‑end APIs, with new support for AMD ROCm.

Key highlights include:

Support for Python function conversion.

Stabilized APIs for FFT (torch.fft) and linear‑algebra functions (torch.linalg).

Complex‑tensor autograd support.

Performance improvements for Hessian and Jacobian matrix calculations.

Enhanced distributed training: more reliable NCCL, pipeline parallelism, RPC profiling, and gradient‑compression communication hooks.

Since PyTorch 1.6, features are classified as Stable, Beta, or Prototype.

Major updates were also made to auxiliary libraries such as TorchCSPRNG, TorchVision, TorchText, and TorchAudio.

New and Updated APIs

torch.fft – NumPy‑compatible FFT operations

The FFT feature, introduced as Beta in 1.7, becomes stable in 1.8, offering hardware‑accelerated, autograd‑compatible FFT functionality comparable to NumPy’s np.fft.

torch.linalg – NumPy‑style linear algebra functions

torch.linalg mirrors NumPy’s np.linalg module, providing operations such as Cholesky decomposition, determinant, and eigenvalue calculations.

Python code conversion with torch.fx

This Beta feature enables Python‑level code transformation, allowing developers to perform Conv/BN fusion, graph‑mode quantization, vmap implementation, and other custom graph transformations using torch.fx.

Distributed Training Enhancements

Pipeline parallelism

A new Beta API makes it easy to incorporate pipeline parallelism into the training loop.

DDP communication hooks

DDP communication hooks provide a generic interface for controlling gradient communication between workers.

Additional prototype features include:

ZeroRedundancyOptimizer – reduces per‑thread memory usage.

Process Group NCCL Send/Recv – enables collective operations at the Python level.

CUDA‑support in RPC using TensorPipe – speeds up RPC on multi‑GPU machines.

Remote Module – allows remote workers to be manipulated like local modules.

PyTorch Mobile Support

New mobile tutorials demonstrate image‑segmentation (DeepLabV3) on iOS and Android, and the release includes demo apps for image segmentation, object detection, neural machine translation, question answering, and vision transformers.

The Mobile Lite Interpreter is also introduced, reducing the runtime binary size.

Performance Optimizations

Benchmark utilities are added, and a new automatic quantization API – FX Graph Mode Quantization – is made available.

Hardware Support

Two Beta features expand the PyTorch dispatcher for new C++ backends and add AMD ROCm support, which is currently limited to Linux systems.

Reference links:

https://pytorch.org/blog/pytorch-1.8-released/

https://github.com/pytorch/pytorch

https://pytorch.org/

https://twitter.com/cHHillee/status/1367621538791317504