PopTorch

PopTorch is a library by Graphcore that enables PyTorch models to run on Graphcore IPU hardware. It provides PyTorch-compatible wrappers and utilities that compile and execute neural networks on IPUs via the Poplar toolchain, allowing developers to leverage IPU performance without rewriting code.

Core components include trainingModel and inferenceModel wrappers that convert a PyTorch module into an IPU-optimized version. An Options object exposes IPU-specific settings such as device iterations, micro-batch size, and replication/pipeline configurations. PopTorch manages data transfer and memory to optimize performance and minimize host-device communication.

PopTorch integrates with standard PyTorch APIs, including DataLoader and optimizers, preserving familiar workflows while targeting IPUs. It supports both training and inference and works with common loss functions and metrics.

Usage pattern: define a PyTorch model, create a PopTorch Options object with desired IPU settings, wrap the model with trainingModel or inferenceModel, and run training or inference using conventional PyTorch data loading and loops. The program executes on Graphcore IPUs, either on-device or via Graphcore cloud offerings.

Overview: PopTorch is part of Graphcore’s software stack and interfaces with the Poplar compiler to enable efficient IPU execution. It aims to simplify porting PyTorch workloads to IPUs and to promote IPU-accelerated machine learning in research and production settings. PopTorch is maintained by Graphcore with community and enterprise support.