TorchNPU

As a core component of the Ascend for PyTorch community, TorchNPU (formerly known as Ascend Extension for PyTorch) is an AI adaptation framework tailored by Ascend specifically for PyTorch. It enables the PyTorch framework to directly leverage Ascend NPUs, providing developers with robust AI computing power.

Key Features of TorchNPU 26.1.0

The original Ascend Extension for PyTorch and torch_npu have now been unified under the name TorchNPU.
Added support for Ascend 950DT products.
Enabled DVM graphless operator fusion in eager mode. View details
Added support for LibTorch Stable ABI on the NPU. View details
Key compute APIs now support heterogeneous inputs.
Added support for operator compilation under all-axis dynamic input tensor scenarios.

Quick Start

TorchNPU Overview

Release Notes

TorchNPU Documentation

**Table 1** Installation Guide
Installation Guide
Software Installation

**Table 2** User Guide
User Guide
TorchNPU Overview
Quick Start
Torch.compile
Companion Software Libraries
Troubleshooting

**Table 3** API
API
Custom APIs
Environment Variables

**Table 4** Developer Guide
Developer Guide
Framework Features
Model Migration and Debugging
Model Performance Tuning

**Table 5** References
References
Ascend Product Overview
Security Statement
Glossary