Ascend Extension for PyTorch

Familiarize yourself with Ascend Extension for PyTorch (the torch_npu plugin) and learn how to develop and tune models with it.

Ascend Extension for PyTorch
Get to know the key features and components of Ascend Extension for PyTorch.
Release Notes
Check the version information of Ascend Extension for PyTorch.
Quick Start
Migrate an example CNN model to run on the NPU.

Environment Preparation

Software Installation
Set up the PyTorch training environment on an Ascend device.

Model Development

Model Training Acceleration Library

MindSpeed
Accelerate end-to-end LLM, multimodal, and reinforcement learning training with optimized computing, memory, and communication.

Adapted Traditional Model Training

Supported Models
Find the list of adapted models. Refer to the README of each model for training and inference instructions.

Unadapted Traditional Model Migration

PyTorch Model Migration and Tuning
Migrate traditional models trained on other platforms to the NPU.
PyTorch Framework Feature Guide
Use the custom features provided by Ascend Extension for PyTorch.

Suites and Third-Party Libraries

Support Matrix for Suites and Third-Party Libraries
Learn about the model usites and acceleration libraries supported on Ascend devices, third-party libraries natively supported by Ascend, and Ascend in-house plugins.

APIs

PyTorch Native API Support
Check the support status of native PyTorch APIs on Ascend devices.
Ascend Extension for PyTorch Custom APIs
Browse the function prototypes, descriptions, parameters, and call examples of custom APIs.

Reference

Environment Variables
Find the environment variables available for training and inference.
Troubleshooting
Troubleshoot exceptions that occur during training and inference.
Security Statement
Follow the security hardening recommendations for Ascend Extension for PyTorch and related components.