TorchNPU

Overview

TorchNPU is a PyTorch adapter plugin that brings Ascend NPU support into the PyTorch ecosystem, enabling developers to harness the high-performance computing power of Ascend AI processors.

Ascend is a full-stack AI computing infrastructure for industry applications and services based on Huawei Ascend processors and software. For more information about Ascend, see Ascend Community.

Installation

From Binary

Provide users with wheel package to quickly install TorchNPU. Before installing TorchNPU, complete the installation of CANN according to Ascend Auxiliary Software. To obtain the CANN installation package, refer to the CANN Installation.

1. Install PyTorch

   Install PyTorch through pip.

   For Aarch64:

   pip3 install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu
   

   For x86:

   pip3 install torch==2.11.0+cpu --index-url https://download.pytorch.org/whl/cpu
   

2. Install TorchNPU dependencies

   Run the following command to install dependencies.

   pip3 install pyyaml
   pip3 install setuptools
   

   If the installation fails, use the download link or visit the PyTorch official website to download the installation package of the corresponding version.

   OS arch	Python version	link
   x86	Python3.10	link
   x86	Python3.11	link
   x86	Python3.12	link
   aarch64	Python3.10	link
   aarch64	Python3.11	link
   aarch64	Python3.12	link

3. Install TorchNPU

   pip3 install torch-npu==2.11.0rc1
   

From Source

In some special scenarios, users may need to compile TorchNPU by themselves.Select a branch in table Ascend Auxiliary Software and a Python version in table PyTorch and Python Version Matching Table first. The docker image is recommended for compiling TorchNPU through the following steps(It is recommended to mount the working path only and avoid the system path to reduce security risks.), the generated .whl file path is ./dist/. Note that gcc version has the following constraints if you try to compile without using docker image: we recommend to use gcc 13.3 for both ARM and X86.

1. Clone TorchNPU

   git clone https://gitcode.com/ascend/pytorch.git -b v2.11.0 --depth 1
   

2. Build Docker Image

   cd pytorch/ci/docker/{arch} # {arch} for X86 or ARM
   docker build -t manylinux-builder:v1 .
   

3. Enter Docker Container

   docker run -it -v /{code_path}/pytorch:/home/pytorch manylinux-builder:v1 bash
   # {code_path} is the torch_npu source code path
   

4. Compile TorchNPU

   Take Python 3.10 as an example.

   cd /home/pytorch
   bash ci/build.sh --python=3.10
   

Tips

If you would like to compile with new C++ ABI, then first run this command, at this point, the recommended compilation environment is same to community torch package: glibc 2.28, gcc 13.3

export _GLIBCXX_USE_CXX11_ABI=1

Meanwhile, we support configuring -fabi-version using the following variables，require consistency with the community torch package

export _ABI_VERSION=18

Getting Started

Prerequisites

Initialize CANN environment variable by running the command as shown below.

# Default path, change it if needed.
source /usr/local/Ascend/ascend-toolkit/set_env.sh

Quick Verification

You can quickly experience Ascend NPU by the following simple examples.

import torch
- import torch_npu # No longer needed in torch_npu 2.5.1 and later versions

x = torch.randn(2, 2).npu()
y = torch.randn(2, 2).npu()
z = x.mm(y)

print(z)

User Manual

Refer to API of TorchNPU for more detailed information.

PyTorch and Python Version Matching Table

Ascend Auxiliary Software

PyTorch Extension versions follow the naming convention {PyTorch version}-{Ascend version}, where the former represents the PyTorch version compatible with the PyTorch Extension, and the latter is used to match the CANN version. The detailed matching is as follows:

Hardware support

The Ascend training device includes the following models, all of which can be used as training environments for PyTorch models

The Ascend inference device includes the following models, all of which can be used as inference environments for large models

Pipeline Status

Due to the asynchronous development mechanism of upstream and downstream, incompatible modifications in upstream may cause some functions of TorchNPU to be unavailable (only upstream and downstream development branches are involved, excluding stable branches). Therefore, we built a set of daily tasks that make it easy to detect relevant issues in time and fix them within 48 hours (under normal circumstances), providing users with the latest features and stable quality.

OS	CANN Version(Docker Image)	Upstream Branch	Downstream Branch	Period	Status
openEuler 24.03	CANN 8.5	main	master	UTC 1200 daily

Suggestions and Communication

Everyone is welcome to contribute to the community. If you have any questions or suggestions, you can submit Github Issues. We will reply to you as soon as possible. Thank you very much.

Branch Maintenance Policies

The version branches of AscendPyTorch have the following maintenance phases:

PyTorch Maintenance Policies

Reference Documents

For more detailed information on installation guides, model migration, training/inference tutorials, and API lists, please refer to the TorchNPU on the HiAI Community.

License

TorchNPU has a BSD-style license, as found in the LICENSE file.