torch-npu 2.9.1

NPU bridge for PyTorch

Ascend Extension for PyTorch

Overview

This repository develops the Ascend Extension for PyTorch named torch_npu to adapt Ascend NPU to PyTorch so that developers who use the PyTorch can obtain powerful compute capabilities 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 torch_npu. Before installing torch_npu, 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==2.9.1

For x86:

pip3 install torch==2.9.1+cpu  --index-url https://download.pytorch.org/whl/cpu

2. Install torch-npu 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
aarch64	Python3.10	link
aarch64	Python3.11	link

3. Install torch-npu

pip3 install torch-npu==2.9.1

From Source

In some special scenarios, users may need to compile torch-npu 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 torch-npu 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 torch-npu

   git clone https://gitcode.com/Ascend/pytorch.git -b v2.9.1 --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 torch-npu

   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 Ascend Extension for PyTorch for more detailed information.

PyTorch and Python Version Matching Table

PyTorch Version	Python Version
PyTorch1.11.0	Python3.7.x(>=3.7.5),Python3.8.x,Python3.9.x,Python3.10.x
PyTorch2.1.0	Python3.8.x,Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.2.0	Python3.8.x,Python3.9.x,Python3.10.x
PyTorch2.3.1	Python3.8.x,Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.4.0	Python3.8.x,Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.5.1	Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.6.0	Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.7.1	Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.8.0	Python3.9.x,Python3.10.x,Python3.11.x
PyTorch2.9.0	Python3.10.x,Python3.11.x
PyTorch2.9.1	Python3.10.x,Python3.11.x

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:

CANN Version	Supported PyTorch Version	Supported Extension Version	Github Branch
CANN 26.0.0	2.9.1	2.9.1	v2.9.1-26.0.0
CANN 8.5.0	2.9.0	2.9.0	v2.9.0-7.3.0
	2.8.0	2.8.0.post2	v2.8.0-7.3.0
	2.7.1	2.7.1.post2	v2.7.1-7.3.0
	2.6.0	2.6.0.post5	v2.6.0-7.3.0
CANN 8.3.RC1	2.8.0	2.8.0	v2.8.0-7.2.0
	2.7.1	2.7.1	v2.7.1-7.2.0
	2.6.0	2.6.0.post3	v2.6.0-7.2.0
	2.1.0	2.1.0.post17	v2.1.0-7.2.0
CANN 8.2.RC1	2.6.0	2.6.0	v2.6.0-7.1.0
	2.5.1	2.5.1.post1	v2.5.1-7.1.0
	2.1.0	2.1.0.post13	v2.1.0-7.1.0
CANN 8.1.RC1	2.5.1	2.5.1	v2.5.1-7.0.0
	2.4.0	2.4.0.post4	v2.4.0-7.0.0
	2.3.1	2.3.1.post6	v2.3.1-7.0.0
	2.1.0	2.1.0.post12	v2.1.0-7.0.0
CANN 8.0.0	2.4.0	2.4.0.post2	v2.4.0-6.0.0
	2.3.1	2.3.1.post4	v2.3.1-6.0.0
	2.1.0	2.1.0.post10	v2.1.0-6.0.0
CANN 8.0.RC3	2.4.0	2.4.0	v2.4.0-6.0.rc3
	2.3.1	2.3.1.post2	v2.3.1-6.0.rc3
	2.1.0	2.1.0.post8	v2.1.0-6.0.rc3
CANN 8.0.RC2	2.3.1	2.3.1	v2.3.1-6.0.rc2
	2.2.0	2.2.0.post2	v2.2.0-6.0.rc2
	2.1.0	2.1.0.post6	v2.1.0-6.0.rc2
	1.11.0	1.11.0.post14	v1.11.0-6.0.rc2
CANN 8.0.RC1	2.2.0	2.2.0	v2.2.0-6.0.rc1
	2.1.0	2.1.0.post4	v2.1.0-6.0.rc1
	1.11.0	1.11.0.post11	v1.11.0-6.0.rc1
CANN 7.0.0	2.1.0	2.1.0	v2.1.0-5.0.0
	2.0.1	2.0.1.post1	v2.0.1-5.0.0
	1.11.0	1.11.0.post8	v1.11.0-5.0.0
CANN 7.0.RC1	2.1.0	2.1.0.rc1	v2.1.0-5.0.rc3
	2.0.1	2.0.1	v2.0.1-5.0.rc3
	1.11.0	1.11.0.post4	v1.11.0-5.0.rc3
CANN 6.3.RC3.1	1.11.0	1.11.0.post3	v1.11.0-5.0.rc2.2
CANN 6.3.RC3	1.11.0	1.11.0.post2	v1.11.0-5.0.rc2.1
CANN 6.3.RC2	2.0.1	2.0.1.rc1	v2.0.1-5.0.rc2
	1.11.0	1.11.0.post1	v1.11.0-5.0.rc2
	1.8.1	1.8.1.post2	v1.8.1-5.0.rc2
CANN 6.3.RC1	1.11.0	1.11.0	v1.11.0-5.0.rc1
	1.8.1	1.8.1.post1	v1.8.1-5.0.rc1
CANN 6.0.1	1.5.0	1.5.0.post8	v1.5.0-3.0.0
	1.8.1	1.8.1	v1.8.1-3.0.0
	1.11.0	1.11.0.rc2（beta)	v1.11.0-3.0.0
CANN 6.0.RC1	1.5.0	1.5.0.post7	v1.5.0-3.0.rc3
	1.8.1	1.8.1.rc3	v1.8.1-3.0.rc3
	1.11.0	1.11.0.rc1（beta)	v1.11.0-3.0.rc3
CANN 5.1.RC2	1.5.0	1.5.0.post6	v1.5.0-3.0.rc2
	1.8.1	1.8.1.rc2	v1.8.1-3.0.rc2
CANN 5.1.RC1	1.5.0	1.5.0.post5	v1.5.0-3.0.rc1
	1.8.1	1.8.1.rc1	v1.8.1-3.0.rc1
CANN 5.0.4	1.5.0	1.5.0.post4	2.0.4.tr5
CANN 5.0.3	1.8.1	1.5.0.post3	2.0.3.tr5
CANN 5.0.2	1.5.0	1.5.0.post2	2.0.2.tr5

Hardware support

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

Product series	Product model
Atlas Training series products	Atlas 800（model: 9000）
	Atlas 800（model：9010）
	Atlas 900 PoD（model：9000）
	Atlas 300T（model：9000）
	Atlas 300T Pro（model：9000）
Atlas A2 Training series products	Atlas 800T A2
	Atlas 900 A2 PoD
	Atlas 200T A2 Box16
	Atlas 300T A2

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

Product series	Product model
Atlas 800I A2 Inference product	Atlas 800I A2

Pipeline Status

Due to the asynchronous development mechanism of upstream and downstream, incompatible modifications in upstream may cause some functions of torch_npu 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 22.03 SP2	CANN 7.1	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:

Status	Duration	Description
Planning	1-3 months	Plan features.
Development	6-12 months	Develop new features and fix issues, regularly release new versions. Different strategies are adopted for different versions of PyTorch, with a regular branch development cycle of 6 months and a long-term support branch development cycle of 12 months.
Maintained	1 year/3.5 years	Regular Release branch for 1 year, Long Term Support branch maintenance for 3.5 years. Fix major issues, do not incorporate new features, and release patch versions based on the impact of fixed bugs.
End Of Life (EOL)	N/A	Do not accept any modification to a branch.

PyTorch Maintenance Policies

PyTorch	Maintenance Policies	Status	Launch Date	Subsequent Status	EOL Date
2.9.0	Regular Release	Development	2026/01/15	Expected to enter maintenance status from July 15, 2026	-
2.8.0	Regular Release	Development	2025/10/15	Expected to enter maintenance status from March 15, 2026	-
2.7.1	Long Term Support	Development	2025/10/15	Expected to enter maintenance status from October 15, 2026
2.6.0	Regular Release	Development	2025/07/25	Expected to enter maintenance status from January 15, 2026	-
2.5.1	Regular Release	Maintained	2024/11/08	Expected to enter maintenance free status from August 8, 2026
2.4.0	Regular Release	Maintained	2024/10/15	Expected to enter maintenance free status from June 15, 2026
2.3.1	Regular Release	Maintained	2024/06/06	Expected to enter maintenance free status from June 7, 2026
2.2.0	Regular Release	EOL	2024/04/01		2025/10/14
2.1.0	Long Term Support	Maintained	2023/10/15	Expected to enter maintenance free status from December 30, 2026
2.0.1	Regular Release	EOL	2023/7/19		2024/3/14
1.11.0	Long Term Support	EOL	2023/4/19		2025/10/25
1.8.1	Long Term Support	EOL	2022/4/10		2023/4/10
1.5.0	Long Term Support	EOL	2021/7/29		2022/7/29

Reference Documents

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

Document Name	Document Link
Installation Guide	link
Network Model Migration and Training	link
Operator Adaptation	link
API List (PyTorch and Custom Interfaces)	link

License

Ascend Extension for PyTorch has a BSD-style license, as found in the LICENSE file.