torch_npu插件有如下两种安装方式：

快速安装：通过wheel格式的二进制软件包直接安装。
源码编译安装：编译安装适用于二次开发场景，用户可以选择需要的分支自行编译torch_npu。

获取安装命令

执行安装命令前，请参见安装前必读完成环境变量配置。

表1 安装torch_npu插件
PyTorch版本	torch_npu插件版本	Python版本	系统架构	CANN版本	安装方式	安装命令
2.1.0	6.0.0	Python 3.8	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.8	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.8	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.8	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.9	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.9	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.10	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.10	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.11	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.1.0/torch_npu-2.1.0.post10-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.1.0.post10-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.11	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
2.3.1	6.0.0	Python 3.8	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.8	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.8	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.8	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.9	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.9	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.10	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.10	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.11	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.3.1/torch_npu-2.3.1.post4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.3.1.post4-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.11	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
2.4.0	6.0.0	Python 3.8	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.8	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.8	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.8	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.9	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.9	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.9	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.10	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.10	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.10	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	AArch64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
	6.0.0	Python 3.11	AArch64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”
	6.0.0	Python 3.11	X86_64	8.0.0	Pip	# 下载插件包 wget https://gitee.com/ascend/pytorch/releases/download/v6.0.0-pytorch2.4.0/torch_npu-2.4.0.post2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl # 安装命令 pip3 install torch_npu-2.4.0.post2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
	6.0.0	Python 3.11	X86_64	8.0.0	Source	# 参见下方“源码编译安装torch_npu插件”

更多PyTorch插件版本可单击Link查询。

执行如下命令，若返回True则说明安装成功。

python3 -c "import torch;import torch_npu;print(torch_npu.npu.is_available())"

若PyTorch版本为2.1.0，出现“找不到google或者protobuf或者protobuf版本过高”报错时，需执行如下命令：
```
pip3 install protobuf==3.20
```

源码编译安装torch_npu插件

以下操作步骤以安装PyTorch 2.1.0版本为例。

方式一（推荐）：容器场景
1. 下载torch_npu源码。
```
git clone https://gitee.com/ascend/pytorch.git -b v2.1.0-6.0.0 --depth 1  
```
  下载对应PyTorch版本分支代码，进入插件根目录，以v2.1.0-6.0.0为例，其他版本请参见《版本说明》中的“版本配套关系”章节下载对应PyTorch版本分支代码。
2. 构建镜像。
```
cd pytorch/ci/docker/{arch} 
docker build -t manylinux-builder:v1 .
```
  - {arch}表示CPU架构（X86或ARM）。
  - 注意不要遗漏命令结尾的“.”。
3. 进入Docker容器，并将torch_npu源代码挂载至容器内。
```
docker run -it -v /{code_path}/pytorch:/home/pytorch manylinux-builder:v1 bash
```
  {code_path}表示torch_npu源代码路径，请根据实际情况进行替换。
4. 编译生成二进制安装包。
```
cd /home/pytorch
bash ci/build.sh --python=3.8
```
  如需指定其他Python版本请使用--python=3.9、--python=3.10或--python=3.11。
5. 在运行环境中安装生成的插件torch_npu包，如果使用非root用户安装，需要在命令后加--user。
```
# 请用户根据实际情况更改命令中的torch_npu包名
pip3 install --upgrade dist/torch_npu-2.1.0.post10-cp38-cp38-linux_aarch64.whl
```
方式二：物理机及虚拟机场景
1. 安装依赖。
  选择编译安装方式安装时需要安装系统依赖，根据不同类型的操作系统，选择对应的命令安装所需依赖。
  - openEuler系列（openEuler、CentOS、Kylin、BCLinux、BC-Linux-for-Euler、UOS20 1050e、UOS20 1020e、UOSV20、AntOS、CTyunOS、CULinux、Tlinux）：
    1. 安装依赖（除gcc和cmake以外）。
      yum install -y patch libjpeg-turbo-devel dos2unix openblas git
    2. 安装gcc和cmake。
      根据实际情况，安装对应gcc和cmake版本，版本信息及安装指导请参见以下说明。
  - Debian系列（Ubuntu、Debian、UOS20、UOS20 SP1）：
    1. 安装依赖（除gcc和cmake以外）。
      apt-get install -y patch build-essential libbz2-dev libreadline-dev wget curl llvm libncurses5-dev libncursesw5-dev xz-utils tk-dev liblzma-dev m4 dos2unix libopenblas-dev git
    2. 安装gcc和cmake。
      根据实际情况，安装对应gcc和cmake版本，版本信息及安装指导请参见以下说明。
  AArch64架构安装gcc10.2.0版本，X86_64架构安装gcc9.4.0版本，cmake则需为3.18.0版本及以上，安装指导可参见安装10.2.0版本gcc和安装3.18.0版本cmake。
2. 编译生成torch_npu插件的二进制安装包。
  1. 下载对应PyTorch版本分支代码，进入插件根目录，以v2.1.0-6.0.0为例，其他版本请参见《版本说明》中的“版本配套关系”章节下载对应PyTorch版本分支代码。
```
git clone -b v2.1.0-6.0.0 https://gitee.com/ascend/pytorch.git 
cd pytorch
```
  2. 编译生成二进制安装包。
```
# 如需指定其他Python版本请使用--python=3.9、--python=3.10或--python=3.11
bash ci/build.sh --python=3.8
```
3. 安装pytorch/dist目录下生成的插件torch_npu包，如果使用非root用户安装，需要在命令后加--user。
```
# 请用户根据实际情况更改命令中的torch_npu包名
pip3 install --upgrade dist/torch_npu-2.1.0.post10-cp38-cp38-linux_aarch64.whl
```

验证是否成功安装

方法一

执行如下命令：

python3 -c "import torch;import torch_npu; a = torch.randn(3, 4).npu(); print(a + a);"

显示如下类似信息说明PyTorch框架与插件安装成功。

tensor([[-0.6066,  6.3385,  0.0379,  3.3356],
        [ 2.9243,  3.3134, -1.5465,  0.1916],
        [-2.1807,  0.2008, -1.1431,  2.1523]], device='npu:0')

方法二

import torch
import torch_npu

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

print(z)

输出如下类似信息说明PyTorch框架与插件安装成功。

tensor([[-0.0515,  0.3664],
        [-0.1258, -0.5425]], device='npu:0')