Compiling the Source Code and Installing OpPlugin

Installation instructions

Hardware mapping table

The Ascend training device includes the following models, which can be used as the training environment of the PyTorch model.

Product range	Product Model
Atlas Training Series	Atlas 800 training server (model: 9000)
	Atlas 800 training server (model: 9010)
	Atlas 900 PoD (model: 9000)
	Atlas 300T training card (model: 9000)
	Atlas 300T Pro training card (model: 9000)
Atlas A2 Training Series	Atlas 800T A2 training server
	Atlas 900 A2 PoD cluster base unit
	Atlas 200T A2 Box16 Heterogeneous Subrack
Atlas A3 Training Series	Atlas 800T A3 training server
	Atlas 900 A3 SuperPoD Super Node

The Ascend inference device includes the following models, which can be used as the inference environment for large models.

Product range	Product Model
Atlas 800I A2 inference product	Atlas 800I A2 inference server

Software mapping table

PyTorch	TorchNPU	OpPlugin	Python	GCC
2.7.1	v2.7.1	master	3.9, 3.10, 3.11, 3.12, 3.13	11.2
2.8.0	v2.8.0	master	3.9, 3.10, 3.11, 3.12, 3.13	13.3
2.9.0	v2.9.0	master	3.10, 3.11, 3.12, 3.13	13.3
2.10.0	v2.10.0	master	3.10, 3.11, 3.12, 3.13	13.3
2.11.0	v2.11.0	master	3.10, 3.11, 3.12, 3.13	13.3
2.12.0	v2.12.0	master	3.10, 3.11, 3.12, 3.13	13.3
2.13.0	master	master	3.10, 3.11, 3.12, 3.13	13.3

Installation Dependency

You are advised to use the development image provided by TorchNPU for compilation. The image can be obtained in either of the following ways: directly pull the pre-built image from the Ascend image repository, or build the image by yourself using the Dockerfile.

Pull the image directly

A ready-to-use development image is provided for compiling OpPlugin. You can pull it directly from the Ascend image repository: torch-npu-devel

Taking the Atlas A2 training products as an example, run the following command to pull the image:
```
docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/torch-npu-devel:2.13.0-cann9.1.0-910b-manylinux_2_28
```
After the image is pulled, use the following command to start and access the Docker container and mount the OpPlugin source code to the container:
```
docker run -it -v /{code_path}/op-plugin:/home/op-plugin swr.cn-south-1.myhuaweicloud.com/ascendhub/torch-npu-devel:2.13.0-cann9.1.0-910b-manylinux_2_28 bash
```
Build the image by yourself

This repository (op-plugin/docker/devel) provides a ready-to-use Dockerfile that can automatically detect the architecture to pull the base image. You can read the README file in that directory for more information and build your own development environment as instructed.
```
cd op-plugin/docker/devel
export DOCKER_BUILDKIT=1
docker build -t op-plugin-builder:v1 .
```
Alternatively, use the one-click container creation script builder.sh:
```
cd op-plugin/docker/devel
export DOCKER_BUILDKIT=1
bash builder.sh --cann
```
After the image is built, use the following command to start and access the Docker container and mount the OpPlugin source code to the container:
```
docker run -it -v /{code_path}/op-plugin:/home/op-plugin op-plugin-builder:v1 bash
```

{code_path} indicates the path of the OpPlugin source code. Replace it based on the actual situation.

Note

The container started directly can only be used for compiling the OpPlugin plug-in.
To run OpPlugin in the NPU environment inside the container, ensure that the driver exists on the host (which can be confirmed by the npu-smi command) and mount the driver when starting the container. For details, see the README in the directory (op-plugin/docker/devel).
In the container scenario, the image and source code need to be obtained from the external network. For network issues such as proxy configuration, see the Docker official documentation.

In the physical machine and virtual machine scenarios, you need to install the system dependencies and the official PyTorch framework by yourself. For details about the dependency installation, see the TorchNPU.

Operation Procedure

Script for configuring CANN environment variables.
```
source <CANN_install_dir>/<CANN_path>/set_env.sh
```
The default path of the environment variable script is /usr/local/npu/ascend-toolkit/set_env.sh. The path of ascend-toolkit depends on the name of the installed CANN software.
Generate the binary installation package of the plug-in.

Download the branch code of the corresponding OpPlugin version and go to the root directory of the plug-in.
```
git clone --branch master https://gitcode.com/ascend/op-plugin.git
cd op-plugin
```
Run the following commands to compile and build the plug-in. The following uses PyTorch 2.13.0 and Python 3.10 as an example.
```
bash ci/build.sh --python=3.10 --pytorch=v2.13.0-26.2.0
```
[!NOTICE]

For details about the GCC and Python versions during compilation, see the Software mapping table During the compilation, the build folder is created in the root directory of the plug-in, and the source code of the corresponding TorchNPU version is downloaded for collaborative compilation. If the build/pytorch directory exists, the TorchNPU source code will not be downloaded repeatedly during OpPlugin compilation. To download the latest TorchNPU source code, delete the build/pytorch directory.

After the compilation is complete, install the TorchNPU package generated in the dist directory. If a non-root user is used to install the TorchNPU package, add the following information to the end of the command:--user.

pip3 install --upgrade dist/torch_npu-{torch_npu_version}-{Python_version}-{arch}.whl
# Replace it with the name of the generated whl package. {torch_npu_version} indicates the TorchNPU version, {Python_version} indicates the Python version, and {arch} indicates the target architecture.
# A typical whl package name is similar to torch_npu-2.13.0rc1-cp310-cp310-linux_aarch64.whl.

Uninstalling

Run the following command to uninstall the torch:

pip uninstall torch_npu