Installation Guide

This document describes how to prepare a MindIE SD environment in container and bare-metal setups. If you want to run an example end to end, read this guide together with Quick Start.

Option 1: Container image installation

This section describes how to prepare and use a MindIE container image.

Install the driver and firmware.

Make sure the host already has the NPU driver and firmware installed. If not, see the CANN Software Installation Guide and choose the appropriate installation scenario.
- Installation method: install on a physical machine.
- Operating system: choose the operating system used in your environment.
- Service scenario: choose training, inference, and development/debugging.
Install Docker on the host yourself. Docker 24.x.x or later is recommended. Make sure the environment can access the network before configuring package sources.
Obtain the MindIE container image.
- Open the Ascend image repository page.
- Sign in with a Huawei account.
- On the image version tab, select the image that matches your hardware form factor and download it.
- Follow the download instructions shown by the site.

Use the image.

The following options usually need to be adjusted for your environment:

Item	Location in command	Description
Container name	`--name <container-name>` and `docker exec <container-name>`	Replace with your custom container name.
Image name	`mindie:2.2.RC1-800I-A2-py311-openeuler24.03-lts` at the end of the command	Replace with the local image name and tag shown by `docker images`.
Mount paths	The host-side path of each `-v` option, such as `/path-to-weights`	Replace with the real paths on the host machine.

Start the container with a command like this:

docker run -it -d --net=host --shm-size=1g \
    --name <container-name> \
    --device=/dev/davinci_manager:rwm \
    --device=/dev/hisi_hdc:rwm \
    --device=/dev/devmm_svm:rwm \
    --device=/dev/davinci0:rwm \
    -v /usr/local/Ascend/driver:/usr/local/Ascend/driver:ro \
    -v /usr/local/Ascend/firmware/:/usr/local/Ascend/firmware:ro \
    -v /usr/local/sbin:/usr/local/sbin:ro \
    -v /path-to-weights:/path-to-weights:ro \
    mindie:2.2.RC1-800I-A2-py311-openeuler24.03-lts bash

Note The image name and tag above are only examples. Use docker images on the host to inspect the images already available on your machine.

The --device mounts are set to rwm rather than rw or r. This avoids known runtime and npu-smi failures on Atlas 800I A2 and Atlas 800I A3 systems when NPUs are mounted with insufficient permissions.

Enter the container:

docker exec -it <container-name> bash

Install other dependencies.

Install the model-specific dependencies before running inference. For Wan2.1, for example:

git clone https://modelers.cn/MindIE/Wan2.1.git
cd Wan2.1
pip install -r requirements.txt

Install gcc and g++ if they are missing from the container, then export the include path:

yum install gcc g++ -y
export CPLUS_INCLUDE_PATH=/usr/include/c++/12/:/usr/include/c++/12/aarch64-openEuler-linux/:$CPLUS_INCLUDE_PATH

Option 2: Bare-metal installation

This section describes how to prepare a full development environment on a physical machine, including the driver and firmware, CANN, PyTorch, Torch NPU, and MindIE SD installation paths.

Install the driver and firmware.

Make sure the host already has the NPU driver and firmware installed. If not, follow the CANN Software Installation Guide and choose the matching installation scenario.
- Installation method: install on a physical machine.
- Operating system: choose the operating system used in your environment.
- Service scenario: choose training, inference, and development/debugging.
Install CANN.

Required CANN packages include:
- the CANN Toolkit development package
- the CANN Kernels operator package
Follow the CANN Software Installation Guide and choose the physical-machine CANN installation path.
Install PyTorch and Torch NPU.

Required packages include:
- the PyTorch wheel package, version 2.1.0
- the torch_npu wheel package
- Install PyTorch by following the Ascend Extension for PyTorch installation guide.
- Install torch_npu by following the optional extension module guide in the same documentation set.
Note If gcc and g++ are missing, install them and export the include path:
```
yum install gcc g++ -y
export CPLUS_INCLUDE_PATH=/usr/include/c++/12/:/usr/include/c++/12/aarch64-openEuler-linux/:$CPLUS_INCLUDE_PATH
```
Install other environment dependencies.

Install the model-specific dependencies required for inference:
```
pip install -r requirements.txt
```
Install MindIE SD.

MindIE SD does not need to be installed separately when you install the full MindIE package. To install the MindIE package:
1. Upload the MindIE package to any path on the target machine, for example /home/package, then grant execute permission:
```
cd /home/package
chmod +x Ascend-mindie_<version>_linux-<arch>_<abi>.run
```
2. Export the ascend-toolkit environment variables, for example with the default root installation path:
```
source /usr/local/Ascend/ascend-toolkit/set_env.sh
```
3. Install the package:
```
./package-name.run --install --quiet
```
4. Verify the installation:
```
python3 -c "import torch, torch_npu, mindiesd; print(torch_npu.npu.is_available())"
```