cann-nnal-installer

CANN Toolkit+Kernels+NNAL Installation and Deployment

Install the complete CANN components (Toolkit, Kernels, NNAL) into a specified path in an Ascend NPU environment.

Core Principle: The installation order must be Toolkit → Kernels → NNAL and cannot be reversed. NNAL depends on the Toolkit's environment variables.

When to Invoke

The user requests installation of CANN (including toolkit, kernels, nnal)
The user needs to install CANN to a custom path
The user is setting up an Ascend development environment in a Docker container
The user needs to install a specific version of CANN (such as 9.0.0-beta.2)

Pre-checks

Step 1: Check the NPU driver

npu-smi info

Expected result: Output NPU device information (chip model, VRAM, etc.). Exception handling: If command not found, the Ascend driver and firmware need to be installed first; refer to the ascend-npu-driver-install skill.

Step 2: Confirm the chip model

Obtain the chip model from the npu-smi info output to select the corresponding Kernels package:

Chip Model	Kernels Package Name Suffix
910B / 910B3	`910b`
910	`910`
310P	`310p`
310B	`310b`

Step 3: Confirm the system architecture

uname -m

aarch64: ARM architecture, select _linux-aarch64.run
x86_64: x86 architecture, select _linux-x86_64.run

Installation Method 1: Download run packages from the official website for installation

Step 4: Obtain the CANN installation packages

Visit the Ascend Community Download Center to obtain the installation packages:

https://www.hiascend.com/developer/download/community/result?module=cann&cann=<版本号>

Packages to download(3):

Package Name	Format	Description
`Ascend-cann-toolkit_<版本>_linux-<架构>.run`	run package	Development toolkit (must be installed first)
`Ascend-cann-kernels-<芯片>_<版本>_linux-<架构>.run`	run package	Operator library (depends on Toolkit)
`Ascend-cann-nnal_<版本>_linux-<架构>.run`	run package	Acceleration library (depends on Toolkit environment variables)

Note: The official website download links require login authentication and cannot be directly wget. Alternatives:

Download after logging in through a browser, then upload to the server
Use "Method 2: Extract from Docker image" below

Step 5: Install Toolkit

INSTALL_PATH="<用户指定安装路径>"

chmod +x Ascend-cann-toolkit_*.run
./Ascend-cann-toolkit_*.run --install --install-path="${INSTALL_PATH}"

Default path：/usr/local/Ascend Custom path example：/home/user/run/cann900beta2

Step 6: Install Kernels

chmod +x Ascend-cann-kernels-*.run
./Ascend-cann-kernels-*.run --install --install-path="${INSTALL_PATH}"

Step 7: Configure Toolkit environment variables (prerequisite for NNAL installation)

source "${INSTALL_PATH}/ascend-toolkit/set_env.sh"

Key: The NNAL installation script depends on environment variables such as ASCEND_TOOLKIT_HOME , so they must be sourced first.

Step 8: Install NNAL

chmod +x Ascend-cann-nnal_*.run
./Ascend-cann-nnal_*.run --install --install-path="${INSTALL_PATH}"

Installation Method 2: Extract from Docker image

When the run packages cannot be downloaded directly, the installed CANN files can be extracted from Huawei's official CANN Docker image.

Step 4: Confirm the local image

docker images | grep cann

Huawei official CANN image tag format:quay.io/ascend/cann:<版本>-<芯片>-<OS>-py<Python版本>

Example:quay.io/ascend/cann:9.0.0-beta.2-910-openeuler24.03-py3.11

If there is no local image, pull it first:

docker pull quay.io/ascend/cann:9.0.0-beta.2-910-openeuler24.03-py3.11

Step 5: Create a temporary container and extract the files

INSTALL_PATH="<用户指定安装路径>"
CANN_IMAGE="quay.io/ascend/cann:9.0.0-beta.2-910-openeuler24.03-py3.11"

docker create --name cann_temp "${CANN_IMAGE}"

docker cp cann_temp:/usr/local/Ascend/cann-9.0.0-beta.2 "${INSTALL_PATH}/"

docker cp cann_temp:/usr/local/Ascend/nnal "${INSTALL_PATH}/"

docker rm cann_temp

Directory structure inside the image：

Path inside the image	Content	Target path
`/usr/local/Ascend/cann-<版本>/`	Toolkit + Kernels	`${INSTALL_PATH}/cann-<版本>/`
`/usr/local/Ascend/nnal/`	NNAL (atb acceleration library)	`${INSTALL_PATH}/nnal/`

Step 6: Fix the hardcoded paths in set_env.sh

The set_env.sh in the Docker image contains hardcoded paths /usr/local/Ascend/cann-<版本>, which need to be replaced with the actual installation paths:

sed -i "s|/usr/local/Ascend/cann-<版本>|${INSTALL_PATH}/cann-<版本>|g" \
  "${INSTALL_PATH}/cann-<版本>/set_env.sh"

Also check for hardcoded paths in other scripts:

grep -rl '/usr/local/Ascend/cann-<版本>' "${INSTALL_PATH}/" --include='*.sh' | \
  xargs sed -i "s|/usr/local/Ascend/cann-<版本>|${INSTALL_PATH}/cann-<版本>|g"

Environment variable configuration

Step 9: Load environment variables

source "${INSTALL_PATH}/cann-<版本>/set_env.sh"
source "${INSTALL_PATH}/nnal/atb/set_env.sh"

Persistent configuration(write to the shell configuration file):

cat >> ~/.bashrc << EOF
source ${INSTALL_PATH}/cann-<版本>/set_env.sh
source ${INSTALL_PATH}/nnal/atb/set_env.sh
EOF

Environment variable description

Variable name	Source	Description
`ASCEND_TOOLKIT_HOME`	toolkit set_env.sh	Toolkit installation root path
`ASCEND_HOME_PATH`	toolkit set_env.sh	Same as ASCEND_TOOLKIT_HOME
`ASCEND_OPP_PATH`	toolkit set_env.sh	OPP operator path
`ATB_HOME_PATH`	nnal atb set_env.sh	ATB acceleration library path
`LD_LIBRARY_PATH`	Two set_env.sh	Dynamic library search path
`PYTHONPATH`	toolkit set_env.sh	Python module search path

Installation verification

Step 10: Verify installation integrity

source "${INSTALL_PATH}/cann-<版本>/set_env.sh"
source "${INSTALL_PATH}/nnal/atb/set_env.sh"

echo "=== Toolkit Version ==="
cat "${INSTALL_PATH}/cann-<版本>/<架构>-linux/ascend_toolkit_install.info" | grep version

echo "=== Kernels Version ==="
cat "${INSTALL_PATH}/cann-<版本>/<架构>-linux/ascend_ops_install.info" | grep version

echo "=== NNAL Version ==="
cat "${INSTALL_PATH}/nnal/atb/latest/version.info"

echo "=== Key Binaries ==="
which bisheng && echo "bisheng OK"
which atc && echo "atc OK"

echo "=== Environment Variables ==="
echo "ASCEND_TOOLKIT_HOME=${ASCEND_TOOLKIT_HOME}"
echo "ATB_HOME_PATH=${ATB_HOME_PATH}"

echo "=== NPU Status ==="
npu-smi info | head -5

Verification checklist

Toolkit version information is correct
Kernels(ops) version information is correct
NNAL(atb) version information is correct
bisheng command is available
atc command is available
ASCEND_TOOLKIT_HOME points to the correct path
ATB_HOME_PATH points to the correct path
npu-smi info device information is output normally

Installation verification report format

CANN INSTALLATION REPORT
========================
Install Path:  <安装路径>
Toolkit:       <版本号>
Kernels:       <版本号>
NNAL:          <版本号>
Architecture:  <aarch64/x86_64>
Chip:          <芯片型号>

VERDICT: SUCCESS / PARTIAL / FAILED

CHECKS:
  [1] Toolkit installed:    PASS / FAIL
  [2] Kernels installed:    PASS / FAIL
  [3] NNAL installed:       PASS / FAIL
  [4] Environment vars:     PASS / FAIL
  [5] bisheng available:    PASS / FAIL
  [6] atc available:        PASS / FAIL
  [7] NPU accessible:       PASS / FAIL

ISSUES: <count>
  [CRITICAL] <finding>
  [WARNING]  <finding>

Common issues

1. Insufficient permissions reported when installing the run package

Cause: Insufficient permissions on the installation directory.

Solution：

chmod 755 <安装路径的父目录>

2. NNAL installation failed, reporting that the Toolkit environment variables cannot be found

Cause: The Toolkit's set_env.sh was not sourced first.

Solution：

source "${INSTALL_PATH}/ascend-toolkit/set_env.sh"
# 或（Docker镜像提取方式）
source "${INSTALL_PATH}/cann-<版本>/set_env.sh"

3. The bisheng/atc command cannot be found after extracting the Docker image

Cause: The paths in set_env.sh have not been corrected.

Solution: Run the sed command in Step 6 to replace the hardcoded paths.

4. The official website download link cannot be wget directly

Cause: The Ascend community download links require login authentication, and the URLs stored in OBS carry signatures and expiration times.

Solution: Use "Method 2: Extract from the Docker image".

5. The Kernels package name does not match the chip model

Cause: For the 910B chip, select the kernels package with the 910b suffix, not 910。

Solution: Select the correct kernels package based on the chip model confirmed in Step 2.

Precautions

Installation order: Must be Toolkit → Kernels → NNAL, and the order cannot be reversed
Environment variables: Before installing NNAL, you must first source the Toolkit's set_env.sh
Path consistency: The three components must be installed under the same path
Version matching: The Toolkit, Kernels, and NNAL versions must be consistent
Driver compatibility: The CANN version must be compatible with the host NPU driver version

About this skill

When to Invoke

Pre-checks

Step 1: Check the NPU driver

Step 2: Confirm the chip model

Step 3: Confirm the system architecture

Installation Method 1: Download run packages from the official website for installation

Step 4: Obtain the CANN installation packages

Step 5: Install Toolkit

Step 6: Install Kernels

Step 7: Configure Toolkit environment variables (prerequisite for NNAL installation)

Step 8: Install NNAL

Installation Method 2: Extract from Docker image

Step 4: Confirm the local image

Step 5: Create a temporary container and extract the files

Step 6: Fix the hardcoded paths in set_env.sh

Environment variable configuration

Step 9: Load environment variables

Environment variable description

Installation verification

Step 10: Verify installation integrity

Verification checklist

Installation verification report format

Common issues

1. Insufficient permissions reported when installing the run package

2. NNAL installation failed, reporting that the Toolkit environment variables cannot be found

3. The bisheng/atc command cannot be found after extracting the Docker image

4. The official website download link cannot be wget directly

5. The Kernels package name does not match the chip model

Precautions

Reference links