FAQ

Installation and Environment

Q: Why does import mm fail with the error ModuleNotFoundError: No module named 'mm'?

Cause: Python cannot find the installed mm package (the Wheel package has not been installed, or it was installed in a different Python environment).

Solution:

pip3 show mm
python3 -c "import mm; print('mm import: OK')"

If pip3 show mm produces no output, complete the installation by following the run package installation or Wheel package installation procedure.

Q: Why does import mm fail when libcore.so or libascendcl.so cannot be found?

Cause:

• Wheel installation: CANN environment variables are not loaded, the Wheel package version does not match the required version, or the installation is incomplete.
• .run installation: source set_env.sh has not been executed, or the installation directory is not complete.

Solution (Wheel installation):

source /usr/local/Ascend/ascend-toolkit/set_env.sh   # Default CANN installation path; modify it according to the actual installation path
pip3 install --force-reinstall --no-deps /path/to/mm-*.whl
python3 -c "import mm; print('mm import: OK')"

The Wheel package does not require MULTIMODAL_SDK_HOME to be set. Native libraries are bundled in the Wheel package and are automatically loaded when mm is imported.

Solution (.run installation):

source ${MULTIMODAL_SDK_HOME}/script/set_env.sh
python3 -c "import mm; print('mm import: OK')"

If ${MULTIMODAL_SDK_HOME} is not set, execute the command using the actual installation path. For example:

source /home/work/Mind_SDK/multimodal/script/set_env.sh

Q: Why does npu-smi info produce no output or report an error?

Cause: The NPU driver or firmware is not properly installed, or the CANN environment variables have not taken effect.

Solution:

1. Ensure that Ascend HDK 26.1.0 and CANN 9.1.0 (or compatible patch versions) are installed.
2. Load the CANN environment variables: source /usr/local/Ascend/ascend-toolkit/set_env.sh (modify the path according to the actual installation).
3. Run npu-smi info again. If the issue persists, restart the host machine and retry.

Q: Why is there a torch/transformers version conflict?

Cause: The versions of third-party dependencies do not match the Multimodal SDK requirements.

Solution: Install the required versions according to Installation Guide > Other Dependencies:

pip3 install transformers==4.51.3 pillow==11.2.1 numpy==1.26.4

Install torch and TorchNPU according to the compatibility table for vllm-ascend==v0.8.5rc1. For details, see Installation Guide > Other Dependencies.

Q: Why does torchvision report that lzma is missing even though the lzma module is installed?

Install the lzma module:

pip3 install backports.lzma

Go to the Python library directory. Python 3.11.4 is used as an example below.

cd /xx/xx/python-3.11.4/lib/python3.11

Modify lzma.py by changing the following:

from _lzma import *
from _lzma import _encode_filter_properties, _decode_filter_properties

to:

from backports.lzma import *
from backports.lzma import _encode_filter_properties, _decode_filter_properties

Docker and Quick Start

Q: Why can't the Docker container find test images stored on the host?

Cause: The host directory has not been mounted into the container, or TEST_IMAGE uses the host path instead of the path inside the container.

Solution: Add a volume mount when starting the container and use the mounted path inside the container:

docker run ... -v /path/to/testdata:/data ...
export TEST_IMAGE="/data/test.jpg"

For details, see Quick Start > Step 2: Starting the Container.

Q: Why can't the container access the NPU?

Cause: The device number specified by --device /dev/davinci* does not match the host device number, or the driver-related directories have not been mounted.

Solution:

1. Run npu-smi info on the host to ensure that the NPU is available.
2. Change the number in --device /dev/davinci0 to the actual device number (for example, davinci1).
3. Ensure that all driver mount items listed in quick_start are included in the docker run command.

Q: Why does a file permission error (0x102003EE) occur when reading a file?

Cause: The API requires the file owner to be the current user and the permissions to be no more permissive than 640.

Solution:

chmod 640 /path/to/your/file.jpg

Running and Troubleshooting

• Encountered an error code? See Appendix > Error Codes.
• Having environment variable issues? See Appendix > Environment Variables.
• Having issues with the installation steps? See Installation Guide > FAQ.