CUDA libcublas.so.11 Error when using GPUs inside ONNX Docker Container

Viewed 310

While programming with python3.6 on DGX Station (NVIDIA) based on ONNX runtime environment;

Using following libraries;

mxnet==1.5.x onnxruntime-gpu==1.7.x

I see following error

OSError: libcublas.so.11: cannot open shared object file: No such file or directory
1 Answers

Fixed it by updating CUDA to 11.0

apt update
apt install cuda-11.0
pip3 uninstall mxnet-cu92 //and reinstall mxnet-cu102
pip3 uninstall onnxruntime-gpu //and resinstall latest version

then a new error started appearing

onnxruntime cuda failure 100 no cuda-capable device is detected

then it was noted that docker was not started with nvidia runtime so following was added to the docker-compose file

runtime: nvidia

and all started working

Related