CUDA error code=35(cudaErrorInsufficientDriver) "cudaSetDevice(device)"

/usr/local/cuda-11.4/compat/libcuda.so
/usr/local/cuda-11.4/targets/x86_64-linux/lib/stubs/libcuda.so
/usr/lib/x86_64-linux-gnu/libcuda.so
root@42675e553dec:/home/nvidia/gitlab/nni/build# ll /usr/lib/x86_64-linux-gnu/libcuda.so*
lrwxrwxrwx 1 root root       12 Aug  1 06:44 /usr/lib/x86_64-linux-gnu/libcuda.so -> libcuda.so.1*
-r-xr-xr-x 1 root root   154088 Feb 26 10:56 /usr/lib/x86_64-linux-gnu/libcuda.so.1*
-rwxrwxrwx 1 1000 1000 22037736 Oct 19  2022 /usr/lib/x86_64-linux-gnu/libcuda.so.470.161.03*
root@42675e553dec:/home/nvidia/gitlab/nni/build#

$ nvidia-smi
Wed Feb 19 04:42:14 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 430.64       Driver Version: 430.64       CUDA Version: 10.1     | <- driver version
|-------------------------------+----------------------+----------------------+     is
                                                                                   insufficient for
Cuda compilation tools, release 10.2, V10.2.89 <---------------------------------- runtime version
Because CUDA 10.2 requires a 440.33.01 or newer driver, and your driver is 430.xx. Too old.

If you don’t need CUDA 10.2, the simplest solution is probably to install CUDA 10.1 without 
installing any new driver, and just keep your 430.xx driver. Toolkit runfile and package manager 
install types both have methods for installing the toolkit without the driver.

lix19937 / tensorrt-insight

CUDA error code=35(cudaErrorInsufficientDriver) "cudaSetDevice(device)" #37