Nvidia-docker: NPP libraries aren't mount for container with capabilities=compute
This is a feature request issue.
1. Please add NPP libraries to nvidia/cuda:10.2 image
CUDA libraries are mounted within nvidia/cuda:10.2, but NPP libraries are not, despite that they are shipped with CUDA toolkit.
2. Repro:
docker run --gpus 'all,"capabilities=compute,utility,video"' -it nvidia/cuda:10.2-base /bin/bash
>find / -name libcuda.so
/usr/lib/x86_64-linux-gnu/libcuda.so
>find / -name libnpps.so
Not found
rarzumanyan
All 2 comments
Why was this closed? Is NPP now present in the container?
zanbri
on 28 Apr 2020
@zanbri since @rarzumanyan works at NVIDIA, I reached out to him to explain.
So, this is normal, NPP is a library that is not tied to your device driver version; hence it is not mounted from the host and instead should come from inside the container image.
You might want nvidia/cuda:10.2-devel or nvidia/cuda:10.2-runtime if you want NPP inside the container image.
flx42
on 28 Apr 2020
👍3
Was this page helpful?
0 / 5 - 0 ratings
Related issues
brandon-northcutt
路
34Comments
jwatte
路
35Comments
cancan101
路
25Comments
JingL1014
路
25Comments
scosaje
路
44Comments
Most helpful comment
@zanbri since @rarzumanyan works at NVIDIA, I reached out to him to explain.
So, this is normal, NPP is a library that is not tied to your device driver version; hence it is not mounted from the host and instead should come from inside the container image.
You might want
nvidia/cuda:10.2-develornvidia/cuda:10.2-runtimeif you want NPP inside the container image.