Host libraries for nvidia-container-runtime

[jeroen.torrekens]

Whenever you start a container and use the nvidia-container-runtime, there are libraries and files mapped into your container automatically. The configuration is located in /etc/nvidia-container-runtime/host-files-for-container.d/:

root@ubuntu:~# ls /etc/nvidia-container-runtime/host-files-for-container.d/
cuda.csv  cudnn.csv  l4t.csv  tensorrt.csv  visionworks.csv

For some containers cuda.csv is used, for others the l4t.csv is used, of others none of these are used. I was wondering if anybody knows what the decision making process is here. What config is used for which container? Documentation and insert your favorite search engine came up empty handed.

Thanks in advance!

[dusty_nv]

@jeroen.torrekens

For some containers cuda.csv is used, for others the l4t.csv is used, of others none of these are used. I was wondering if anybody knows what the decision making process is here. What config is used for which container?

Hi @jeroen.torrekens, for JetPack 4.x all of the .csv files should get used by --runtime nvidia (as long as the container is derived from l4t-base). On JetPack 5.x, only l4t.csv gets used (because CUDA/cuDNN/TensorRT/ect are installed inside the containers on JetPack 5 for portability). Does that answer your question?

[jeroen.torrekens]

Hi dusty_nv! That indeed explains the behavior we’re seeing. Thanks a lot!