2

I am trying to use the docker image with new drivers >460 for CUDA 11 but my host machine has cuda-10.

According to the image below (https://github.com/NVIDIA/nvidia-docker) docker uses the drivers present in the host hence I am getting error as kernel mismatch required 450+ kernel version 418.

I have installed the drivers on my container. I there any way to use the bare metal gpus from the container itself and bypass the host drivers??

enter image description here

talonmies
  • 70,661
  • 34
  • 192
  • 269
vizakshat
  • 191
  • 1
  • 8
  • 2
    In a word, no. CUDA must use the driver installed on the host. There is a CUDA compatibility method to do something like it, but it has a variety of requirements that have to be satisfied. [This question/answer](https://stackoverflow.com/questions/57015346/can-i-run-a-docker-container-with-cuda-10-when-host-has-cuda-9/57023447#57023447) provides a summary, your question is arguably a duplicate of that. one. – Robert Crovella May 29 '21 at 17:51
  • 2
    with a more advanced container orchestration system like kubernetes, its possible to update the GPU driver in the host from an [operator](https://github.com/NVIDIA/gpu-operator) However this doesn't change the fact that a running container that is using the GPU depends on the GPU driver installed in the host machine. – Robert Crovella May 29 '21 at 18:10

0 Answers0