Hello,
To control which GPUs will be made accessible inside the container, should we use
NVIDIA_VISIBLE_DEVICES or CUDA_VISIBLE_DEVICES ?
Are there similar variables or not at all ?
Is NVIDIA_VISIBLE_DEVICES to be used by admin when providing container and let CUDA_VISIBLE_DEVICES available for user ?
Regards,
Bernard
have you got any answer about this?
Good morning, No I didn’t got any answer.
What I can say is that I my case, running Docker container, I set both NVIDIA_VISIBLE_DEVICES and NVIDIA_VISIBLE_DEVICES with GPU-IDs to control the GPUs available inside the container,
That sounds to do the trick, even if I change the NVIDIA_VISIBLE_DEVICES and/or NVIDIA_VISIBLE_DEVICES values inside the container, I have always the same GPUs as initially specified = exactly what I want. (Tested by using export …+ nvidia-smi )
Regards