search

rapidsai / ci-imgs

https://hub.docker.com/u/rapidsai
12 stars 16 forks source link

Use Conda to install CUDA Toolkit (for 12.x) #174

Open jakirkham opened 4 months ago

jakirkham commented 4 months ago

Now that the CUDA Toolkit is fully supplied via Conda packages starting with CUDA 12.0, there is no need to depend on the nvidia/cuda images (except for CUDA 11.x)

Given this, it might make sense to change the FROM logic here to simply use the Linux distro image we would like for CUDA 12.x

https://github.com/rapidsai/miniforge-cuda/blob/0e20886991c4fcb178d73bd0f1cf3224d1b0e0a8/Dockerfile#L3

The CUDA 11.x case would still need the nvidia/cuda images

jakirkham commented 4 months ago

@ajschmidt8 do you have thoughts/preferences on how we go about this?

ajschmidt8 commented 4 months ago

no thoughts. just that the change should be broadcasted loudly in case there are any unexpected issues.

vyasr commented 3 months ago

We still need a container with the CUDA driver installed, so we would presumably have to install that manually into the image.