Update service base image

[elisabettai]

The images fail to build in the CI with this error:
docker.io/nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04: not found

Because this image doesn't exist anymore: nvidia/cuda:11.2.1-cudnn8-runtime-ubuntu18.04 (see Dockerhub list here).

Should we go with nvidia/cuda:11.2.2-cudnn8-runtime @alessandrofasse?

Here's the table of compatibilities for tensorflow and tensorflow-gpu: https://www.tensorflow.org/install/source#gpu
Is there something similar for pytorch?

[alessandrofasse]

I think upgrading is a good idea. There we a lot of changes. With respect to pytorch such a table does not exists right away. However, I checked on both websites.

If use the image

11.8.0-cudnn8-runtime-ubuntu22.04

we could install the most recent versions of tensorflow in version 2.14.0 and pytorch in version 2.1.0 (https://pytorch.org/get-started/locally/)

How does that sound?

[elisabettai]

It sounds good @alessandrofasse, thanks. The biggest question a think is to figure out if users that have scripts running with the current versions will be able to use them as they are if we upgrade.

Once we answer that, we can decide if we go with a new patch service version or a minor/major one.

[alessandrofasse]

But I think all our containers are version controlled? I.e. others can still use their scripts and if they update they need to fix the scripts

Yes. How it works for updating services in oSPARC is the following:

• We update the patch version of the service -> Users can update their containers in one click and they should expect to be able to run their scripts as they were before.
• We update minor/major version of the service -> Users cannot expect that their scripts will still work.