The docker image is too large

[JudithBernett]

Describe the bug

The runner is running out of space and therefore, the Github action is no longer working. We should think about how we can reduce the size of the image

[nictru]

I think a different strategy could be adding this to the docker publish workflow

[nictru]

Sorry for closing, misclick

[JudithBernett]

I was just thinking about it, and I think the action is actually no longer needed because Drevalpy is now available on Bioconda. Then, biocontainers should automatically make a container, right? One can also make the Seqera container then. However, this container is huuuge because it does not have the slim build, so I still don't know what we should do

[nictru]

The biocontainers container seems to be 674MB, which does not seem overly huge if it also contains all the CUDA stuff

[JudithBernett]

damn that's true … but apparently the seqera image 12GB? I have to investigate this all

[nictru]

I guess this is actually a central problem here:
The build that is performed as part of the docker publish workflow is very large, because it is based on the poetry.lock file. There, the exact versions of e.g. pytorch including CUDA is defined and installed. When installing the drevalpy pip or conda package, it automatically installs the variant (cpu/cuda) that matches the device that the build is performed on.

Since both biocontainers and seqera containers are built on devices without GPU, it installs the CPU-version of the pytorch dependencies.

damn that's true … but apparently the seqera image 12GB? I have to investigate this all

If the seqera image is 12GB then maybe it actually installs the CUDA dependencies, but still my insights from above might be relevant for you