New model ready for testing!

[github-actions[bot]]

This model is ready for testing. If you are assigned to this issue, please try it out using the CLI, Google Colab and DockerHub and let us know if it works!

[miquelduranfrigola]

@GemmaTuron this model has been incorporated by @DhanshreeA. It has relevance for the GRADIENT project. I therefore assign it to myself too.

[miquelduranfrigola]

Works for me in Codespaces.

However, upload to DockerHub has failed? https://github.com/ersilia-os/eos7d58/actions/runs/7828896799/job/21359729719

Importantly - we should not be getting a "New model ready for testing" issue if DockerHub upload is not yet complete... @DhanshreeA , I am opening a separate issue correspondingly to deal with this: https://github.com/ersilia-os/ersilia/issues/959

[GemmaTuron]

nice @DhanshreeA ! Is it a memory issue?

[DhanshreeA]

nice @DhanshreeA ! Is it a memory issue?

@GemmaTuron it appears to be that. The checkpoints are all very very small. One checkpoint .pt file is 1.8 MB, so all 10 of them make up 18MB total. The admet_ai repository itself is also very small, 38MB, however that's not a good indicator for anything since it doesn't account for dependencies. The fun thing is that the conda environment for model eos7d58 on my machine is 5.7GB. I've noticed that Chemprop comes with a bunch of dependencies which definitely makes for a bloated environment every time we use a model based on it. And this is on top of conda environments being pretty large to begin with.

I'll build the Docker image locally and see how much space that's taking up and there's a few things we can do from there:

One of course is to push the image manually, and for a longer term solution, I'll dig into reducing the size of conda environments/conda based Docker images.

[miquelduranfrigola]

Thanks @DhanshreeA - good stuff. I completely agree that the focus should be on reducing size of conda environments for a longer term solution. Note that a related issue is already open with @mjwarren3, perhaps we can join efforts: https://github.com/ersilia-os/ersilia/issues/900

[GemmaTuron]

Hi @DhanshreeA and @miquelduranfrigola !

I've used @HellenNamulinda edit to the workflow file to clear the disk and it worked. Should we incorporate it in all models?

[HellenNamulinda]

I also agree with @DhanshreeA on reducing the size of conda environments. Some of these models have docker images that are abnormally big, especially for the amd64 architecture. Forexample the recent model eos18ie, amd64 is 7.14GB, yet arm64 is 1.93GB

I also think the machine used to build the docker image has a great impact on the size of final image. I will explore this.

[GemmaTuron]

The model works nicely! I'll close this issue, and we can continue discussing about Docker sizes in the related issue