Adding Docker image to container registry

[VRehnberg]

Are there any plans on adding the AlphaFold3 container image to any container registry like DockerHub? I believe it could simplify the installation a lot for some users.

Would as a bonus also improve reproducibility.

Example use case

Users of an academic HPC compute cluster running Apptainer/Singularity, set-up so that users can install their own containers. Currently they would have two alternatives:

1. Installation as recommended:
   1. Install Docker on their own computer (possibly first need to install VM to install docker in)
   2. Build image according to suggestions here
   3. Convert to singularity/apptainer image
   4. Transfer to cluster and hope that it works
2. Direct singularity install
   1. Try to translate Dockerfile to Apptainer/Singularity definition file
   2. Try building, if it doesn't work go back to 1

However, if you added it to a container registry the step would instead be:

apptainer pull alphafold3.sif registry.io/deepmind/alphafold3:latest

[jwdelta]

I agree - installing on HPC clusters is challenging. The Triton package installation fails on virtual machines (both Ubuntu and Red Hat) running on M1 MacBooks, resulting in Docker build failures. I'm not sure about the reason. Maybe NVIDIA GPU and CUDA are also needed for this step, which MacBooks don't have.

I eventually succeeded using this guide: #23
My code is:

pip3 install spython
spython recipe docker/Dockerfile > docker/Singularity.def
apptainer build --fakeroot $APPTAINER_CACHEDIR/alphafold3.0.1.sif docker/Singularity.def

Hope this helps others installing AlphaFold3 on HPC.

[saadkhi]

Hello @VRehnberg, I want to work on this issue. Can you briefy which file I have to look upon for the enhancement

[VRehnberg]

@saadkhi build a docker image (files https://github.com/google-deepmind/alphafold3/tree/main/docker) and add to a container registry such as https://hub.docker.com/ under an official account for AlphaFold3 (perhaps https://hub.docker.com/u/google or perhaps you want a separate account for google-deepmind).

After that, this section https://github.com/google-deepmind/alphafold3/blob/main/docs/installation.md#installing-docker could be very much simplified.

[serenalotreck]

@saadkhi @VRehnberg any progress on this front? My local machine doesn't have enough RAM to build the image myself, and the unofficial versions available on DockerHub are out of date with the current code, which causes the issue in #361

[shivamgcodes]

Hey, i have built and uploaded the docker image, pls let me know if any changes are required with the naming / tagging or anything else. - https://hub.docker.com/repository/docker/shivamgupta579/alphafold3/general .

Also, please let me know if i should do any further changes or raise a PR.

[VRehnberg]

The main challenge with this issue is to make it an official image, something endorsed by the developers so that it can be trusted without looking through the recipe in detail. Ideally also then a new image with every release (as part of CI/CD or manual effort).

@shivamgcodes it seems to be private, I can't view it at least. If you want to share it with others you have to make it public. As for tagging I can't see what you have but I believe customary is a tag for the specific version you built (if not a release version include date and/or first part of commit hash).

@serenalotreck guessing its the hmmer build that has the most memory requirements. You can probably reduce this by building it with less processes, so instead of make -j do make -j 1 or just make. Or just doing apt-get install --yes hmmer instead of the manual build of hmmer.

[shivamgcodes]

I am not sure why its showing up as private, it shows the visibility as Public to me. I will change the tag to reflect the release version.
I would love to add that to the CI/CD pipeline! Should i create an issue for that?

Also, can you point me to some basic tests, or like a 'hello world' of alphafold3 world, so that i can validate the docker images, at least on a personal level

[VRehnberg]

Ah, it was just an issue with the link. https://hub.docker.com/r/shivamgupta579/alphafold3/tags works.

Also, can you point me to some basic tests, or like a 'hello world' of alphafold3 world, so that i can validate the docker images, at least on a personal level

The files run_alphafold_test.py and run_alphafold_data_test.py sounds promising.

Should i create an issue for that?

I should clarify, I'm not in any way affiliated with DeepMind or the AlphaFold team. I'm just a guy that wanted something and created an issue asking for it.

[shivamgcodes]

sorry for the wrong link before, and thanks for putting the right one here.

Thanks for clarifying, i did not know about that, maybe i will wait for someone from the team to tell me about the next steps. I hope you find the docker image useful.

[saadkhi]

@saadkhi build a docker image (files https://github.com/google-deepmind/alphafold3/tree/main/docker) and add to a container registry such as https://hub.docker.com/ under an official account for AlphaFold3 (perhaps https://hub.docker.com/u/google or perhaps you want a separate account for google-deepmind).

After that, this section https://github.com/google-deepmind/alphafold3/blob/main/docs/installation.md#installing-docker could be very much simplified.

Yeah, I made and added but any error related to memory overload encountered me so I back off. If guidance will be provided, I'll back to solve where I left.

[jurgjn]

Docker images for all releases, and a recent checkout. No guarantees, but we do use these ourselves..