Docker image missing cuda libs about alphafold

[carbocation]

I am also encountering the same warning message. For additional context, I am attempting to run this on a GCP machine with the nvidia-gpu-cloud-image-20200629 boot disk image, with two nvidia-tesla-v100 GPUs attached.

[YoshitakaMo]

I see libcusolver.so.10 (a different version than listed in the warning), in the 11.0-runtime image (the Dockerfile is based on 11.0-base). And cudnn would require an additional install step.

See #11. Installation of libcudnn and libcusolver is not included in the Dockerfile.

[chrisroat]

Thanks for pointing me to the duplicate. It looks like #11 is an error/crash and using a different cuda version then provided in the repo (11.1 vs 11.0).

In our setup, with the provided 11.0 image, I get the above reported warnings, but the process still completes.

I hope DeepMind wants the world to be able to reproduce what they did, so I'm curious what their recommended solution is.

[carbocation]

I am also curious as to whether people have had success getting the released Dockerfile to work with GPUs within a CUDA 11.0-based environment.

So far, I have been able to get this Dockerfile working without GPU support (at least, with the same "Skipping registering GPU devices" warnings noted above). I have also gotten this to work without GPU warnings by upgrading the Dockerfile to CUDA 11.1. But ideally I would like to run this with GPU support in a CUDA 11.0-based environment, which seems like it should be possible given the description of this release.

[chrisroat]

Given the requirement of tensorflow-2.5.0 and the TF compatibility matrix, it seems it should be using cuda-11.2.

The solution I am now trying is #36 , which updates the base image to use nvidia's cuda 11.2 runtime, which brings in the missing libraries and uses versions in agreement with the compatibility matrix.

[tfgg]

Hi,

Thank you for the report and discussion. Please see if acf25fc fixes this issue and reopen if not. This change moves to using the cudnn base image and works in testing (though we also didn't encounter this issue in testing, so I'm not sure exactly what's causing it).