search

google-deepmind / alphafold

Open source code for AlphaFold 2.
Apache License 2.0
12.93k stars 2.29k forks source link

OOM error for 4000 length monomer strucuture prediction #824

Open CHUht opened 1 year ago

CHUht commented 1 year ago

Hello,

I’m trying to predict monomer structure of length 4000 protein on 80G A100 NVIDIA GPU, the readme says the maximum length of prediction is more than 5000, while an OOM error occurs as below, are there some special settings need to be set to run prediction of length 5000 protein?

‘’‘ log BufferAssignment OOM Debugging. BufferAssignment stats: parameter allocation: 3.25GiB constant allocation: 38.6KiB maybe_live_out allocation: 4.18GiB preallocated temp allocation: 77.25GiB preallocated temp fragmentation: 976B (0.00%) total allocation: 84.69GiB total fragmentation: 4.10GiB (4.85%) '''

Environment settings: Alphafold : commit id 3f31725591b0c6b0b6d63214235d3abe6b81742c, Mon Mar 27 jax: 0.3.25 jaxlib: 0.3.24+cuda11.cudnn805 80G A100 NVIDIA GPU

Htomlinson14 commented 1 year ago

Hi! Are you using the docker image?

CHUht commented 1 year ago

I‘m using without docker, the docker script is not supported on my machine. Is there any difference between with and without docker running script?

Htomlinson14 commented 1 year ago

These environment variables should be helpful: https://github.com/deepmind/alphafold/blob/main/docker/run_docker.py#L245

The reason for my question was just to check these are being set. It should be possible to run without the Docker image, but in this case you will need to ensure the runtime environment has been set up correctly.