Running AlphaFold on models larger than AMD GPU available memory

[dipietrantonio]

Dear developers,

I am facing an out-of-memory issue (follow the link for a detailed log) when running ColabFold (which is a repackaging of AlphaFold) on AMD GPUs.

I have found a comment on this repository that suggests GPU managed memory can be used to leverage host memory together with GPU memory when the latter is not enough to hold the dataset. However, the same configuration does not work on AMD GPUs.

Do you have any suggestions on how to overcome the memory issue?

Thank you!

[cgseitz]

What have you tried so far? When I faced an issue like this, I used the reduced_dbs flag, which reduced the GPU memory required and still generated useful outputs for my project. See here

[dipietrantonio]

We are working with AMD because their JAX implementation did not support Unified Memory to take advantage of the host memory.

[smilenaderi]

@dipietrantonio Can you please share how were you able to run alphafold on AMD GPUs? Thanks

[dipietrantonio]

@smilenaderi we can only run it on small problem sizes, the ones that do not require more memory than it is available on the GPUs. We are using it through the ColabFold repackaging. Here is the Dockerfile:

https://github.com/PawseySC/pawsey-containers/blob/cdp-colabfold/colabfold/colabfold.dockerfile