Gustaf Ahdritz

Could you share the full error trace?

GPU memory is the bottleneck for AlphaFold2/OpenFold, not system RAM. I suspect that's what's causing your crash. Could you report which GPU you were assigned in your Colab session? In...

Do you have `gcc` and `nvcc` on your PATH? You need both in order to run `setup.py` and train the model. On many systems, you can just run `module load...

During inference, whether the rigid object has grad enabled doesn't matter since grad computation is disabled globally.

Are you using the `multimer` branch? That's still experimental (we'll get to this eventually), and so small discrepancies like this are to be expected.

Are you regenerating PDB alignments? There's no need to do that; we've pre-computed them all. See the RODA repository linked in the README.

OpenFold is based on PyTorch, not TensorFlow, and doesn't recognize those flags. PyTorch does not have CUDA Unified Memory support. However, we do have a bunch of optimizations that make...

Sorry for the delayed response---I haven't seen this particular issue. I'll leave this open in case anyone else has. In any case, you can always just train without DeepSpeed---simply remove...

I believe this might cause issues with the custom CUDA kernels, and it will definitely cause a slowdown, but it's worth a try.

Hm. PDB70 should be downloaded by `download_alphafold_dbs.sh` and the ColabFold database by `download_mmseqs_dbs.sh`. If those two didn't end up in your `data/` directory, the downloads must have failed for some...