Couldn't use amber in colabfold_batch, after running colabfold_search

[lzhangUT]

Hi, I have followed all the steps on your GitHub using pip install (not localcolabfold), and everything worked fine. I downloaded all the data to my machine, and was able to run colabfold_search for MSA. after MSA search, I then do colabfold_batch to get he prediction, if i use this:

This needs a lot of CPU

colabfold_search input_sequences.fasta /path/to/db_folder msas

This needs a GPU

colabfold_batch msas predictions

it worked fine. however, I want to add some arguments to the colabfold_batch for the second step, like amber. then I did colabfold_batch --amber --use-gpu-relaxation msas predictions the error is shown below: it seems like model is running okay, but the relaxation can not be run.

any help would be greatly appreciated.

(edited: please don't tag random people)

[milot-mirdita]

Can you please run:

python3 PATH_YOUR_CONDA_ENV/lib/python3.8/site-packages/simtk/testInstallation.py

[lzhangUT]

this is what I get, along with my cuda version. Thanks for your help.

[lzhangUT]

and also cuda diver verson:

[milot-mirdita]

Can you try installing localcolabfold, that might fix the cuda issue with openmm

[lzhangUT]

The same issue/error occurs, after I install localcolabfold. anything with the minimization.py?

[milot-mirdita]

This looks like an issue with your CUDA installation/your gpu. The testInstallation script above should not show any cuda errors. I would recommend to stop the gpu relaxation parameter for now.

[lzhangUT]

But while I was doing colabfold_batch msas predictions command, I checked my GPU usage, it seems like it was fully used at that time. If CUDA is the issue, the model inference would not use gpu at all, right? If I am understanding correctly. That’s also another question I have, I have a test fasta with 545 aa, I run the msa search first, then do the colabfold_batch msas predictions command. Each model takes about 450s to finish, does that sound reasonable, was it using the GPU? BTW, the GPU  I am using is  a nvidia p40 GPU (will use another one later, right now just for testing colabfold).    From: Milot MirditaSent: Tuesday, July 26, 2022 11:21 PMTo: sokrypton/ColabFoldCc: lzhangUT; AuthorSubject: Re: [sokrypton/ColabFold] Couldn't use amber in colabfold_batch, after running colabfold_search (Issue #275) This looks like an issue with your CUDA installation/your gpu. The testInstallation script above should not show any cuda errors. I would recommend to stop the gpu minimization parameter for now.—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***> 

[lzhangUT]

It is like this while running colabfold_batch    From: Milot MirditaSent: Tuesday, July 26, 2022 11:21 PMTo: sokrypton/ColabFoldCc: lzhangUT; AuthorSubject: Re: [sokrypton/ColabFold] Couldn't use amber in colabfold_batch, after running colabfold_search (Issue #275) This looks like an issue with your CUDA installation/your gpu. The testInstallation script above should not show any cuda errors. I would recommend to stop the gpu minimization parameter for now.—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you authored the thread.Message ID: ***@***.***> 

[milot-mirdita]

the prediction step and relaxation step have completely separate implementations. You might be running the prediction successfully on the GPU but then facing issues with the relaxation stage.

[lzhangUT]

Okay, thank you very much for your quick reply. You mentioned ‘ I would recommend to stop the gpu minimization parameter for now.’ how do I implement that in the command, colabfol_batch —amber msas predictions —cpu? I don’t see other arguments for minimizing parameters. Thanks

-Li

On Jul 27, 2022, at 00:11, Milot Mirdita @.***> wrote:

 the prediction step and relaxation step have completely separate implementations. You might be running the prediction successfully on the GPU but then facing issues with the relaxation stage.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.

[milot-mirdita]

Don't use the --use-gpu-relaxation parameter.

The --cpu parameter will cause alphafold to run on CPU instead of GPU and thus be extremely slow.