ValueError: Cannot create a tensor proto whose content is larger than 2GB. about alphafold HOT 4 CLOSED

Hi, how big is the protein you’re trying to inference? Unfortunately there’s no easy workaround for this but if the large inputs are the result of a long input sequence you will probably find that the model inference stage would fail too due to the GPU running out of memory.

from alphafold.

Hi,
Thanks for the reply. It is quite big, ~2700 residues. If there is no tweak on the tf side, then I might have to be more restrictive with the sequence search/alignment step.
Still, if that is the case, you might want to consider an if else statement checking for the length of the alignments, right?
I will just close this then, but if you have any additional suggestion, please just let me know.
Cheers,

from alphafold.

Unfortunately it is a hard limit on the TF side indeed. That’s also right around the upper limit of what we ran for the AlphaFold database (see the second FAQ here: https://alphafold.ebi.ac.uk/faq) so you may still have luck fitting it onto a GPU but it might be a tight fit. We generally see MSA construction working at that size so it’s unlikely to be very far over the limit — removing just a few residues might make the sequence fit. You could also try predicting shorter overlapping subsequences — that is the approach we took for sequence lengths over 2700 for the database release. One final suggestion, if you haven’t already tried it, is to run with --preset=reduced_dbs which should reduce the size of your MSA. This is the setup that was used to produce the AlphaFold database predictions.

from alphafold.

Cool, I will give it a shot. Thanks!

from alphafold.