Open ylx6266 opened 2 years ago
How large is the structure you are trying to predict?
How large is the structure you are trying to predict?
The complex I want to predict consists of a protein with 872 residues and a DNA with 12bp.
I had the same problem, and my GPU mem is 8GB. Does it have to run on a multi-GPU server?
I have encountered this same issue. It was running on a 3090 GPU with 24GB mem.
Running on a GPU with more memory solved the problem for me when I encountered this error.
Running on a GPU with more memory solved the problem for me when I encountered this error.
I forced this program to use CPU, which can avoid this error, but it was too slow.
As others have pointed out, for large complexes the memory requirements might be high, using a higher memory card (or CPU only) may be necessary.
I have a memory-optimized version I will try to push in the next couple of weeks (need to make sure results are the same).
More of a FYI: I'm getting CUDA out of memory
for a complex of a 505 aa dimer (1010 aa in total) with a 76 bases RNA molecule (x2 = 152 in total) on an NVIDIA V100 with 32GB memory, which is somewhat disappointing as I was hoping to also model considerable larger complexes than that in the future.
If anyone is encountering this error, I just solved it yesterday by:
export 'PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:256'
and then running the script as usual. I have a 16GB GPU and trying to solve for 1030nts long RNA structure.
Thanks,
Kamil
met the same problem, hope a new version of multi-gpu in the future
FYI I tested a similar system to the one described by @mf-rug (2x500aa + 76b RNA) on a 16GB GPU, using @kcygan recommendation of setting the PYTORCH_CUDA_ALLOC_CONF environment. I still got the CUDA out of memory error.
I also tried suggestions from stackoverflow
Alas none of these helped.
However running predict.py on 8 CPUs with 64GB memory on an HPC completed in ~ 5hrs. Good enough for me in this instance.
Hi same problem here, I have the following card NVIDIA GeForce RTX 2080 Ti , I know 11Mb is not a lot, but maybe there is a workaround?
I have many CPUs', but I am unsure how to make the modeling itself with CPUs, can anybody give accurate instructions if this is possible?
Thanks!! Fabian
Trying to solve a big protein/RNA complex...
If you're struggling to run RosettaFold2NA locally, feel free to try https://www.tamarind.bio/rosettafold2na. Tamarind is an online platform for bioinformatics tools that offers structural biology workflows, including RoseTTAFold2NA, for free.
RuntimeError: CUDA out of memory. Tried to allocate 1.54 GiB (GPU 0; 10.75 GiB total capacity; 6.28 GiB already allocated; 1.44 GiB free; 7.32 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF