search

steineggerlab / foldseek

Foldseek enables fast and sensitive comparisons of large structure sets.
https://foldseek.com
GNU General Public License v3.0
822 stars 101 forks source link

cmake with CUDA fails #324

Open mohmhm1 opened 3 months ago

mohmhm1 commented 3 months ago

Expected Behavior

Foldseek compiled

Current Behavior

Fails during make step

Steps to Reproduce (for bugs)

cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_CUDA=1 -DCUDAToolkit_ROOT=/usr/local/cuda .. make -j 16

Foldssek Output (for bugs)

https://gist.github.com/mohmhm1/12e59b073256287283922b457f6bf486

Context

Trying to build from source for GPU acceleration on Ubuntu 22.04, CUDA 12.3, NVIDIA A10. The build fails with the errors attached in gist.

milot-mirdita commented 3 months ago

We are re-engineering the GPU integration currently. We hope to release the updated version next month that should resolve most of the current GPU integration weirdeness.

I'd recommend to use the CPU only ProstT5 version for now. It obviously slower, but not unusable so. Or use the PyTorch ProstT5 version at https://github.com/mheinzinger/ProstT5

mohmhm1 commented 3 months ago

Thanks for the update!

I would tried to run cpu based createdb using prosT5 weights on a 6M seq FASTA but it's very slow. Was thinking i can distribute the jobs across a fleet of instances and then combine the db after. Any advice on if the merge is possible?