Look into singularity images with higher libtorch/cuda-toolkit support
Closed this issue · 2 comments
cissieAB commented
Move all the stuff into a container should make things easier.
nathanwbrei commented
Yes please! Are you documenting anywhere the horrible constraint propagation graph? I.e. what compiler version/language standard do we need for each choice of libtorch, CUDA, and GPU hardware?
cissieAB commented
https://github.com/nathanwbrei/phasm/blob/cissie_pinn/docs/farm_guide.md
Above it's a guide on installing all the stuff on the farm. For this guide
- CUDA: 11.4.2
- libtorch: libtorch-shared-with-deps-1.12.0+cu113.zip
- cudnn: 8.6.0
- gcc/10.2.0, cmake/3.21.1, default pin/dwarf versions
I did not include the correct JANA version (2.6.0) in the above guide. Will fix it later.
The docker image can provide higher CUDA/libtorch versions than the above