How to build DGL from source for the NVIDIA 5070 Ti (CUDA 13) and Pytorch 2.10.0

Hi everyone,

I have got DGL to build from source for the NVIDIA 5070 Ti and CUDA 13.0 with Pytorch 2.10.0. This comes with some help from ChatGPT and a lot of tinkering and refining and so I thought I would document this here for the benefit of anyone else following suit - I see many open issues here with similar requests with no replies.

Note:

• My end goal has been to get RFantibody that depends on DGL working so some of the choices I have made (such as turning off HugeCTR) are specifically targeted with this in mind. You may need to tinker some more for your use case.
• I have built this on Debian Stable.

1. Deviations from DGL source:

a) Patches: I have patched the following files :

Patch CMakeLists.txt in the root DGL folder with the following: CMakeLists.txt.patch

Patch cmake/modules/CUDA.cmake with the following: CUDA.cmake.patch

b) I am also using GKlib and METIS from their source repo and not the ones provided along with DGL - the CMakeLists.txt changes take care of that.

c) I have disabled the following files (by renaming them with a .disabled suffix) : ./src/graph/metis_partition.cc.disabled ./src/graph/transform/metis_partition_hetero.cc.disabled

• the above may no longer be necessary since we are building and installing Metis and Gklib from source. However, I haven't tested things with these enabled once I got things working. You can let me know your experience if you choose to keep these on :)

1. Build Steps :

Having done the above, I basically followed the following steps

• Followed https://www.dgl.ai/dgl_docs/install/index.html#install-from-source to create a conda environment for the GPU
• Activated environment
• Installed Pytorch with CUDA in this environment - pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
• Build and install GKLib shared library from source in /usr/local - https://github.com/KarypisLab/GKlib make config shared=1 prefix=/usr/local
• Build and install Metis shared library from source in /usr/local - https://github.com/KarypisLab/METIS.git make config shared=1 prefix=/usr/local
• Back in DGL source directory
• mkdir build && cd build
• cmake .. -DUSE_CUDA=ON -DUSE_OPENMP=ON -DBUILD_TORCH=ON -DUSE_METIS=ON -DBUILD_HUGECTR=OFF -DCMAKE_BUILD_TYPE=Release -DCMAKE_POLICY_VERSION_MINIMUM=3.5
• In build directory run make
• Then followed the DGL document to setup python cd ~/code/dgl/python pip install -e .
• Added dgl/build to LD_LIBRARY_PATH in ~/.bashrc
• Test with python -c "import dgl"

Hope the above helps :)

Regards, Jeetu