I wanted to share that AMD has sponsored a port of DGL to run on AMD GPUs using HIP. It’s available at GitHub - nod-ai/dgl at hip-ready (note you want the hip-ready or hipify-inplace branch. The master branch is just a mirror of upstream). It still has some rough edges, but we’ve gotten ~all the C++ and Python unit tests passing for the PyTorch backend on CDNA2 and CDNA3 architectures as well as running some models like RFDiffusion. Those interested, please give it a try and report any issues you run into.
See also https://github.com/dmlc/dgl/issues/7838
Tagging people who’ve asked about this previously in Running DGL with ROCm (AMD GPUS) and Inquiry about the progress of the inclusion of rocm options to dgl in case you’re still interested: @paulamartingonzalez, @rafasupremoyo