Inquiry about the progress of the inclusion of rocm options to dgl

I would like to know if the plans to add rocm version to dgl are on the table. torch and pytorch geometric have included the option to use amd gpu.

I would like to know if that architecture option available by torch would be able to be used in dgl at some point because right now gives errors.

Sorry and thank you for your patience

Hi, the team currently does not have plan to add ROCM support. The main bottleneck is the lack of relevant expertise for long-term maintenance. If you know someone who would be interested in working with us, please let us know.

Thank you for your response. Regretfully I am working on my phd right now, and I was wondering about the progress for using it on research.

Thank you so much, maybe the progress given the time left would be using other libraries for the current experiment.

I could only help from time to time. Good luck with your other future endeavours.

In regard with someone to help, I don’t know experts in the field. Torch uses the same interface and devices as the one in cuda. Also pytorch geometric used little transformations to their interface, the repo from looong in their page have littles changes that allow the use of rocm. Maybe this can be of help.

One question, do you know a way of using dgl with my amd card, without needing to recompile dgl with external libraries as hipify?

Sorry for being so insistent, but I have only at home a amd card. I use a nvidia card at work, but I would like to paralelize experiments.

Sorry and thank you for your time.

DGL has no AMD card support. You might want to look into projects such as ZLUDA so that the compiled CUDA binaries can be run on your AMD card. However, it is not guaranteed to work.