When I read the code distributed training, I found that when doing graph partition, the input graph is converted to homogeneous graph in the first place, and when using DistGraph load the partition graph, the loaded graph is also homogeneous graph, even though the input training graph before partition is hetero. I am confused that why it is necessary to convert the input graph into homogeneous graph but not keey using the original type? Is there any limitation that only homogeneous graph work during the distributed training?
Thank you:)