I have been learning about multi-GPU knowledge graph embedding training recently, and I saw Marius support multi-GPU training. But the configuration file stores relation embedding on CPU. I want to know if Marius can store relations on GPU. If yes, how does Marius store them? Does it duplicate relations on different GPUs and writeback to CPU global relations like DGL-KE, or store different parts of relations on different GPUs?
I have been learning about multi-GPU knowledge graph embedding training recently, and I saw Marius support multi-GPU training. But the configuration file stores relation embedding on CPU. I want to know if Marius can store relations on GPU. If yes, how does Marius store them? Does it duplicate relations on different GPUs and writeback to CPU global relations like DGL-KE, or store different parts of relations on different GPUs?