I only changed the runtime to TPU and it worked perfectly for me. If dealing with a dense graph data especially with language model embedding for your data.x features, it increases the data size and TPU has the best RAM capacity to rescue you.