Faster distributed GPU training with Reduction Server on Vertex AI