What are the primary bottlenecks in multi-GPU training and inference, and how do you configure your infrastructure to overcome them at Decagon?