Dynamic Resource Allocation lets Kubernetes match device claims to GPU properties. Here is what DRA changes for AI training and inference, and where it still falls short.