The separation between scheduling decisions and binding is particularly interesting. It improves resilience while making GPU allocation and scheduling failures easier to trace.