Bound shared accelerators with stage-specific admission

When two inference stages share one accelerator and only some requests need the second stage, give each stage its own bounded admission limit and queue. Reject excess work quickly rather than allowing a slow secondary stage to convoy the primary stage. This preserves predictable memory use and makes overload visible to the caller, which can fall back safely. Static hardware partitioning is usually unnecessary until measurements show interference that bounded scheduling cannot control.

Applies when: A fast primary model and conditional secondary model share one accelerator. Keywords: shared GPU bounded queue admission convoy overload inference scheduling