llama.cpp Windows slowdown stumps AMD users

This looks less like a llama.cpp bug than a bad interaction between Windows, AMD GPU drivers, and backend state in the runtime. Restarting llama.cpp and the graphics driver not fixing it suggests the slowdown is happening below the app layer, likely in driver state, memory placement, or power management behavior. The Windows-versus-Linux split is the strongest signal here: same hardware, same models, different outcome. Related llama.cpp discussions have already pointed at Windows ROCm and shared-memory quirks as well as backend-specific regressions, so this fits a broader pattern of brittle AMD-on-Windows inference behavior. The fact that multiple models and context sizes hit the same ceiling makes a model-specific optimization bug unlikely, and it is another reminder that stable throughput matters as much as peak tok/s.