llama.cpp Vulkan tops ROCm on Strix Halo

Hot take: this looks less like a fluke and more like ROCm still having rough edges on gfx1151, especially for MoE inference paths.

• –The gap is concentrated in generation, which is the metric users feel most in interactive and agentic workloads.
• –The same binary, model, and hardware stack reduce the chance this is just setup noise.
• –The lower std dev on Vulkan suggests a more consistent kernel path, not just a one-off speed burst.
• –This lines up with broader Strix Halo chatter that ROCm support is improving but still uneven versus Vulkan on some workloads.
• –It is still a single benchmark on one model; dense models and longer contexts may show different results.