do small edge boxes make sense with llama.cpp

[visionforge]

I think the bottleneck is somewhere else. I've seen people blame VRAM headroom when the real issue was much more mundane. Worth testing the simpler explanation before adopting the more interesting one.