VLA on Raspberry Pi?

I'm GPU-poor and I'm not ashamed to admit it. I took @HuggingFace SmolVLA, a 500M-parameter robot model people normally run on GPUs, and squeezed it onto boards that don't cost a metaphorical arm and a leg. 60 seconds of inference per action chunk down to 1.7 seconds, real-time on Radxa Orion O6. Almost the same time (1.8 seconds) on @raspberrypi Pi 5. Sources and Links: https://huggingface.co/blog/smolvla https://huggingface.co/HuggingFaceVLA/smolvla_libero https://github.com/VinRobotics/vla.cpp https://github.com/AIWintermuteAI/vla.cpp/pulls (not organized PRs) https://radxa.com/products/orion/o6n/ (not the same as I sued, but same CPU and cheaper) https://www.youtube.com/watch?v=JnldWEexR10