Kuma: compiling PyTorch models into self-contained WebGPU executables [P]

I've been experimenting with a compiler/runtime project that I'm not entirely sure is a good idea, so I'd love some feedback from people who've worked on deployment systems. The idea is to compile an exported PyTorch model into a self-contained package that contains: graph binary weights backend kernels (currently WGSL) runtime metadata A lightweight runtime loads that package and executes it directly in the browser with WebGPU. No Python, no server inference, and no dependency on a heavyweight runtime. Rig