Hybrid LLM runtime — minimal VRAM, always-on GPU prefill, optimised CPU inference - View it on GitHub