Predicting MoE experts speeds up GPU offload 5x

User achieved 150-200 tokens/s (up from 30) by predicting which MoE experts to load next on a 3060 12GB with Qwen3.6 35b A3B. The technique aims to reduce PCIe idle time by preloading likely experts.