CUDA fork of antirez ds4 that runs giant MoE language models on a single consumer GPU by streaming routed experts from SSD.