I wrote a custom CUDA inference engine to run Qwen3.5-27B on $130 mining cards