Husky: A model-specific inference engine up to 4.5x faster than Apple's MLX | Dark Hacker News