Official PyTorch implementation for Hogwild! Inference: Parallel LLM Generation with a Concurrent Attention Cache