LayerSkip is an innovative technology designed to enhance the efficiency of large language models (LLMs) by enabling early exit inference and self-speculative decoding. These methods allow AI systems to produce faster results by optimizing the inference process. To fully experience the speedup benefits of LayerSkip, users need access to models specifically trained with this method. Facebook Research has open-sourced several Llama models pre-trained on HuggingFace using the LayerSkip technique. Users must request access to these models, follow the necessary steps to obtain a user token, and execute commands to utilize the checkpoints effectively. By reducing the number of necessary computations, LayerSkip's self-speculative decoding facilitates faster token generation during model inference. This approach supports various language modeling tasks, such as predicting missing parts of text with sampled or greedy decoding methods. Although classification tasks do not benefit from speed increases using this method, generative tasks can witness significant improvements. LayerSkip integrates with the Eleuther Language Model Evaluation Harness, allowing a wide range of tasks to be tested with different models and datasets, optimizing the hyperparameters for speedup accordingly. Beyond its core model, LayerSkip's implementation extends to existing platforms such as gpt-fast and HuggingFace models, offering flexible configurations and optimizations like quantization and tensor parallelism. Although convergence between autoregressive and self-speculative decoding outputs can only be verified without sampling, the technology lays potential for further optimization of AI applications. LayerSkip, licensed under a CC-by-NC license, is still evolving, and contributions are welcomed to further its development while adopting it in AI research contexts.