Meta presents Layer Skip Enabling Early Exit Inference and Self-Speculative Decoding We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs). First, during training we apply layer dropout, with low dropout rates for