LayerSkip: Enabling Early Exit Inference and Self-Speculative Decoding

TL;DR

LayerSkip introduces a training and inference method for large language models that enables early exit during inference, significantly speeding up tasks like summarization and coding without extra model complexity.

Contribution

The paper proposes a novel training recipe and a self-speculative decoding approach that allow early exit inference in LLMs, improving speed without additional model modules.

Findings

Achieved up to 2.16x speedup on summarization tasks

Improved early exit accuracy during training

Reduced memory footprint compared to other speculative decoding methods

Abstract

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 earlier layers and higher dropout rates for later layers, and an early exit loss where all transformer layers share the same exit. Second, during inference, we show that this training recipe increases the accuracy of early exit at earlier layers, without adding any auxiliary layers or modules to the model. Third, we present a novel self-speculative decoding solution where we exit at early layers and verify and correct with remaining layers of the model. Our proposed self-speculative decoding approach has less memory footprint than other speculative decoding approaches and benefits from shared compute and activations of the draft and verification stages. We run experiments on different Llama model sizes on different types of training: pretraining from scratch, continual pretraining, finetuning on specific data domain, and finetuning on specific task. We implement our inference solution and show speedups of up to 2.16x on summarization for CNN/DM documents, 1.82x on coding, and 2.0x on TOPv2 semantic parsing task. We open source our code and checkpoints at https://github.com/facebookresearch/LayerSkip.