AdaSkip: Adaptive Sublayer Skipping for Accelerating Long-Context LLM Inference

Long-context large language models (LLMs) inference is increasingly critical, motivating a number of studies devoted to alleviating the substantial storage and computational costs in such scenarios. Layer-wise skipping methods are promising optimizations but rarely explored in long-context inference. We observe that existing layer-wise skipping strategies have several limitations when applied in long-context inference, including the inability to adapt to model and context variability, disregard for sublayer significance, and inapplicability for the prefilling phase. This paper proposes AdaSkip, an adaptive sublayer skipping method specifically designed for long-context inference. AdaSkip adaptively identifies less important layers by leveraging on-the-fly similarity information, enables sublayer-wise skipping, and accelerates both the prefilling and decoding phases. The effectiveness of AdaSkip is demonstrated through extensive experiments on various long-context benchmarks and models, showcasing its superior inference performance over existing baselines.