River-LLM: Large Language Model Seamless Exit Based on KV Share

TL;DR

River-LLM introduces a training-free, efficient early exit framework for decoder-only LLMs, significantly reducing inference latency by sharing KV caches and guiding exit decisions without accuracy loss.

Contribution

It proposes a novel KV-Shared Exit River mechanism enabling seamless token-level early exit without additional training, improving practical speedup in LLM inference.

Findings

Achieves 1.71 to 2.16 times practical speedup.

Maintains high generation quality across tasks.

Eliminates costly KV cache recovery operations.

Abstract

Large Language Models (LLMs) have demonstrated exceptional performance across diverse domains but are increasingly constrained by high inference latency. Early Exit has emerged as a promising solution to accelerate inference by dynamically bypassing redundant layers. However, in decoder-only architectures, the efficiency of Early Exit is severely bottlenecked by the KV Cache Absence problem, where skipped layers fail to provide the necessary historical states for subsequent tokens. Existing solutions, such as recomputation or masking, either introduce significant latency overhead or incur severe precision loss, failing to bridge the gap between theoretical layer reduction and practical wall-clock speedup. In this paper, we propose River-LLM, a training-free framework that enables seamless token-level Early Exit. River-LLM introduces a lightweight KV-Shared Exit River that allows the backbone's missing KV cache to be naturally generated and preserved during the exit process, eliminating the need for costly recovery operations. Furthermore, we utilize state transition similarity within decoder blocks to predict cumulative KV errors and guide precise exit decisions. Extensive experiments on mathematical reasoning and code generation tasks demonstrate that River-LLM achieves 1.71 to 2.16 times of practical speedup while maintaining high generation quality.