VLCache: Computing 2% Vision Tokens and Reusing 98% for Vision-Language Inference

This paper presents VLCache, a cache reuse framework that exploits both Key-Value (KV) cache and encoder cache from prior multimodal inputs to eliminate costly recomputation when the same multimodal inputs recur. Unlike previous heuristic approaches, we formally identify the cumulative reuse error effect and demonstrate how to minimize the non-prefix cache reuse error effectively. We further analyze the varying importance of model layers and propose a dynamic, layer-aware recomputation strategy to balance accuracy and efficiency. Experimental results show that VLCache achieves an accuracy on par with full recomputation, while requiring only 2-5% of the tokens to compute, yielding 1.2x-16x TTFT speedups. We develop an experimental implementation of the proposed VLCache pipeline based on SGLang, enabling significantly faster inference in practical deployments.

Optimal Expert Selection for Distributed Mixture-of-Experts at the Wireless Edge

The emergence of distributed Mixture-of-Experts (DMoE) systems, which deploy expert models at edge nodes, offers a pathway to achieving connected intelligence in sixth-generation (6G) mobile networks and edge artificial intelligence (AI). However, current DMoE systems lack an effective expert selection algorithm to address the simultaneous task-expert relevance and channel diversity inherent in these systems. Traditional AI or communication systems focus on either performance or channel conditions, and direct application of these methods leads to high communication overhead or low performance. To address this, we propose the DMoE protocol to schedule the expert inference and inter-expert transmission. This protocol identifies expert selection and subcarrier allocation as key optimization problems. We formulate an expert selection problem by incorporating both AI performance and channel conditions, and further extend it to a Joint Expert and Subcarrier Allocation (JESA) problem for comprehensive AI and channel management within the DMoE framework. For the NP-hard expert selection problem, we introduce the Dynamic Expert Selection (DES) algorithm, which leverages a linear relaxation as a bounding criterion to significantly reduce search complexity. For the JESA problem, we discover a unique structural property that ensures asymptotic optimality in most scenarios. We propose an iterative algorithm that addresses subcarrier allocation as a subproblem and integrates it with the DES algorithm. The proposed framework effectively manages the tradeoff between task relevance and channel conditions through a tunable importance factor, enabling flexible adaptation to diverse scenarios. Numerical experiments validate the dual benefits of the proposed expert selection algorithm: high performance and significantly reduced cost.