Weikuan Yu

Oteo Mamo, Olga Kogiou, Hyunjin Yi et al.

Efficient inference with Large Language Models (LLMs) increasingly relies on Key-Value (KV) caches to store previously computed key and value vectors at each layer. These caches are essential to minimize redundant computation during autoregressive token generation, lowering computational complexity from quadratic to linear. However, the growth of KV caches has posed significant system-level challenges, particularly as model sizes increase, context lengths grow, and concurrent requests compete for limited memory resources. Even though several recent frameworks for KV cache management have emerged, their comparative trade-offs in memory consumption and inference performance have not been fully understood, especially under varying request sizes and model configurations. In this work, we conduct an empirical study of three state-of-the-art KV cache management frameworks: vLLM, InfiniGen, and H2O. These frameworks employ techniques such as tensor offloading, token eviction heuristics, and speculative scheduling to balance memory usage and performance. We evaluate their performance in terms of a range of metrics such as latency, throughput, and memory usage across a spectrum of key parameters including request rates, model sizes, and sparsity levels. Our results pinpoint the conditions for each framework to perform the best, revealing the most suitable selection and configuration of KV cache strategies under memory and performance constraints.

Bodon Jeong, Hongsu Byun, Youngjae Kim et al.

The increasing deployment of Large Language Model (LLM) inference on edge AI systems demands efficient execution under tight memory budgets. A key challenge arises from Key-Value (KV) caches, which often exceed available device memory. Although NVMe-based offloading offers scalable capacity, existing file-based designs rely heavily on the kernel page cache, leading to cache thrashing, unpredictable latency, and high software overhead under memory pressure. We present DUAL-BLADE, a dual-path KV residency framework that dynamically assigns KV tensors to either a page-cache path or an NVMe-direct path based on runtime memory availability. The NVMe-direct path bypasses the filesystem by mapping KV tensors to contiguous logical block address (LBA) regions, enabling low-overhead direct storage access. DUAL-BLADE further incorporates adaptive pipeline parallelism to overlap storage I/O with GPU DMA, improving inference throughput. Our evaluation shows that DUAL-BLADE substantially mitigates I/O bottlenecks, reducing prefill and decode latency by up to 33.1% and 42.4%, respectively, while improving SSD utilization by 2.2x across diverse memory budgets.

Chashi Mahiul Islam, Oteo Mamo, Samuel Jacob Chacko et al.

Vision-language models (VLMs) have advanced multimodal reasoning but still face challenges in spatial reasoning for 3D scenes and complex object configurations. To address this, we introduce SpatialViLT, an enhanced VLM that integrates spatial features like depth maps, 3D coordinates, and edge maps through a multi-task learning framework. This approach enriches multimodal embeddings with spatial understanding. We propose two variants: SpatialViLT and MaskedSpatialViLT, focusing on full and masked object regions, respectively. Additionally, SpatialEnsemble combines both approaches, achieving state-of-the-art accuracy. Our models excel in spatial reasoning categories such as directional, topological, and proximity relations, as demonstrated on the challenging Visual Spatial Reasoning (VSR) dataset. This work represents a significant step in enhancing the spatial intelligence of AI systems, crucial for advanced multimodal understanding and real-world applications.

Subhadeep Bhattacharya, Weikuan Yu, Fahim Tahmid Chowdhury

Large neural network models present a hefty communication challenge to distributed Stochastic Gradient Descent (SGD), with a communication complexity of O(n) per worker for a model of n parameters. Many sparsification and quantization techniques have been proposed to compress the gradients, some reducing the communication complexity to O(k), where k << n. In this paper, we introduce a strategy called two-level gradient averaging (A2SGD) to consolidate all gradients down to merely two local averages per worker before the computation of two global averages for an updated model. A2SGD also retains local errors to maintain the variance for fast convergence. Our theoretical analysis shows that A2SGD converges similarly like the default distributed SGD algorithm. Our evaluation validates the theoretical conclusion and demonstrates that A2SGD significantly reduces the communication traffic per worker, and improves the overall training time of LSTM-PTB by 3.2x and 23.2x, respectively, compared to Top-K and QSGD. To the best of our knowledge, A2SGD is the first to achieve O(1) communication complexity per worker for distributed SGD.