Sangseok Yun

Mingyu Sung, Il-Min Kim, Sangseok Yun et al.

Diffusion models have emerged as state-of-the-art in image generation, but their practical deployment is hindered by the significant computational cost of their iterative denoising process. While existing caching techniques can accelerate inference, they often create a challenging trade-off between speed and fidelity, suffering from quality degradation and high computational overhead. To address these limitations, we introduce H2-Cache, a novel hierarchical caching mechanism designed for modern generative diffusion model architectures. Our method is founded on the key insight that the denoising process can be functionally separated into a structure-defining stage and a detail-refining stage. H2-cache leverages this by employing a dual-threshold system, using independent thresholds to selectively cache each stage. To ensure the efficiency of our dual-check approach, we introduce pooled feature summarization (PFS), a lightweight technique for robust and fast similarity estimation. Extensive experiments on the Flux architecture demonstrate that H2-cache achieves significant acceleration (up to 5.08x) while maintaining image quality nearly identical to the baseline, quantitatively and qualitatively outperforming existing caching methods. Our work presents a robust and practical solution that effectively resolves the speed-quality dilemma, significantly lowering the barrier for the real-world application of high-fidelity diffusion models. Source code is available at https://github.com/Bluear7878/H2-cache-A-Hierarchical-Dual-Stage-Cache.

Mingyu Sung, Vikas Palakonda, Suhwan Im et al.

Large language models (LLMs) have achieved near-human performance across diverse reasoning tasks, yet their deployment on resource-constrained Internet-of-Things (IoT) devices remains impractical due to massive parameter footprints and memory-intensive autoregressive decoding. While split computing offers a promising solution by partitioning model execution between edge devices and cloud servers, existing approaches fail to address the unique challenges of autoregressive inference, particularly the iterative token generation process and expanding key-value (KV) cache requirements. This work introduces the first autoregressive-aware split computing framework designed explicitly for LLM deployment on edge devices. Our approach makes three key contributions. First, we develop one-point split compression (OPSC), a mixed-precision quantization scheme that prevents out-of-memory failures by strategically partitioning models into front-end and back-end segments with different precision levels. Second, we propose a two-stage intermediate compression pipeline that combines threshold splitting (TS) and token-wise adaptive bit quantization (TAB-Q) to preserve accuracy-critical activations while dramatically reducing communication overhead. Third, we formulate a unified optimization framework that jointly selects optimal split points, quantization settings, and sequence lengths to satisfy strict memory and latency constraints. Extensive evaluations across diverse LLMs and hardware platforms demonstrate superior performance compared to state-of-the-art quantization methods, including SmoothQuant, OmniQuant, and Atom. The framework achieves a 1.49 inference speedup and significant communication overhead reduction while maintaining or improving model accuracy.

Mingyu Sung, Suhwan Im, Daeho Bang et al.

Modern DNNs often rely on edge-cloud model partitioning (MP), but widely used schemes fix shallow, static split points that underutilize edge compute and concentrate latency and energy on the server. The problem is exacerbated in autoregressive (AR) LLM inference, where per-token forward passes repeatedly generate bulky intermediate features (IFs). We introduce SLICER, a retraining-free, architecture-agnostic framework that compresses IFs to reduce both communication and server load in split computing. SLICER combines (i) asymmetric top-K filtering (ATKF) to sparsify low-magnitude activations, (ii) magnitude-splitting (MS) to group the remaining non-zeros into equal-cardinality blocks, and (iii) adaptive bit quantization (ABQ) that selects per-block bitwidths under a distortion budget. Across standard vision and LLM workloads (e.g., ImageNet/COCO; HellaSwag, PIQA, ARC-E/C, GSM8K, HumanEval), SLICER reduces uplink volume by up to 10x and server GPU time by up to 4.4x, while keeping task quality within ~0-3 pp of baseline. In multi-device settings and AR LLMs, SLICER scales by shifting meaningful compute to the edge and lowering bits-per-token and server time per token, stabilizing per-step traffic. The codec attaches to off-the-shelf models without retraining or architectural changes, offering a plug-and-play path to scalable, low-latency distributed inference. Code is provided in the supplementary material.

Milad Khademi Nori, Sangseok Yun, Il-Min Kim

Federated Learning (FL) has recently received a lot of attention for large-scale privacy-preserving machine learning. However, high communication overheads due to frequent gradient transmissions decelerate FL. To mitigate the communication overheads, two main techniques have been studied: (i) local update of weights characterizing the trade-off between communication and computation and (ii) gradient compression characterizing the trade-off between communication and precision. To the best of our knowledge, studying and balancing those two trade-offs jointly and dynamically while considering their impacts on convergence has remained unresolved even though it promises significantly faster FL. In this paper, we first formulate our problem to minimize learning error with respect to two variables: local update coefficients and sparsity budgets of gradient compression who characterize trade-offs between communication and computation/precision, respectively. We then derive an upper bound of the learning error in a given wall-clock time considering the interdependency between the two variables. Based on this theoretical analysis, we propose an enhanced FL scheme, namely Fast FL (FFL), that jointly and dynamically adjusts the two variables to minimize the learning error. We demonstrate that FFL consistently achieves higher accuracies faster than similar schemes existing in the literature.