arXiv subjects

Joo-Young Kim

Publications and source records attributed to Joo-Young Kim.

At least 19 recordsLinked to original sources

arXiv2026

PATTON: Enabling Commodity PIM for Production LLM Serving

Processing-in-Memory (PIM) is promising for accelerating memory-bound decode attention, but attention acceleration alone is insufficient for production LLM serving, where engines dynamically allocate, populate, share, cache, and reclaim logical KV cache blocks. Supporting this lifecycle on commodity PIM requires efficient physical memory allocation, block-to-address mapping, and command generation. For the Value cache, these requirements create a fundamental conflict among GEMV efficiency, single-token write efficiency, and memory capacity: GEMV-optimized layouts scatter newly generated Value vectors across rows, making writes costly, while finer-grained memory sharing improves capacity utilization but fragments GEMV reductions. We present PATTON, a PIM runtime that integrates production LLM serving engines with commodity PIM. PATTON introduces hierarchical granule allocation: block-sized Key and Value granules map one-to-one to logical token blocks, fixing their physical placements and commands, while coarser granules group blocks for efficient GEMV execution and memory utilization. A Commit Zone stages partial Value blocks for efficient single-token writes before committing them to GEMV-optimized locations. PATTON tracks these placements to generate KV cache writes and QK-transpose/SV commands. Across attention execution and runtime-induced prefill recomputation, PATTON achieves an average 1.95x speedup and 4.83x higher energy efficiency over evaluated baselines, requires no PIM processing-unit modifications, and maintains a KV cache hit rate comparable to the native GPU KV cache in vLLM.

cs.AR↗

arXiv2026

Distance Is Not Enough: Forget-Retain Alignment Gap Predicts LLM Relearning Robustness

Machine unlearning aims to make a model forget specific data, yet unlearned LLMs often fail to stay unlearned: brief fine-tuning can revive removed knowledge. Existing robustness predictors rely on global weight-space displacement, but distance alone can be misleading when random or destructive updates collapse performance. We argue that relearning robustness depends on update structure: robust unlearning should affect forget-critical weights while sparing retain-critical ones. We introduce the Forget-Retain Alignment Gap (FRAG), a training-free predictor that scores an update's forget-retain alignment without running a relearning attack, and separates selective from dense updates more reliably than global distance. Building on the forget-critical, retain-sparing principle, Forget-Retain Pruning (FRP) improves relearning robustness. Our results suggest that weight selectivity better explains robustness than distance alone. Code is available at https://github.com/Yi1-Chen/FRAG.

cs.AI↗

arXiv2026

APT: Accelerating Diffusion Transformers via Attention Probability-Guided Pruning and Quantization

Recent advances in generative AI have significantly increased the demand for high-resolution image and video generation, positioning diffusion models as a core technology. Among them, Diffusion Transformers (DiTs) have emerged as the state-of-the-art (SOTA) models due to their scalability and output quality. However, self-attention in DiTs incurs significant computational overhead, leading to excessively long latency as the complexity grows with the fourth power of the output resolution. While prior works have attempted to mitigate this cost using sparsity and quantization techniques, they fall short of effectively reducing the computational cost in high-resolution DiTs. In this paper, we present APT, a software-hardware co-designed accelerator for high-resolution DiTs. APT leverages attention probabilities as a unified importance metric to jointly optimize computation through fine-grained pruning and adaptive precision scaling. At the algorithm level, we propose Attention Probability-guided Adaptive Dual Thresholding (APDT), which dynamically performs element selection and precision assignment using dual thresholds. To ensure compatibility with memory-efficient FlashAttention, we introduce Timestep-Aware FlashAttention (TAFA), which predicts attention probabilities across timesteps by exploiting temporal similarity. At the architecture level, we co-design a specialized accelerator that efficiently supports irregular sparsity and dual-precision execution, featuring dynamic mask management, address translation, dual-precision compute units, and a tile-based dataflow. Finally, we evaluate APT on SOTA DiT models, including PixArt-$α$, Stable Diffusion 3, and FLUX. APT achieves up to 8.16$\times$ speedup and 14.98$\times$ higher energy efficiency over NVIDIA A100, and up to 3.01$\times$ speedup and 2.04$\times$ higher energy efficiency over EXION, a SOTA diffusion model accelerator.

cs.AR↗

arXiv2026

NOVA: Technology-Architecture Co-Design of Near-Memory Processing for Attention-SSM-MoE Hybrid LLM Inference

The rapid evolution of hybrid large language models (LLMs), which interleave grouped-query-attention (GQA), state-space model (SSM), and Mixture-of-Experts (MoE) layers, introduces two fundamental challenges for near-memory processing (NMP) architectures. First, the Technology Wall: the conventional 6F^2 DRAM cell is approaching its physical scaling limits at 10nm-class nodes, making it difficult to meet the memory capacity demands of MoE models with hundreds of experts. Second, the Architecture Wall: existing NMP designs target narrow arithmetic intensity (Op/B) ranges and cannot efficiently support the heterogeneous compute characteristics of hybrid LLMs, spanning memory-bound SSM layers, compute-intensive GQA layers, and large Op/B variations across experts. We propose NOVA, a technology-architecture co-designed NMP system that overcomes both walls. On the technology side, NOVA combines a 4F^2 vertical channel transistor (VCT) DRAM cell with a peri-over-cell (POC) structure to achieve approximately 2x memory density at iso-area over conventional 6F^2-based DRAM, enabling continued scaling into sub-10nm nodes. On the architecture side, NOVA repurposes the POC peripheral-die (peri-die) to host processing units, forming a 2-tier NMP architecture: Tier-1 (peri-die NMP) for low-to-mid Op/B operations, and Tier-2 (base-die NMP) for mid-to-high Op/B operations. Parallel execution across tiers supports diverse compute patterns for hybrid LLMs, maximizing inference performance. Evaluated on state-of-the-art hybrid and MoE LLMs including Nemotron3-Nano, Nemotron3-Super, Falcon-H1R, and Qwen3, NOVA achieves on average 4.5x higher throughput, 69.8% lower end-to-end latency, and 5x better energy efficiency over a GPU baseline, with only 3.9% area overhead and no loss in memory capacity.

cs.AR↗

arXiv2026

Beyond Capacity: Scalable MoE LLM Inference via High-Bandwidth Flash with Direct GPU and HBM Paths

Modern mixture-of-experts (MoE) language models increasingly strain the capacity and cost efficiency of high-bandwidth memory (HBM), as rapidly growing expert weights must be provisioned close to GPUs. High-bandwidth flash (HBF) offers substantially greater capacity, but conventional designs typically deliver HBF-resident expert weights to the GPU through HBM, leaving an additional direct GPU-HBF connection underutilized. We explore an HBF organization that simultaneously exploits two independent expert-delivery routes: a direct path that transfers expert weights from HBF to the GPU and a relay path that transfers them from HBF through the HBM base die to the GPU. Whole experts are assigned to one of the two routes, and transfers over both routes proceed concurrently, increasing aggregate expert-delivery bandwidth without replicating expert weights or introducing a shared relay bottleneck. Early expert determination identifies upcoming experts ahead of their conventional execution point, allowing HBF read latency to overlap with preceding computation, while separate management of immutable expert weights and mutable KV-cache data reduces interference between the two traffic classes. We evaluate the architecture using an event-driven continuous-batching LLM serving simulator with empirically measured GPU compute latencies. Across representative MoE workloads, concurrently utilizing the direct GPU-HBF and HBF-HBM-GPU routes consistently improves expert-delivery efficiency over designs restricted to either route alone. For a representative workload, the proposed architecture can achieve 1.94$\times$ higher throughput and 1.90$\times$ end-to-end speedup over a design that delivers all HBF-resident expert weights to the GPU through the HBM base die.

cs.AR↗

arXiv2026

3DLS: A 3D Logic-Stacked Architecture for Disaggregated LLM Serving

Large language model (LLM) serving increasingly combines prefill-decode (PD) disaggregation with tensor parallelism (TP) to support large models and long contexts. In conventional 2D/2.5D chiplet architectures, layer-wise prefill-to-decode KV-cache transfer decode-side TP collectives share the same lateral die-to-die (D2D) interconnect, creating mixed-traffic contention on the decode critical path. This contention increases communication latency, prolongs token generation intervals, and degrades end-to-end serving performance. We propose 3DLS, a logic-on-logic 3D-stacked chiplet architecture that separates traffic classes by routing KV-cache transfers through vertical interconnects while preserving decode-side TP collectives on the lateral D2D fabric. 3DLS achieves up to 1.49$\times$ throughput and 60.2\% lower end-to-end (E2E) latency over the shared-fabric planar baseline, and still achieves up to 1.17$\times$ throughput and 31.4\% lower E2E latency over a workload-aware priority-managed planar baseline. These results highlight that physical isolation is an important design principle for future chiplet-based PD-disaggregated LLM serving systems.

cs.AR↗

arXiv2026

ZK-Flex: A Flexible and Scalable Framework for Accelerating Zero-Knowledge Proofs

Zero-knowledge proofs (ZKP) allows a prover to convince a verifier of computational correctness without revealing private data, ensuring both privacy and verifiability. However, proof generation is highly compute-intensive, dominated by polynomial (POLY) and elliptic-curve (EC) operations. These workloads pose two key challenges for hardware acceleration: (1) efficiently supporting diverse large-precision modular multiplications, and (2) maintaining high utilization across workloads that dynamically shift between POLY and EC stages. Existing reconfigurable accelerators address these issues only partially, remaining limited in precision scalability, algorithmic flexibility, and resource efficiency. To overcome these limitations, we propose ZK-Flex, a flexible and scalable software-hardware co-designed framework for accelerating ZKP proof generation. The software layer incorporates POLY and EC optimizers that reduce computation through hardware- and workload-aware algorithmic choices, while the hardware integrates TCore, a Toom-Cook-based multi-precision core with a flexible NoC and a linked-list memory mechanism that improves parallelism under limited memory capacity. Across representative ZKP benchmarks, ZK-Flex achieves 5 to 11 times speedup and up to 3.8 times higher area efficiency over the state of the art, establishing a new foundation for high-performance, reconfigurable ZKP acceleration.

cs.AR↗

arXiv2026

Cassandra: Enabling Reasoning LLMs at Edge via Self-Speculative Decoding

Speculative decoding has emerged as a promising lossless approach for accelerating Large Language Models (LLMs). As reasoning LLMs increasingly suffer from decode-stage overhead and approximation-based methods degrade accuracy, lossless speculative decoding has become essential for efficient inference. However, existing methods still struggle to deliver strong low-batch performance without additional training, limiting practical deployment on consumer devices. To address this challenge, we propose Cassandra, an algorithm-hardware co-designed self-speculative decoding framework optimized for low-batch scenarios. Cassandra constructs a high-performance, training-free draft model through fine-grained data selection. Using optimized pruning and mantissa truncation, it identifies the most salient values in both model weights and the Key-Value (KV) cache, enabling rapid candidate token generation before full-precision parallel verification. Unlike prior self-speculative decoding methods based on layer skipping or structured KV compression, Cassandra achieves significantly higher efficiency. To further reduce the overhead of format conversion between Cassandra representations and standard floating-point formats, we also introduce a lightweight encoder-decoder hardware module designed for seamless integration with commercial GPUs and NPUs. Experimental results show that Cassandra achieves up to 2.41x speedup over the BF16 baseline without additional training. Furthermore, on Llama 3 8B running on an NVIDIA GeForce RTX 4090, Cassandra generates 1.81x more tokens under the same memory budget compared to Eagle-3, a state-of-the-art speculative decoding method.

cs.AR↗

arXiv2026

DiSC: Resolution-Scalable Acceleration of Diffusion Models by Exploiting Sparsity and Cached Token Reuse with Hash-based Distribution

Transformer-based diffusion models offer superior scalability and performance but suffer from high computational overhead due to the iterative nature and quadratic complexity of self-attention at high resolutions. In this paper, we propose DiSC, a resolution-scalable, sparsity-aware hardware accelerator. At the software level, DiSC introduces two algorithms: Cached Token Reuse (CTR), and Softmax Thresholding with Sparsity Mask Reuse (ST). CTR introduces a mechanism that translates spatial variations in the input latent difference across steps into a token-level reuse decision, effectively eliminating redundant token computation. ST induces sparsity in attention operations by reusing a generated sparsity pattern, leveraging temporal similarity to bypass costly prediction overhead. Together, these algorithms provide resolution-scalable computational benefits and yield a moderate sparsity and hybrid dense-sparse workload. To exploit this efficiently, we design a specialized hardware architecture and unified dataflow. This architecture avoids dedicated sparsity-handling components; instead, a hash-based distribution over on-chip memory banks allows DiSC to reuse its existing compute engines for sparse operations, efficiently exploiting the induced sparsity with minimal hardware overhead. Evaluated on DiT and PixArt-Sigma, DiSC achieves 3.47-4.74x and 2.48-3.50x speedups over NVIDIA A100 and H100 GPUs, respectively, with energy savings ranging from 46.4% to 68.1%.

cs.AR↗

arXiv2026

MASQ: Accelerating Masked Diffusion via Stage-Wise Multi-Precision Quantization

Masked diffusion enables region-specific image synthesis but suffers from computational redundancy, since the entire image is processed each timestep even though only the masked region requires generation. To address this, we introduce MASQ, a hardware-software co-designed accelerator for masked diffusion. Our approach performs stage-wise MXINT8/4/2 precision assignment that dynamically reflects spatial and semantic importance, complemented by timestep-aware scheduling and optimized non-matrix operations. MASQ features a block-wise multi-precision compute engine and mask management unit, efficiently handling our approach. It achieves up to 16.06x and 5.39x speedup and 4.18x and 4.93x energy-efficiency gain over A100 and Orin NX, respectively, while preserving quality.

cs.AR↗

arXiv2026

Rethinking Token Reduction for Diffusion Models via Output-Similarity-Awareness

Diffusion Transformers (DiTs) achieve superior image generation quality but suffer from quadratic computational complexity relative to token count. While various token reduction (TR) methods have been proposed to mitigate this cost, they overlook the primary objective of generative models: minimizing recovery error, which requires reflecting output token similarity. They rely solely on input token similarity inherited from reduction-only ViT paradigms, leading to a fundamental misalignment with this objective. To bridge this gap, we propose DiTo, a novel TR paradigm that shifts the focus toward output-centric token reduction. Based on the observation that output token similarity is consistently preserved across adjacent timesteps, DiTo utilizes prior-step similarities as an effective proxy to establish token correspondences at a Matching timestep, which are then reused across multiple subsequent Reduction timesteps. To optimize this interleaved scheduling, we propose Pair Match Ratio (PMR)-guided Interval Scheduling to determine the optimal matching frequency. Furthermore, to mitigate localized approximation errors and resulting blocking artifacts caused by repeated reuse, we propose Frequency-aware Token Matching by incorporating a selection-frequency penalty. Extensive experiments demonstrate that DiTo consistently outperforms existing TR methods with 1.6-3.9 dB higher PSNR at comparable speedups, achieving a superior Pareto frontier.

cs.CV↗

arXiv2026

ORBIS: Output-Guided Token Reduction with Distribution-Aware Matching for Video Diffusion Acceleration

Diffusion Transformer (DiT) has emerged as a powerful model architecture for generating high-quality images and videos. In the case of video DiT, 3D Spatio-Temporal Attention increases token length in proportion to the number of frames, sharply increasing computational cost. Token reduction methods mitigate this cost by exploiting spatial redundancy, but existing approaches rely on inaccurate similarity estimates and lightweight matching algorithms, resulting in poor matching quality and only marginal acceleration. To overcome these limitations, we propose ORBIS, an SW-HW co-designed accelerator for video DiT. ORBIS leverages the output activation from the previous timestep to obtain more accurate inter-token similarity, substantially improving matching quality and enabling a higher token reduction ratio. We further introduce a Distribution-Aware Token Matching (DATM) algorithm that captures global token distribution and explicitly minimizes token-pair loss for additional gains. To fully hide DATM latency, we design specialized, deeply pipelined hardware and minimize its hardware cost through quantization, occupying only 2.4% of total area with negligible accuracy loss. Extensive experiments show that ORBIS achieves about 2x higher token reduction ratio than the state-of-the-art approach, AsymRnR, while delivering up to 4.5x speedup and 79.3% energy reduction compared to an NVIDIA A100 GPU.

cs.CV↗

arXiv2026

CoX-MoE: Coalesced Expert Execution for High-Throughput MoE Inference with AMX-Enabled CPU-GPU Co-Execution

The Mixture-of-Experts (MoE) architecture improves computational efficiency via sparse expert activation, but throughput-oriented inference faces substantial GPU memory pressure due to a significant parameter size and intermediate data. Prior works attempt to mitigate this using expert offloading with micro-batching or by offloading computation to the CPU. However, the fragmented workload resulting from micro-batching degrades operational intensity, causing expert execution to become memory-bound. Meanwhile, CPU offloading is constrained by slow PCIe transfers and its limited applicability to attention computation in the decode stage. Consequently, these inefficiencies prevent effective system utilization, severely restricting the end-to-end throughput of MoE inference. To address these challenges, this paper proposes CoX-MoE, an Advanced Matrix Extensions (AMX)-enabled CPU-GPU collaborative system that comprehensively optimizes MoE inference by combining coalesced expert execution with strategic workload orchestration for higher throughput. CoX-MoE introduces (i) a coalescing-aware orchestration policy to jointly optimize resource allocation by adopting ordinary batch, instead of micro-batch, for expert computation and selective attention offloading, and (ii) a static expert-aware stratification scheme that pre-assigns frequently activated experts to the GPU, mitigating PCIe transfer overhead and balancing workload for the CPU and GPU during inference. Compared to state-of-the-art frameworks, CoX-MoE delivers significant gains, achieving up to 7.1x and 2.4x higher throughput than FlexGen and MoE-Lightning, respectively.

cs.LG↗

arXiv2026

Reformulating KV Cache Eviction Problem for Long-Context LLM Inference

Large language models (LLMs) support long-context inference but suffer from substantial memory and runtime overhead due to Key-Value (KV) Cache growth. Existing KV Cache eviction methods primarily rely on local attention weights, neglecting the influence of value representations, output projection, and inter-head interactions. In this work, we reformulate KV Cache eviction from a conventional head-wise, weight-averaging approach into an output-aware, layer-wise matrix multiplication approximation problem. We introduce LaProx, a novel eviction strategy that explicitly models the multiplicative interaction between attention maps and projected value states to accurately quantify token contributions while accounting for inter-head dependencies. Building on this metric, we propose the first unified eviction strategy that assigns globally comparable importance scores to tokens, enabling model-wide selection instead of local, head-wise decisions. Experimental results across 19 datasets on long-context benchmarks LongBench and Needle-In-A-Haystack demonstrate that our approach maintains model performance with only 5\% of the KV cache and consistently outperforms prior works across all configurations. Notably, our method achieves up to 2$\times$ accuracy loss reduction under extreme compression scenarios compared to existing state-of-the-art baselines with minimal overhead.

cs.CL↗

arXiv2026

Training-free Adjustable Polynomial Graph Filtering for Ultra-fast Multimodal Recommendation

Multimodal recommender systems improve the performance of canonical recommender systems with no item features by utilizing diverse content types such as text, images, and videos, while alleviating inherent sparsity of user-item interactions and accelerating user engagement. However, current neural network-based models often incur significant computational overhead due to the complex training process required to learn and integrate information from multiple modalities. To address this challenge, we propose a training-free multimodal recommendation method grounded in graph filtering, designed for multimodal recommendation systems to achieve efficient and accurate recommendation. Specifically, the proposed method first constructs multiple similarity graphs for two distinct modalities as well as user-item interaction data. Then, it optimally fuses these multimodal signals using a polynomial graph filter that allows for precise control of the frequency response by adjusting frequency bounds. Furthermore, the filter coefficients are treated as hyperparameters, enabling flexible and data-driven adaptation. Extensive experiments on real-world benchmark datasets demonstrate that the proposed method not only improves recommendation accuracy by up to 22.25% compared to the best competitor but also dramatically reduces computational costs by achieving the runtime of less than 10 seconds.

cs.IR↗

arXiv2026

FastSTAR: Spatiotemporal Token Pruning for Efficient Autoregressive Video Synthesis

Visual Autoregressive modeling (VAR) has emerged as a highly efficient alternative to diffusion-based frameworks, achieving comparable synthesis quality. However, as this paradigm extends to Spacetime Autoregressive modeling (STAR) for video generation, scaling resolution and frame counts leads to a "token explosion" that creates a massive computational bottleneck in the final refinement stages. To address this, we propose FastSTAR, a training-free acceleration framework designed for high-quality video generation. Our core method, Spatiotemporal Token Pruning, identifies essential tokens by integrating two specialized terms: (1) Spatial similarity, which evaluates structural convergence across hierarchical scales to skip computations in regions where further refinement becomes redundant, and (2) Temporal similarity, which identifies active motion trajectories by assessing feature-level variations relative to the preceding clip. Combined with a Partial Update mechanism, FastSTAR ensures that only non-converged regions are refined, maintaining fluid motion while bypassing redundant computations. Experimental results on InfinityStar demonstrate that FastSTAR achieves up to a 2.01x speedup with a PSNR of 28.29 and less than 1% performance degradation, proving a superior efficiency-quality trade-off for STAR-based video synthesis.

cs.CV↗

arXiv2026

POP: Online Structural Pruning Enables Efficient Inference of Large Foundation Models

Large foundation models (LFMs) achieve strong performance through scaling, yet current structural pruning methods derive fixed pruning decisions during inference, overlooking sparsity patterns that emerge in the autoregressive token generation. In this paper, we propose POP (Partition-guided Online Pruning), an efficient online structural pruning framework that enables context-conditioned dynamic pruning with minimal computational overhead. POP partitions model channels into retained, candidate, and pruned regions, where prefilling defines a coarse pruning partition, and the decoding stage generates a fine-grained mask within the candidate region, avoiding full-channel re-evaluation. The coarse pruning partition preserves consistently important weights, while the fine-grained masking provides context-conditioned variation during decoding. Moreover, POP is a lightweight, plug-and-play method that requires no preprocessing, including offline calibration, retraining, or learning predictors. Extensive evaluations across diverse LFMs, including large language models (LLMs), mixture-of-experts models (MoEs), and vision-language models (VLMs), demonstrate that POP consistently delivers higher accuracy than existing pruning approaches while incurring smaller computational overhead and minimizing inference latency.

cs.AI↗

arXiv2025

V-Rex: Real-Time Streaming Video LLM Acceleration via Dynamic KV Cache Retrieval

Streaming video large language models (LLMs) are increasingly used for real-time multimodal tasks such as video captioning, question answering, conversational agents, and augmented reality. However, these models face fundamental memory and computational challenges because their key-value (KV) caches grow substantially with continuous streaming video input. This process requires an iterative prefill stage, which is a unique feature of streaming video LLMs. Due to its iterative prefill stage, it suffers from significant limitations, including extensive computation, substantial data transfer, and degradation in accuracy. Crucially, this issue is exacerbated for edge deployment, which is the primary target for these models. In this work, we propose V-Rex, the first software-hardware co-designed accelerator that comprehensively addresses both algorithmic and hardware bottlenecks in streaming video LLM inference. At its core, V-Rex introduces ReSV, a training-free dynamic KV cache retrieval algorithm. ReSV exploits temporal and spatial similarity-based token clustering to reduce excessive KV cache memory across video frames. To fully realize these algorithmic benefits, V-Rex offers a compact, low-latency hardware accelerator with a dynamic KV cache retrieval engine (DRE), featuring bit-level and early-exit based computing units. V-Rex achieves unprecedented real-time of 3.9-8.3 FPS and energy-efficient streaming video LLM inference on edge deployment with negligible accuracy loss. While DRE only accounts for 2.2% power and 2.0% area, the system delivers 1.9-19.7x speedup and 3.1-18.5x energy efficiency improvements over AGX Orin GPU. This work is the first to comprehensively tackle KV cache retrieval across algorithms and hardware, enabling real-time streaming video LLM inference on resource-constrained edge devices.

eess.IV↗