arXiv subjects

Sungjoo Yoo

Publications and source records attributed to Sungjoo Yoo.

At least 19 recordsLinked to original sources

arXiv2026

GraniKV: Asymmetric Granularity KV-Cache Paging for Multi-Agent Systems with Long Shared Prefix

Production paged-serving engines apply uniform paging granularity to the KV cache, even though the two regions of a multi-agent workload have opposite storage requirements: a long shared prefix demands contiguity, while the per-request suffix demands fine-grained allocation. We present \textbf{GraniKV}, a KV-cache layer that allocates the shared prefix in a contiguous HOT pool and the suffix in a token-level COLD pool, combined with a per-step dispatcher which selects the appropriate backend among dual backends for each regime (compute-, memory-, or communication-bound). To the best of our knowledge, GraniKV is the first system to apply asymmetric paging granularity to the KV cache of a production paged-serving engine. At $L_p{=}16$\,K shared tokens GraniKV reaches $\mathbf{2.16\times}$, $\mathbf{1.98\times}$, and $\mathbf{1.57\times}$ output-token throughput over the production baseline on Llama-3.1-8B/TP=1, Qwen-2.5-14B/TP=2, and Qwen-2.5-32B/TP=4. The gain decomposes: cascade attention integration contributes the majority at saturation; the asymmetric storage layer adds $1.05$--$1.15\times$ end-to-end while being what makes the batched-GEMM prefix backend possible at all. Under heterogeneous multi-agent serving with \emph{distinct} prompts of different lengths, the attribution inverts: GraniKV sustains $\mathbf{1.95\times}$ while batch-global cascade collapses to parity --- the storage layer alone carries the win in the regime that motivates the paper.

cs.LG↗

arXiv2026

KV-Rescue: Recovering Reasoning Language Model KV Eviction Loss via Stepwise Interleaving

KV-cache eviction caps the memory cost of long reasoning traces but is inherently lossy because the model decodes from a partial view of its history. Under aggressive budgets, this not only lowers accuracy but can also cause runaway degeneration, where the model produces incoherent or repetitive tokens until reaching the length limit. We characterize much of this loss as an information gapf caused by missing context, rather than a capability gap caused by limited model capacity. An evicted 7B model and a full-context 1.5B model make complementary errors, and an oracle choice between their answers recovers 79% of the accuracy gap to the full-KV 7B model. Based on this observation, we propose KV-Rescue, a training-free inference framework that bridges the information gap introduced by KV eviction using a lightweight full-context helper. KV-Rescue interleaves reasoning steps from the two models into a shared trajectory. An online detector uses entropy and compressibility to terminate the generation of incoherent or repetitive base-model candidates early. Across five math benchmarks with Qwen2.5-Math 7B and 72B, KV-Rescue recovers an average of 87% of the accuracy lost to eviction at eviction budget B=64. A decode-cost analysis further shows that preventing runaway degeneration cuts base-model token generation by 43% on average.

cs.AI↗

arXiv2026

Output-Aware Rotation for INT2 KV-Cache Quantization

The key-value (KV) cache has become a major memory and bandwidth bottleneck in long-context large language model inference, making ultra-low-bit quantization increasingly important. However, existing rotation-based INT2 methods optimize cache statistics or proxy errors before the complete attention readout, even though the model is ultimately affected by the error propagated through attention and the output projection $W_O$. To address this mismatch, we propose \textit{OptR}, an output-aware rotation method that minimizes post-$W_O$ attention-output error. OptR decomposes the post-$W_O$ attention-output error into key- and value-induced terms and learns per-head orthogonal corrections through the full INT2 quantization and attention path. OptR further applies an attention-equivalent key reparameterization to reduce large channel-wise offsets without changing the softmax distribution. Across three models and five reasoning and coding benchmarks, OptR consistently improves both QuaRot and OSCAR and strengthens long-context retrieval, while preserving the paged KV-cache format with negligible inference overhead.

cs.LG↗

arXiv2026

NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache

Large Language Model (LLM) inference is typically memory-intensive, especially when processing large batch sizes and long sequences, due to the large size of key-value (KV) cache. Vector Quantization (VQ) is recently adopted to alleviate this issue, but we find that the existing approach is susceptible to distribution shift due to its reliance on calibration datasets. To address this limitation, we introduce NSNQuant, a calibration-free Vector Quantization (VQ) technique designed for low-bit compression of the KV cache. By applying a three-step transformation-1) a token-wise normalization (Normalize), 2) a channel-wise centering (Shift), and 3) a second token-wise normalization (Normalize)-with Hadamard transform, NSNQuant effectively aligns the token distribution with the standard normal distribution. This alignment enables robust, calibration-free vector quantization using a single reusable codebook. Extensive experiments show that NSNQuant consistently outperforms prior methods in both 1-bit and 2-bit settings, offering strong generalization and up to 3$\times$ throughput gain over full-precision baselines. Code is available at https://github.com/DHdroid/NSNQuant.

cs.LG↗

arXiv2026

LAQuant: A Simple Overhead-free Large Reasoning Model Quantization by Layer-wise Lookahead Loss

Large reasoning models (LRMs) reach competition-level math and coding accuracy via long autoregressive decoding, making per-token decoding cost a primary deployment concern. Weight quantization is the standard tool for acceleration, but representative recipes -- including state-of-the-art end-to-end (E2E) QAT -- lose accuracy on long-decoding reasoning benchmarks despite preserving perplexity and short-decode accuracy. Through a systematic gradient-direction analysis, we identify two factors driving this gap: (i) KV-cache fidelity preservation under the QAT loss, which E2E supervision attenuates via the softmax Fisher metric; and (ii) Hessian-subspace alignment between calibration data and the deployment distribution. We propose LookAhead Quantization (LAQuant), a layer-wise weight-only QAT method that addresses both factors without online-transform overhead by combining reasoning-domain calibration with a one-layer lookahead loss whose implicit cross-layer co-adaptation preserves the next-layer residual stream. For Qwen3-4B under W3G128 quantization, LAQuant improves AIME25 Pass@1 over ParoQuant by 15.11pp (1.93pp over ParoQuant++ at matched calibration) while achieving a 3.42x decoding speedup over FP16 on RTX A6000, compared with ParoQuant's 3.01x.

cs.LG↗

arXiv2026

MV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction

Establishing consistent correspondences across images is essential for 3D vision tasks such as structure-from-motion (SfM), yet most existing matchers operate in a pairwise manner, often producing fragmented and geometrically inconsistent tracks when their predictions are chained across views. We propose MV-RoMa, a multi-view dense matching model that jointly estimates dense correspondences from a source image to multiple co-visible targets. Specifically, we design an efficient model architecture which avoids high computational cost of full cross-attention for multi-view feature interaction: (i) multi-view encoder that leverages pair-wise matching results as a geometric prior, and (ii) multi-view matching refiner that refines correspondences using pixel-wise attention. Additionally, we propose a post-processing strategy that integrates our model's consistent multi-view correspondences as high-quality tracks for SfM. Across diverse and challenging benchmarks, MV-RoMa produces more reliable correspondences and substantially denser, more accurate 3D reconstructions than existing sparse and dense matching methods. Project page: https://icetea-cv.github.io/mv-roma/.

cs.CV↗

arXiv2026

EntropyCache: Decoded Token Entropy Guided KV Caching for Diffusion Language Models

Diffusion-based large language models (dLLMs) rely on bidirectional attention, which prevents lossless KV caching and requires a full forward pass at every denoising step. Existing approximate KV caching methods reduce this cost by selectively updating cached states, but their decision overhead scales with context length or model depth. We propose EntropyCache, a training-free KV caching method that uses the maximum entropy of newly decoded token distributions as a constant-cost signal for deciding when to recompute. Our design is grounded in two empirical observations: (1) decoded token entropy correlates with KV cache drift, providing a cheap proxy for cache staleness, and (2) feature volatility of decoded tokens persists for multiple steps after unmasking, motivating recomputation of the $k$ most recently decoded tokens. The skip-or-recompute decision requires only $O(V)$ computation per step, independent of context length and model scale. Experiments on LLaDA-8B-Instruct and Dream-7B-Instruct show that EntropyCache achieves $15.2\times$-$26.4\times$ speedup on standard benchmarks and $22.4\times$-$24.1\times$ on chain-of-thought benchmarks, with competitive accuracy and decision overhead accounting for only $0.5\%$ of inference time. Code is available at https://github.com/mscheong01/EntropyCache.

cs.CL↗

arXiv2026

Dense-SfM: Structure from Motion with Dense Consistent Matching

We present Dense-SfM, a novel Structure from Motion (SfM) framework designed for dense and accurate 3D reconstruction from multi-view images. Sparse keypoint matching, which traditional SfM methods often rely on, limits both accuracy and point density, especially in texture-less areas. Dense-SfM addresses this limitation by integrating dense matching with a Gaussian Splatting (GS) based track extension which gives more consistent, longer feature tracks. To further improve reconstruction accuracy, Dense-SfM is equipped with a multi-view kernelized matching module leveraging transformer and Gaussian Process architectures, for robust track refinement across multi-views. Evaluations on the ETH3D and Texture-Poor SfM datasets show that Dense-SfM offers significant improvements in accuracy and density over state-of-the-art methods. Project page: https://icetea-cv.github.io/densesfm/.

cs.CV↗

arXiv2025

Rotate, Clip, and Partition: Towards W2A4KV4 Quantization by Integrating Rotation and Learnable Non-uniform Quantizer

We propose Rotate, Clip, and Partition (RCP), a quantization-aware training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4(2-bit weight, 4-bit activation, and 4-bit KV cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design, by quantitatively analyzing the impact of random rotation on 2-bit weight quantization. Our weight quantizer features Learnable Direct Partitioning (LDP), which introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a specialized GPU kernel that supports GEMV on non-uniform W2A4. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 ppl and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code is available at https://github.com/ songsm921/RCP.

cs.LG↗

arXiv2025

Grouped Sequency-arranged Rotation: Optimizing Rotation Transformation for Quantization for Free

Large Language Models (LLMs) face deployment challenges due to high computational costs, and while Post-Training Quantization (PTQ) offers a solution, existing rotation-based methods struggle at very low bit-widths like 2-bit. We introduce a novel, training-free approach to construct an improved rotation matrix, addressing the limitations of current methods. The key contributions include leveraging the Walsh-Hadamard transform with sequency ordering, which clusters similar frequency components to reduce quantization error compared to standard Hadamard matrices, significantly improving performance. Furthermore, we propose a Grouped Sequency-arranged Rotation (GSR) using block-diagonal matrices with smaller Walsh blocks, effectively isolating outlier impacts and achieving performance comparable to optimization-based methods without requiring any training. Our method demonstrates robust performance on reasoning tasks and Perplexity (PPL) score on WikiText-2. Our method also enhances results even when applied over existing learned rotation techniques.

cs.LG↗

arXiv2025

Gaussian Weight Sampling for Scalable, Efficient and Stable Pseudo-Quantization Training

Ever-growing scale of large language models (LLMs) is pushing for improved efficiency, favoring fully quantized training (FQT) over BF16. While FQT accelerates training, it faces consistency challenges and requires searching over an exponential number of cases, each needing over 200B tokens to ensure stability. Pseudo-quantization training (PQT) addresses the issues of FQT, although it is not well-studied. We explore the practical implications of PQT in detail and propose a noise distribution $R$ that is floating-point (FP)-friendly, with ideal properties including stochastic precision annealing. As a result, the proposed method serves as an effective theoretical foundation for low-precision FP parameters through PQT, utilizing efficient fake quantization via an addition and subsequent FP casting. We demonstrate that Gaussian weight sampling is (1) scalable: supports low-precision FP parameters down to FP6 and high-precision noise up to 9-bit with BF16 operator. The proposed method is (2) efficient: incurring computational overhead as low as 1.40\% on the A100 GPU in terms of Llama2 training tokens per second, and requiring 2 bytes per parameter in GPU memory. We demonstrate that PQT with Gaussian weight sampling is (3) stable: closely following or even surpassing performance of the BF16 baseline while pre-training GPT2 and Llama2 models with up to 1B parameters and 300B tokens.

cs.LG↗

arXiv2024

Baking Relightable NeRF for Real-time Direct/Indirect Illumination Rendering

Relighting, which synthesizes a novel view under a given lighting condition (unseen in training time), is a must feature for immersive photo-realistic experience. However, real-time relighting is challenging due to high computation cost of the rendering equation which requires shape and material decomposition and visibility test to model shadow. Additionally, for indirect illumination, additional computation of rendering equation on each secondary surface point (where reflection occurs) is required rendering real-time relighting challenging. We propose a novel method that executes a CNN renderer to compute primary surface points and rendering parameters, required for direct illumination. We also present a lightweight hash grid-based renderer, for indirect illumination, which is recursively executed to perform the secondary ray tracing process. Both renderers are trained in a distillation from a pre-trained teacher model and provide real-time physically-based rendering under unseen lighting condition at a negligible loss of rendering quality.

cs.CV↗

arXiv2024

Phys3DGS: Physically-based 3D Gaussian Splatting for Inverse Rendering

We propose two novel ideas (adoption of deferred rendering and mesh-based representation) to improve the quality of 3D Gaussian splatting (3DGS) based inverse rendering. We first report a problem incurred by hidden Gaussians, where Gaussians beneath the surface adversely affect the pixel color in the volume rendering adopted by the existing methods. In order to resolve the problem, we propose applying deferred rendering and report new problems incurred in a naive application of deferred rendering to the existing 3DGS-based inverse rendering. In an effort to improve the quality of 3DGS-based inverse rendering under deferred rendering, we propose a novel two-step training approach which (1) exploits mesh extraction and utilizes a hybrid mesh-3DGS representation and (2) applies novel regularization methods to better exploit the mesh. Our experiments show that, under relighting, the proposed method offers significantly better rendering quality than the existing 3DGS-based inverse rendering methods. Compared with the SOTA voxel grid-based inverse rendering method, it gives better rendering quality while offering real-time rendering.

cs.GR↗

arXiv2024

MetaMix: Meta-state Precision Searcher for Mixed-precision Activation Quantization

Mixed-precision quantization of efficient networks often suffer from activation instability encountered in the exploration of bit selections. To address this problem, we propose a novel method called MetaMix which consists of bit selection and weight training phases. The bit selection phase iterates two steps, (1) the mixed-precision-aware weight update, and (2) the bit-search training with the fixed mixed-precision-aware weights, both of which combined reduce activation instability in mixed-precision quantization and contribute to fast and high-quality bit selection. The weight training phase exploits the weights and step sizes trained in the bit selection phase and fine-tunes them thereby offering fast training. Our experiments with efficient and hard-to-quantize networks, i.e., MobileNet v2 and v3, and ResNet-18 on ImageNet show that our proposed method pushes the boundary of mixed-precision quantization, in terms of accuracy vs. operations, by outperforming both mixed- and single-precision SOTA methods.

cs.LG↗

arXiv2024

Geometry Transfer for Stylizing Radiance Fields

Shape and geometric patterns are essential in defining stylistic identity. However, current 3D style transfer methods predominantly focus on transferring colors and textures, often overlooking geometric aspects. In this paper, we introduce Geometry Transfer, a novel method that leverages geometric deformation for 3D style transfer. This technique employs depth maps to extract a style guide, subsequently applied to stylize the geometry of radiance fields. Moreover, we propose new techniques that utilize geometric cues from the 3D scene, thereby enhancing aesthetic expressiveness and more accurately reflecting intended styles. Our extensive experiments show that Geometry Transfer enables a broader and more expressive range of stylizations, thereby significantly expanding the scope of 3D style transfer.

cs.CV↗

arXiv2024

Breaking MLPerf Training: A Case Study on Optimizing BERT

Speeding up the large-scale distributed training is challenging in that it requires improving various components of training including load balancing, communication, optimizers, etc. We present novel approaches for fast large-scale training of BERT model which individually ameliorates each component thereby leading to a new level of BERT training performance. Load balancing is imperative in distributed BERT training since its training datasets are characterized by samples with various lengths. Communication cost, which is proportional to the scale of distributed training, needs to be hidden by useful computation. In addition, the optimizers, e.g., ADAM, LAMB, etc., need to be carefully re-evaluated in the context of large-scale distributed training. We propose two new ideas, (1) local presorting based on dataset stratification for load balancing and (2) bucket-wise gradient clipping before allreduce which allows us to benefit from the overlap of gradient computation and synchronization as well as the fast training of gradient clipping before allreduce. We also re-evaluate existing optimizers via hyperparameter optimization and utilize ADAM, which also contributes to fast training via larger batches than existing methods. Our proposed methods, all combined, give the fastest MLPerf BERT training of 25.1 (22.3) seconds on 1,024 NVIDIA A100 GPUs, which is 1.33x (1.13x) and 1.57x faster than the other top two (one) submissions to MLPerf v1.1 (v2.0). Our implementation and evaluation results are available at MLPerf v1.1~v2.1.

cs.LG↗

arXiv2023

MFOS: Model-Free & One-Shot Object Pose Estimation

Existing learning-based methods for object pose estimation in RGB images are mostly model-specific or category based. They lack the capability to generalize to new object categories at test time, hence severely hindering their practicability and scalability. Notably, recent attempts have been made to solve this issue, but they still require accurate 3D data of the object surface at both train and test time. In this paper, we introduce a novel approach that can estimate in a single forward pass the pose of objects never seen during training, given minimum input. In contrast to existing state-of-the-art approaches, which rely on task-specific modules, our proposed model is entirely based on a transformer architecture, which can benefit from recently proposed 3D-geometry general pretraining. We conduct extensive experiments and report state-of-the-art one-shot performance on the challenging LINEMOD benchmark. Finally, extensive ablations allow us to determine good practices with this relatively new type of architecture in the field.

cs.CV↗

arXiv2023

NIPQ: Noise proxy-based Integrated Pseudo-Quantization

Straight-through estimator (STE), which enables the gradient flow over the non-differentiable function via approximation, has been favored in studies related to quantization-aware training (QAT). However, STE incurs unstable convergence during QAT, resulting in notable quality degradation in low precision. Recently, pseudoquantization training has been proposed as an alternative approach to updating the learnable parameters using the pseudo-quantization noise instead of STE. In this study, we propose a novel noise proxy-based integrated pseudoquantization (NIPQ) that enables unified support of pseudoquantization for both activation and weight by integrating the idea of truncation on the pseudo-quantization framework. NIPQ updates all of the quantization parameters (e.g., bit-width and truncation boundary) as well as the network parameters via gradient descent without STE instability. According to our extensive experiments, NIPQ outperforms existing quantization algorithms in various vision and language applications by a large margin.

cs.LG↗