Baeseong Park

Gunho Park, Baeseong Park, Minsub Kim et al.

Recent advances in self-supervised learning and the Transformer architecture have significantly improved natural language processing (NLP), achieving remarkably low perplexity. However, the growing size of NLP models introduces a memory wall problem during the generation phase. To mitigate this issue, recent efforts have focused on quantizing model weights to sub-4-bit precision while preserving full precision for activations, resulting in practical speed-ups during inference on a single GPU. However, these improvements primarily stem from reduced memory movement, which necessitates a resource-intensive dequantization process rather than actual computational reduction. In this paper, we introduce LUT-GEMM, an efficient kernel for quantized matrix multiplication, which not only eliminates the resource-intensive dequantization process but also reduces computational costs compared to previous kernels for weight-only quantization. Furthermore, we proposed group-wise quantization to offer a flexible trade-off between compression ratio and accuracy. The impact of LUT-GEMM is facilitated by implementing high compression ratios through low-bit quantization and efficient LUT-based operations. We show experimentally that when applied to the OPT-175B model with 3-bit quantization, LUT-GEMM substantially accelerates token generation latency, achieving a remarkable 2.1$\times$ improvement on a single GPU when compared to OPTQ, which relies on the costly dequantization process.

Se Jung Kwon, Jeonghoon Kim, Jeongin Bae et al.

There are growing interests in adapting large-scale language models using parameter-efficient fine-tuning methods. However, accelerating the model itself and achieving better inference efficiency through model compression has not been thoroughly explored yet. Model compression could provide the benefits of reducing memory footprints, enabling low-precision computations, and ultimately achieving cost-effective inference. To combine parameter-efficient adaptation and model compression, we propose AlphaTuning consisting of post-training quantization of the pre-trained language model and fine-tuning only some parts of quantized parameters for a target task. Specifically, AlphaTuning works by employing binary-coding quantization, which factorizes the full-precision parameters into binary parameters and a separate set of scaling factors. During the adaptation phase, the binary values are frozen for all tasks, while the scaling factors are fine-tuned for the downstream task. We demonstrate that AlphaTuning, when applied to GPT-2 and OPT, performs competitively with full fine-tuning on a variety of downstream tasks while achieving >10x compression ratio under 4-bit quantization and >1,000x reduction in the number of trainable parameters.

Gunho Park, Jeongin Bae, Byeongwook Kim et al.

Weight-only quantization is widely used to mitigate the memory-bound nature of LLM inference. Codebook-based methods extend this trend by achieving strong accuracy in the extremely low-bit regime (e.g., 2-bit). However, current kernels rely on dequantization, which repeatedly fetches centroids and reconstructs weights, incurring substantial latency and cache pressure. We present CodeGEMM, a codebook-centric GEMM kernel that replaces dequantization with precomputed inner products between centroids and activations stored in a lightweight Psumbook. At inference, code indices directly gather these partial sums, eliminating per-element lookups and reducing the on-chip footprint. The kernel supports the systematic exploration of latency-memory-accuracy trade-offs under a unified implementation. On Llama-3 models, CodeGEMM delivers 1.83x (8B) and 8.93x (70B) speedups in the 2-bit configuration compared to state-of-the-art codebook-based quantization at comparable accuracy and further improves computing efficiency and memory subsystem utilization.

Sunghyeon Woo, Ahreum Seo, Jaegwang Lee et al.

In multi-model LLM serving, decode execution remains inefficient due to model-specific resource partitioning: since cross-model batching is not possible, memory-bound decoding often suffers from severe GPU underutilization, especially under skewed workloads. We propose Shared Use of Next-token Prediction (SUN), the first approach that enables cross-model sharing of decode execution in disaggregated multi-LLM serving. SUN decomposes a decoder-only Transformer into a prefill module and a decode module, and fine-tunes only the task-specific prefill module, enabling a frozen decode module to be shared across models. This design enables a model-agnostic decode routing policy that balances decode requests across shared workers to maximize utilization. Across diverse tasks and model families, SUN achieves accuracy comparable to full fine-tuning while maintaining system throughput with fewer decode workers. In particular, SUN improves throughput per GPU by up to 2.0x over conventional disaggregation while keeping time-per-output-token (TPOT) within 5%. SUN inherently enables and facilitates low-bit decoding; with Quantized SUN (QSUN), it achieves a 45% speedup with comparable accuracy to SUN while preserving the benefits of shared decoding.

Sunghyeon Woo, Jaeeun Kil, Hoseung Kim et al.

Multi model inference has recently emerged as a prominent paradigm, particularly in the development of agentic AI systems. However, in such scenarios, each model must maintain its own Key-Value (KV) cache for the identical prompt, leading to substantial memory consumption. This explosive growth of KV caches forces LLM serving systems to evict previously stored caches, which in turn introduces significant recomputation overhead whenever the evicted caches are required again. Moreover, prefix caching is inherently infeasible across different models, forcing each model to recompute KV cache for the identical prompt, which leads to significant overhead. To alleviate these issues, we propose Identical Cache Reuse (ICaRus), a novel architecture that allows multiple models to share identical KV caches across all layers. ICaRus is based on the key observation that a decoder-only Transformer can be conceptually decomposed into a logical encoder, which generates KV caches, and a logical decoder, which predicts output tokens from the KV caches. ICaRus fine-tunes only the logical decoder while freezing the logical encoder, enabling multiple models to share an identical KV cache. This eliminates cache memory explosion and unexpected evictions while also allowing cross-model reuse of KV caches for new input tokens, thereby removing redundant recomputation in multi model inference achieving both efficiency and scalability. Moreover, by incorporating lightweight adapters such as LoRA, ICaRus parallelizes KV cache generation and next-token prediction during decoding. ICaRus achieves comparable accuracy to task-specific fine-tuned model across a diverse set of tasks, while allowing multiple specialized models to fully share KV caches. ICaRus achieves up to 11.1x lower P95 latency and 3.8x higher throughput in multi agent workflow with 8 different models, compared to conventional multi model system.

Jeongin Bae, Baeseong Park, Gunho Park et al.

Transformer attention is typically implemented using softmax normalization, which enforces attention weights with unit sum normalization. While effective in many settings, this constraint can limit flexibility in controlling attention magnitudes and may contribute to overly concentrated or unstable attention patterns during training. Prior work has explored modifications such as attention sinks or gating mechanisms, but these approaches provide only limited or indirect control over attention reweighting. We propose Affine-Scaled Attention, a simple extension to standard attention that introduces input-dependent scaling and a corresponding bias term applied to softmax-normalized attention weights. This design relaxes the strict normalization constraint while maintaining aggregation of value representations, allowing the model to adjust both the relative distribution and the scale of attention in a controlled manner. We empirically evaluate Affine-Scaled Attention in large-scale language model pretraining across multiple model sizes. Experimental results show consistent improvements in training stability, optimization behavior, and downstream task performance compared to standard softmax attention and attention sink baselines. These findings suggest that modest reweighting of attention outputs provides a practical and effective way to improve attention behavior in Transformer models.

Sunghyeon Woo, Hoseung Kim, Sunghwan Shim et al.

Multi-agent systems increasingly orchestrate multiple specialized language models to solve complex real-world problems, often invoking them over a shared context. This execution pattern repeatedly processes the same prompt prefix across models. Consequently, each model redundantly executes the prefill stage and maintains its own key-value (KV) cache, increasing aggregate prefill load and worsening tail latency by intensifying prefill-decode interference in existing LLM serving stacks. Disaggregated serving reduces such interference by placing prefill and decode on separate GPUs, but disaggregation does not fundamentally eliminate inter-model redundancy in computation and KV storage for the same prompt. To address this issue, we propose PrefillShare, a novel algorithm that enables sharing the prefill stage across multiple models in a disaggregated setting. PrefillShare factorizes the model into prefill and decode modules, freezes the prefill module, and fine-tunes only the decode module. This design allows multiple task-specific models to share a prefill module and the KV cache generated for the same prompt. We further introduce a routing mechanism that enables effective prefill sharing across heterogeneous models in a vLLM-based disaggregated system. PrefillShare not only matches full fine-tuning accuracy on a broad range of tasks and models, but also delivers 4.5x lower p95 latency and 3.9x higher throughput in multi-model agent workloads.

Sunghyeon Woo, Baeseong Park, Byeongwook Kim et al.

Large language models (LLMs) have achieved significant success across various domains. However, training these LLMs typically involves substantial memory and computational costs during both forward and backward propagation. While parameter-efficient fine-tuning (PEFT) considerably reduces the training memory associated with parameters, it does not address the significant computational costs and activation memory. In this paper, we propose Dropping Backward Propagation (DropBP), a novel approach designed to reduce computational costs and activation memory while maintaining accuracy. DropBP randomly drops layers during backward propagation, which is essentially equivalent to training shallow submodules generated by undropped layers and residual connections. Additionally, DropBP calculates the sensitivity of each layer to assign an appropriate drop rate, thereby stabilizing the training process. DropBP is not only applicable to full fine-tuning but can also be orthogonally integrated with all types of PEFT by dropping layers during backward propagation. Specifically, DropBP can reduce training time by 44% with comparable accuracy to the baseline, accelerate convergence to the same perplexity by 1.5x, and enable training with a sequence length 6.2x larger on a single NVIDIA-A100 GPU. Furthermore, our DropBP enabled a throughput increase of 79% on a NVIDIA A100 GPU and 117% on an Intel Gaudi2 HPU. The code is available at https://github.com/WooSunghyeon/dropbp.

Kang Min Yoo, Jaegeun Han, Sookyo In et al.

We introduce HyperCLOVA X, a family of large language models (LLMs) tailored to the Korean language and culture, along with competitive capabilities in English, math, and coding. HyperCLOVA X was trained on a balanced mix of Korean, English, and code data, followed by instruction-tuning with high-quality human-annotated datasets while abiding by strict safety guidelines reflecting our commitment to responsible AI. The model is evaluated across various benchmarks, including comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness, in both Korean and English. HyperCLOVA X exhibits strong reasoning capabilities in Korean backed by a deep understanding of the language and cultural nuances. Further analysis of the inherent bilingual nature and its extension to multilingualism highlights the model's cross-lingual proficiency and strong generalization ability to untargeted languages, including machine translation between several language pairs and cross-lingual inference tasks. We believe that HyperCLOVA X can provide helpful guidance for regions or countries in developing their sovereign LLMs.

Dongsoo Lee, Se Jung Kwon, Byeongwook Kim et al.

While model compression is increasingly important because of large neural network size, compression-aware training is challenging as it needs sophisticated model modifications and longer training time.In this paper, we introduce regularization frequency (i.e., how often compression is performed during training) as a new regularization technique for a practical and efficient compression-aware training method. For various regularization techniques, such as weight decay and dropout, optimizing the regularization strength is crucial to improve generalization in Deep Neural Networks (DNNs). While model compression also demands the right amount of regularization, the regularization strength incurred by model compression has been controlled only by compression ratio. Throughout various experiments, we show that regularization frequency critically affects the regularization strength of model compression. Combining regularization frequency and compression ratio, the amount of weight updates by model compression per mini-batch can be optimized to achieve the best model accuracy. Modulating regularization frequency is implemented by occasional model compression while conventional compression-aware training is usually performed for every mini-batch.

Baeseong Park, Se Jung Kwon, Daehwan Oh et al.

Even though fine-grained pruning techniques achieve a high compression ratio, conventional sparsity representations (such as CSR) associated with irregular sparsity degrade parallelism significantly. Practical pruning methods, thus, usually lower pruning rates (by structured pruning) to improve parallelism. In this paper, we study fixed-to-fixed (lossless) encoding architecture/algorithm to support fine-grained pruning methods such that sparse neural networks can be stored in a highly regular structure. We first estimate the maximum compression ratio of encoding-based compression using entropy. Then, as an effort to push the compression ratio to the theoretical maximum (by entropy), we propose a sequential fixed-to-fixed encoding scheme. We demonstrate that our proposed compression scheme achieves almost the maximum compression ratio for the Transformer and ResNet-50 pruned by various fine-grained pruning methods.

Byeongwook Kim, Dongsoo Lee, Yeonju Ro et al.

Various post-training uniform quantization methods have usually been studied based on convex optimization. As a result, most previous ones rely on the quantization error minimization and/or quadratic approximations. Such approaches are computationally efficient and reasonable when a large number of quantization bits are employed. When the number of quantization bits is relatively low, however, non-convex optimization is unavoidable to improve model accuracy. In this paper, we propose a new post-training uniform quantization technique considering non-convexity. We empirically show that hyper-parameters for clipping and rounding of weights and activations can be explored by monitoring task loss. Then, an optimally searched set of hyper-parameters is frozen to proceed to the next layer such that an incremental non-convex optimization is enabled for post-training quantization. Throughout extensive experimental results using various models, our proposed technique presents higher model accuracy, especially for a low-bit quantization.

Insoo Chung, Byeongwook Kim, Yoonjung Choi et al.

The deployment of widely used Transformer architecture is challenging because of heavy computation load and memory overhead during inference, especially when the target device is limited in computational resources such as mobile or edge devices. Quantization is an effective technique to address such challenges. Our analysis shows that for a given number of quantization bits, each block of Transformer contributes to translation quality and inference computations in different manners. Moreover, even inside an embedding block, each word presents vastly different contributions. Correspondingly, we propose a mixed precision quantization strategy to represent Transformer weights by an extremely low number of bits (e.g., under 3 bits). For example, for each word in an embedding block, we assign different quantization bits based on statistical property. Our quantized Transformer model achieves 11.8$\times$ smaller model size than the baseline model, with less than -0.5 BLEU. We achieve 8.3$\times$ reduction in run-time memory footprints and 3.5$\times$ speed up (Galaxy N10+) such that our proposed compression strategy enables efficient implementation for on-device NMT.

Dongsoo Lee, Se Jung Kwon, Byeongwook Kim et al.

Quantization based on the binary codes is gaining attention because each quantized bit can be directly utilized for computations without dequantization using look-up tables. Previous attempts, however, only allow for integer numbers of quantization bits, which ends up restricting the search space for compression ratio and accuracy. In this paper, we propose an encryption algorithm/architecture to compress quantized weights so as to achieve fractional numbers of bits per weight. Decryption during inference is implemented by digital XOR-gate networks added into the neural network model while XOR gates are described by utilizing $\tanh(x)$ for backward propagation to enable gradient calculations. We perform experiments using MNIST, CIFAR-10, and ImageNet to show that inserting XOR gates learns quantization/encrypted bit decisions through training and obtains high accuracy even for fractional sub 1-bit weights. As a result, our proposed method yields smaller size and higher model accuracy compared to binary neural networks.

Yongkweon Jeon, Baeseong Park, Se Jung Kwon et al.

The number of parameters in deep neural networks (DNNs) is rapidly increasing to support complicated tasks and to improve model accuracy. Correspondingly, the amount of computations and required memory footprint increase as well. Quantization is an efficient method to address such concerns by compressing DNNs such that computations can be simplified while required storage footprint is significantly reduced. Unfortunately, commercial CPUs and GPUs do not fully support quantization because only fixed data transfers (such as 32 bits) are allowed. As a result, even if weights are quantized into a few bits, CPUs and GPUs cannot access multiple quantized weights without memory bandwidth waste. Success of quantization in practice, hence, relies on an efficient computation engine design, especially for matrix multiplication that is a basic computation engine in most DNNs. In this paper, we propose a novel matrix multiplication method, called BiQGEMM, dedicated to quantized DNNs. BiQGEMM can access multiple quantized weights simultaneously in one instruction. In addition, BiQGEMM pre-computes intermediate results that are highly redundant when quantization leads to limited available computation space. Since pre-computed values are stored in lookup tables and reused, BiQGEMM achieves lower amount of overall computations. Our extensive experimental results show that BiQGEMM presents higher performance than conventional schemes when DNNs are quantized.

Se Jung Kwon, Dongsoo Lee, Byeongwook Kim et al.

Model compression techniques, such as pruning and quantization, are becoming increasingly important to reduce the memory footprints and the amount of computations. Despite model size reduction, achieving performance enhancement on devices is, however, still challenging mainly due to the irregular representations of sparse matrix formats. This paper proposes a new weight representation scheme for Sparse Quantized Neural Networks, specifically achieved by fine-grained and unstructured pruning method. The representation is encrypted in a structured regular format, which can be efficiently decoded through XOR-gate network during inference in a parallel manner. We demonstrate various deep learning models that can be compressed and represented by our proposed format with fixed and high compression ratio. For example, for fully-connected layers of AlexNet on ImageNet dataset, we can represent the sparse weights by only 0.28 bits/weight for 1-bit quantization and 91% pruning rate with a fixed decoding rate and full memory bandwidth usage. Decoding through XOR-gate network can be performed without any model accuracy degradation with additional patch data associated with small overhead.