Dongsoo Lee

Dongjun Kim, Byeonghu Na, Se Jung Kwon et al.

Whereas diverse variations of diffusion models exist, extending the linear diffusion into a nonlinear diffusion process is investigated by very few works. The nonlinearity effect has been hardly understood, but intuitively, there would be promising diffusion patterns to efficiently train the generative distribution towards the data distribution. This paper introduces a data-adaptive nonlinear diffusion process for score-based diffusion models. The proposed Implicit Nonlinear Diffusion Model (INDM) learns by combining a normalizing flow and a diffusion process. Specifically, INDM implicitly constructs a nonlinear diffusion on the \textit{data space} by leveraging a linear diffusion on the \textit{latent space} through a flow network. This flow network is key to forming a nonlinear diffusion, as the nonlinearity depends on the flow network. This flexible nonlinearity improves the learning curve of INDM to nearly Maximum Likelihood Estimation (MLE) against the non-MLE curve of DDPM++, which turns out to be an inflexible version of INDM with the flow fixed as an identity mapping. Also, the discretization of INDM shows the sampling robustness. In experiments, INDM achieves the state-of-the-art FID of 1.75 on CelebA. We release our code at https://github.com/byeonghu-na/INDM.

Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon et al.

Post-training quantization (PTQ) has been gaining popularity for the deployment of deep neural networks on resource-limited devices since unlike quantization-aware training, neither a full training dataset nor end-to-end training is required at all. As PTQ schemes based on reconstructing each layer or block output turn out to be effective to enhance quantized model performance, recent works have developed algorithms to devise and learn a new weight-rounding scheme so as to better reconstruct each layer or block output. In this work, we propose a simple yet effective new weight-rounding mechanism for PTQ, coined \emph{FlexRound}, based on element-wise division instead of typical element-wise addition such that FlexRound enables jointly learning a common quantization grid size as well as a different scale for each pre-trained weight. Thanks to the reciprocal rule of derivatives induced by element-wise division, FlexRound is inherently able to exploit pre-trained weights when updating their corresponding scales, and thus, flexibly quantize pre-trained weights depending on their magnitudes. We empirically validate the efficacy of FlexRound on a wide range of models and tasks. To the best of our knowledge, our work is the first to carry out comprehensive experiments on not only image classification and natural language understanding but also natural language generation. Moreover, we demonstrate, for the first time, that large language models can be efficiently quantized, with only a negligible impact on performance compared to half-precision baselines, achieved by reconstructing the output in a block-by-block manner. Our code is available at \url{https://github.com/onliwad101/FlexRound_LRQ}.

Jung Hyun Lee, Jeonghoon Kim, June Yong Yang et al.

With the commercialization of large language models (LLMs), weight-activation quantization has emerged to compress and accelerate LLMs, achieving high throughput while reducing inference costs. However, existing post-training quantization (PTQ) techniques for quantizing weights and activations of LLMs still suffer from non-negligible accuracy drops, especially on massive multitask language understanding. To address this issue, we propose Low-Rank Quantization (LRQ) - a simple yet effective post-training weight quantization method for LLMs that reconstructs the outputs of an intermediate Transformer block by leveraging low-rank weight-scaling matrices, replacing the conventional full weight-scaling matrices that entail as many learnable scales as their associated weights. Thanks to parameter sharing via low-rank structure, LRQ only needs to learn significantly fewer parameters while enabling the individual scaling of weights, thus boosting the generalization capability of quantized LLMs. We show the superiority of LRQ over prior LLM PTQ works under (i) 8-bit weight and per-tensor activation quantization, (ii) 4-bit weight and 8-bit per-token activation quantization, and (iii) low-bit weight-only quantization schemes. Our code is available at Software.

Seongmin Hong, Seungjae Moon, Junsoo Kim et al.

Transformer is a deep learning language model widely used for natural language processing (NLP) services in datacenters. Among transformer models, Generative Pre-trained Transformer (GPT) has achieved remarkable performance in text generation, or natural language generation (NLG), which needs the processing of a large input context in the summarization stage, followed by the generation stage that produces a single word at a time. The conventional platforms such as GPU are specialized for the parallel processing of large inputs in the summarization stage, but their performance significantly degrades in the generation stage due to its sequential characteristic. Therefore, an efficient hardware platform is required to address the high latency caused by the sequential characteristic of text generation. In this paper, we present DFX, a multi-FPGA acceleration appliance that executes GPT-2 model inference end-to-end with low latency and high throughput in both summarization and generation stages. DFX uses model parallelism and optimized dataflow that is model-and-hardware-aware for fast simultaneous workload execution among devices. Its compute cores operate on custom instructions and provide GPT-2 operations end-to-end. We implement the proposed hardware architecture on four Xilinx Alveo U280 FPGAs and utilize all of the channels of the high bandwidth memory (HBM) and the maximum number of compute resources for high hardware efficiency. DFX achieves 5.58x speedup and 3.99x energy efficiency over four NVIDIA V100 GPUs on the modern GPT-2 model. DFX is also 8.21x more cost-effective than the GPU appliance, suggesting that it is a promising solution for text generation workloads in cloud datacenters.

Jung Hwan Heo, Jeonghoon Kim, Beomseok Kwon et al.

Large Language Models (LLMs) have recently demonstrated remarkable success across various tasks. However, efficiently serving LLMs has been a challenge due to the large memory bottleneck, specifically in small batch inference settings (e.g. mobile devices). Weight-only quantization can be a promising approach, but sub-4 bit quantization remains a challenge due to large-magnitude activation outliers. To mitigate the undesirable outlier effect, we first propose per-IC quantization, a simple yet effective method that creates quantization groups within each input channel (IC) rather than the conventional per-output-channel (per-OC). Our method is motivated by the observation that activation outliers affect the input dimension of the weight matrix, so similarly grouping the weights in the IC direction can isolate outliers within a group. We also find that activation outliers do not dictate quantization difficulty, and inherent weight sensitivities also exist. With per-IC quantization as a new outlier-friendly scheme, we propose Adaptive Dimensions (AdaDim), a versatile quantization framework that can adapt to various weight sensitivity patterns. We demonstrate the effectiveness of AdaDim by augmenting prior methods such as Round-To-Nearest and GPTQ, showing significant improvements across various language modeling benchmarks for both base (up to +4.7% on MMLU) and instruction-tuned (up to +10% on HumanEval) LLMs. Code is available at https://github.com/johnheo/adadim-llm

June Yong Yang, Byeongwook Kim, Jeongin Bae et al.

Key-Value (KV) Caching has become an essential technique for accelerating the inference speed and throughput of generative Large Language Models~(LLMs). However, the memory footprint of the KV cache poses a critical bottleneck in LLM deployment as the cache size grows with batch size and sequence length, often surpassing even the size of the model itself. Although recent methods were proposed to select and evict unimportant KV pairs from the cache to reduce memory consumption, the potential ramifications of eviction on the generative process are yet to be thoroughly examined. In this paper, we examine the detrimental impact of cache eviction and observe that unforeseen risks arise as the information contained in the KV pairs is exhaustively discarded, resulting in safety breaches, hallucinations, and context loss. Surprisingly, we find that preserving even a small amount of information contained in the evicted KV pairs via reduced precision quantization substantially recovers the incurred degradation. On the other hand, we observe that the important KV pairs must be kept at a relatively higher precision to safeguard the generation quality. Motivated by these observations, we propose \textit{Mixed-precision KV cache}~(MiKV), a reliable cache compression method that simultaneously preserves the context details by retaining the evicted KV pairs in low-precision and ensure generation quality by keeping the important KV pairs in high-precision. Experiments on diverse benchmarks and LLM backbones show that our proposed method offers a state-of-the-art trade-off between compression ratio and performance, compared to other baselines.

Jiwoo Kim, Joonhyung Lee, Gunho Park et al.

As large language models (LLMs) continue to scale, the high power consumption of AI accelerators in datacenters presents significant challenges, substantially increasing the total cost of ownership (TCO) for cloud service providers (CSPs) that provide LLM inference. In this work, we analyze the computational characteristics of LLM inference from a TCO perspective and present a generalizable framework to compare AI accelerators across diverse operational requirements. Using this model, we investigate key workload characteristics influencing TCO for AI accelerators from Intel (Gaudi 2 & 3) and NVIDIA (H100 & H200), especially thin GEMM utilization and FP8 quantization. In particular, as FP8 emerges as the baseline precision for next-generation LLMs, understanding how different architectures implement and benefit from low-precision computation is increasingly critical. Throughput on thin GEMMs has a greater impact on TCO than theoretical hardware peak throughput because the memory-bound decode phase is dominated by GEMV-like computations. We find that Gaudi HPUs achieve superior utilization on thin GEMMs compared to their counterparts, especially in FP8-quantized models. Our result underscores the importance of empirical, workload-level analysis in evaluating accelerator performance, rather than relying solely on theoretical hardware specifications. By studying the interaction between power consumption, quantization strategies, and hardware architecture, we provide insights to support informed deployment decisions and guide future accelerator designs aimed at improving the TCO of LLM inference workloads.

Kanghoon Yoon, Minsub Kim, Sungjae Lee et al.

Speculative decoding accelerates LLM inference by verifying candidate tokens from a draft model against a larger target model. Recent judge decoding boosts this process by relaxing verification criteria by accepting draft tokens that may exhibit minor discrepancies from target model output, but existing methods are restricted by their reliance on human annotations or tasks with verifiable ground truths, limiting generalizability across diverse NLP tasks. We propose SelfJudge, which trains judge verifiers via self-supervision of the target model. Our method measures semantic preservation by assessing whether token-substituted responses preserve the meaning of original responses, enabling automatic verifier training across diverse NLP tasks. Our experiments show SelfJudge achieves superior inference-accuracy trade-offs than judge decoding baselines, offering a broadly applicable solution for faster LLM inference.

Jeonghoon Kim, Jung Hyun Lee, Sungdong Kim et al.

Large language models (LLMs) face the challenges in fine-tuning and deployment due to their high memory demands and computational costs. While parameter-efficient fine-tuning (PEFT) methods aim to reduce the memory usage of the optimizer state during fine-tuning, the inherent size of pre-trained LLM weights continues to be a pressing concern. Even though quantization techniques are widely proposed to ease memory demands and accelerate LLM inference, most of these techniques are geared towards the deployment phase. To bridge this gap, this paper presents Parameter-Efficient and Quantization-aware Adaptation (PEQA) - a simple yet effective method that combines the advantages of PEFT with quantized LLMs. By updating solely the quantization scales, PEQA can be directly applied to quantized LLMs, ensuring seamless task transitions. Parallel to existing PEFT methods, PEQA significantly reduces the memory overhead associated with the optimizer state. Furthermore, it leverages the advantages of quantization to substantially reduce model sizes. Even after fine-tuning, the quantization structure of a PEQA-tuned LLM remains intact, allowing for accelerated inference on the deployment stage. We employ PEQA-tuning for task-specific adaptation on LLMs with up to 65 billion parameters. To assess the logical reasoning and language comprehension of PEQA-tuned LLMs, we fine-tune low-bit quantized LLMs using a instruction dataset. Our results show that even when LLMs are quantized to below 4-bit precision, their capabilities in language modeling, few-shot in-context learning, and comprehension can be resiliently restored to (or even improved over) their full-precision original performances with PEQA.

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.

Chanwoo Kim, Dhananjaya Gowda, Dongsoo Lee et al.

In this paper, we review various end-to-end automatic speech recognition algorithms and their optimization techniques for on-device applications. Conventional speech recognition systems comprise a large number of discrete components such as an acoustic model, a language model, a pronunciation model, a text-normalizer, an inverse-text normalizer, a decoder based on a Weighted Finite State Transducer (WFST), and so on. To obtain sufficiently high speech recognition accuracy with such conventional speech recognition systems, a very large language model (up to 100 GB) is usually needed. Hence, the corresponding WFST size becomes enormous, which prohibits their on-device implementation. Recently, fully neural network end-to-end speech recognition algorithms have been proposed. Examples include speech recognition systems based on Connectionist Temporal Classification (CTC), Recurrent Neural Network Transducer (RNN-T), Attention-based Encoder-Decoder models (AED), Monotonic Chunk-wise Attention (MoChA), transformer-based speech recognition systems, and so on. These fully neural network-based systems require much smaller memory footprints compared to conventional algorithms, therefore their on-device implementation has become feasible. In this paper, we review such end-to-end speech recognition models. We extensively discuss their structures, performance, and advantages compared to conventional algorithms.

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.

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.

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

Low-rank approximation is an effective model compression technique to not only reduce parameter storage requirements, but to also reduce computations. For convolutional neural networks (CNNs), however, well-known low-rank approximation methods, such as Tucker or CP decomposition, result in degraded model accuracy because decomposed layers hinder training convergence. In this paper, we propose a new training technique that finds a flat minimum in the view of low-rank approximation without a decomposed structure during training. By preserving the original model structure, 2-dimensional low-rank approximation demanding lowering (such as im2col) is available in our proposed scheme. We show that CNN models can be compressed by low-rank approximation with much higher compression ratio than conventional training methods while maintaining or even enhancing model accuracy. We also discuss various 2-dimensional low-rank approximation techniques for CNNs.

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

Pruning is an efficient model compression technique to remove redundancy in the connectivity of deep neural networks (DNNs). Computations using sparse matrices obtained by pruning parameters, however, exhibit vastly different parallelism depending on the index representation scheme. As a result, fine-grained pruning has not gained much attention due to its irregular index form leading to large memory footprint and low parallelism for convolutions and matrix multiplications. In this paper, we propose a new network pruning technique that generates a low-rank binary index matrix to compress index data while decompressing index data is performed by simple binary matrix multiplication. This proposed compression method finds a particular fine-grained pruning mask that can be decomposed into two binary matrices. We also propose a tile-based factorization technique that not only lowers memory requirements but also enhances compression ratio. Various DNN models can be pruned with much fewer indexes compared to previous sparse matrix formats while maintaining the same pruning rate.