Cloud gaming is an advanced form of Internet service that necessitates local terminals to decode within limited resources and time latency. Super-Resolution (SR) techniques are often employed on these terminals as an efficient way to reduce the required bit-rate bandwidth for cloud gaming. However, insufficient attention has been paid to SR of compressed game video content. Most SR networks amplify block artifacts and ringing effects in decoded frames while ignoring edge details of game content, leading to unsatisfactory reconstruction results. In this paper, we propose a novel lightweight network called Coding Prior-Guided Super-Resolution (CPGSR) to address the SR challenges in compressed game video content. First, we design a Compressed Domain Guided Block (CDGB) to extract features of different depths from coding priors, which are subsequently integrated with features from the U-net backbone. Then, a series of re-parameterization blocks are utilized for reconstruction. Ultimately, inspired by the quantization in video coding, we propose a partitioned focal frequency loss to effectively guide the model's focus on preserving high-frequency information. Extensive experiments demonstrate the advancement of our approach.
Accelerating machine learning inference has been an active research area in recent years. In this context, field-programmable gate arrays (FPGAs) have demonstrated compelling performance by providing massive parallelism in deep neural networks (DNNs). Neural networks (NNs) are computationally intensive during inference, as they require massive amounts of multiplication and addition, which makes their implementations costly. Numerous studies have recently addressed this challenge to some extent using a combination of sparsity induction, quantization, and transformation of neurons or sub-networks into lookup tables (LUTs) on FPGAs. Gradient boosted decision trees (GBDTs) are a high-accuracy alternative to DNNs in a wide range of regression and classification tasks, particularly for tabular datasets. The basic building block of GBDTs is a decision tree, which resembles the structure of binary decision diagrams. FPGA design flows are heavily optimized to implement such a structure efficiently. In addition to decision trees, GBDTs perform simple operations during inference, including comparison and addition. We present TreeLUT as an open-source tool for implementing GBDTs using an efficient quantization scheme, hardware architecture, and pipelining strategy. It primarily utilizes LUTs with no BRAMs or DSPs on FPGAs, resulting in high efficiency. We show the effectiveness of TreeLUT using multiple classification datasets, commonly used to evaluate ultra-low area and latency architectures. Using these benchmarks, we compare our implementation results with existing DNN and GBDT methods, such as DWN, PolyLUT-Add, NeuraLUT, LogicNets, FINN, hls4ml, and others. Our results show that TreeLUT significantly improves hardware utilization, latency, and throughput at competitive accuracy compared to previous works.
End-to-end image and video codecs are becoming increasingly competitive, compared to traditional compression techniques that have been developed through decades of manual engineering efforts. These trainable codecs have many advantages over traditional techniques, such as their straightforward adaptation to perceptual distortion metrics and high performance in specific fields thanks to their learning ability. However, current state-of-the-art neural codecs do not fully exploit the benefits of vector quantization and the existence of the entropy gradient in decoding devices. In this paper, we propose to leverage these two properties (vector quantization and entropy gradient) to improve the performance of off-the-shelf codecs. Firstly, we demonstrate that using non-uniform scalar quantization cannot improve performance over uniform quantization. We thus suggest using predefined optimal uniform vector quantization to improve performance. Secondly, we show that the entropy gradient, available at the decoder, is correlated with the reconstruction error gradient, which is not available at the decoder. We therefore use the former as a proxy to enhance compression performance. Our experimental results show that these approaches save between 1 to 3% of the rate for the same quality across various pretrained methods. In addition, the entropy gradient based solution improves traditional codec performance significantly as well.
Recent years have witnessed the success of foundation models pre-trained with self-supervised learning (SSL) in various music informatics understanding tasks, including music tagging, instrument classification, key detection, and more. In this paper, we propose a self-supervised music representation learning model for music understanding. Distinguished from previous studies adopting random projection or existing neural codec, the proposed model, named MuQ, is trained to predict tokens generated by Mel Residual Vector Quantization (Mel-RVQ). Our Mel-RVQ utilizes residual linear projection structure for Mel spectrum quantization to enhance the stability and efficiency of target extraction and lead to better performance. Experiments in a large variety of downstream tasks demonstrate that MuQ outperforms previous self-supervised music representation models with only 0.9K hours of open-source pre-training data. Scaling up the data to over 160K hours and adopting iterative training consistently improve the model performance. To further validate the strength of our model, we present MuQ-MuLan, a joint music-text embedding model based on contrastive learning, which achieves state-of-the-art performance in the zero-shot music tagging task on the MagnaTagATune dataset. Code and checkpoints are open source in https://github.com/tencent-ailab/MuQ.
Diffusionmodels(DMs)havedemonstratedremarkableachievements in synthesizing images of high fidelity and diversity. However, the extensive computational requirements and slow generative speed of diffusion models have limited their widespread adoption. In this paper, we propose a novel post-training quantization for diffusion models (PQD), which is a time-aware optimization framework for diffusion models based on post-training quantization. The proposed framework optimizes the inference process by selecting representative samples and conducting time-aware calibration. Experimental results show that our proposed method is able to directly quantize full-precision diffusion models into 8-bit or 4-bit models while maintaining comparable performance in a training-free manner, achieving a few FID change on ImageNet for unconditional image generation. Our approach demonstrates compatibility and can also be applied to 512x512 text-guided image generation for the first time.
Transformers have shown impressive results in tabular data generation. However, they lack domain-specific inductive biases which are critical for preserving the intrinsic characteristics of tabular data. They also suffer from poor scalability and efficiency due to quadratic computational complexity. In this paper, we propose TabTreeFormer, a hybrid transformer architecture that integrates inductive biases of tree-based models (i.e., non-smoothness and non-rotational invariance) to effectively handle the discrete and weakly correlated features in tabular datasets. To improve numerical fidelity and capture multimodal distributions, we introduce a novel tokenizer that learns token sequences based on the complexity of tabular values. This reduces vocabulary size and sequence length, yielding more compact and efficient representations without sacrificing performance. We evaluate TabTreeFormer on nine diverse datasets, benchmarking against eight generative models. We show that TabTreeFormer consistently outperforms baselines in utility, fidelity, and privacy metrics with competitive efficiency. Notably, in scenarios prioritizing data utility over privacy and efficiency, the best variant of TabTreeFormer delivers a 44% performance gain relative to its baseline variant.
In this paper, we present an intuitive analysis of the optimization technique based on the quantization of an objective function. Quantization of an objective function is an effective optimization methodology that decreases the measure of a level set containing several saddle points and local minima and finds the optimal point at the limit level set. To investigate the dynamics of quantization-based optimization, we derive an overdamped Langevin dynamics model from an intuitive analysis to minimize the level set by iterative quantization. We claim that quantization-based optimization involves the quantities of thermodynamical and quantum mechanical optimization as the core methodologies of global optimization. Furthermore, on the basis of the proposed SDE, we provide thermodynamic and quantum mechanical analysis with Witten-Laplacian. The simulation results with the benchmark functions, which compare the performance of the nonlinear optimization, demonstrate the validity of the quantization-based optimization.
The rapidly increasing size of large language models (LLMs) presents significant challenges in memory usage and computational costs. Quantizing both weights and activations can address these issues, with hardware-supported fine-grained scaling emerging as a promising solution to mitigate outliers. However, existing methods struggle to capture nuanced block data distributions. We propose BlockDialect, a block-wise fine-grained mixed format technique that assigns a per-block optimal number format from a formatbook for better data representation. Additionally, we introduce DialectFP4, a formatbook of FP4 variants (akin to dialects) that adapt to diverse data distributions. To leverage this efficiently, we propose a two-stage approach for online DialectFP4 activation quantization. Importantly, DialectFP4 ensures energy efficiency by selecting representable values as scaled integers compatible with low-precision integer arithmetic. BlockDialect achieves 10.78% (7.48%) accuracy gain on the LLaMA3-8B (LLaMA2-7B) model compared to MXFP4 format with lower bit usage per data, while being only 5.45% (2.69%) below full precision even when quantizing full-path matrix multiplication. Focusing on how to represent over how to scale, our work presents a promising path for energy-efficient LLM inference.
This paper describes a memory-efficient transformer model designed to drive a reduction in memory usage and execution time by substantial orders of magnitude without impairing the model's performance near that of the original model. Recently, new architectures of transformers were presented, focused on parameter efficiency and computational optimization; however, such models usually require considerable resources in terms of hardware when deployed in real-world applications on edge devices. This approach addresses this concern by halving embedding size and applying targeted techniques such as parameter pruning and quantization to optimize the memory footprint with minimum sacrifices in terms of accuracy. Experimental results include a 52% reduction in memory usage and a 33% decrease in execution time, resulting in better efficiency than state-of-the-art models. This work compared our model with existing compelling architectures, such as MobileBERT and DistilBERT, and proved its feasibility in the domain of resource-friendly deep learning architectures, mainly for applications in real-time and in resource-constrained applications.
Large language models (LLMs) have transformed the way we think about language understanding and generation, enthralling both researchers and developers. However, deploying LLMs for inference has been a significant challenge due to their unprecedented size and resource requirements. While quantizing model weights to sub-byte precision has emerged as a promising solution to ease memory pressure, the group quantization formats commonly used for LLM quantization have significant compute overheads and a resource-intensive dequantization process. As a result, a higher proportion of compute instructions do not perform multiplies, i.e., real work, rendering them unsuitable for meeting the required latency requirements for LLMs deployed on commodity CPUs. In this work, we propose a set of highly optimized kernels to accelerate LLM inference and unleash the full potential of CPUs, particularly Arm CPUs. These kernels amortize the cost of loading the operands and the cost of weight unpacking across multiple output rows. This, along with the introduction of an optimized interleaved group data layout for weights and decompression path optimizations to reduce unnecessary operations and dequantization overhead while maximizing the use of vector and matrix multiply operations, significantly improves the efficiency of MAC operations. Furthermore, we present a groupwise non-uniform codebook-based quantization method for ultra-low-precision quantization of LLMs to better match non-uniform patterns in their weight distributions, demonstrating better throughput during token generation while ensuring better quality than the state-of-the-art. Applying these improvements to 4-bit LLMs results in a 3-3.2x improvement in prompt processing and a 2x improvement in autoregressive decoding on Arm CPUs, compared to LLaMA.cpp-based solution. The optimized kernels are available at https://github.com/ggerganov/llama.cpp.
Acoustic Scene Classification (ASC) identifies an environment based on an audio signal. This paper explores ASC in low-resource conditions and proposes a novel model, DS-FlexiNet, which combines depthwise separable convolutions from MobileNetV2 with ResNet-inspired residual connections for a balance of efficiency and accuracy. To address hardware limitations and device heterogeneity, DS-FlexiNet employs Quantization Aware Training (QAT) for model compression and data augmentation methods like Auto Device Impulse Response (ADIR) and Freq-MixStyle (FMS) to improve cross-device generalization. Knowledge Distillation (KD) from twelve teacher models further enhances performance on unseen devices. The architecture includes a custom Residual Normalization layer to handle domain differences across devices, and depthwise separable convolutions reduce computational overhead without sacrificing feature representation. Experimental results show that DS-FlexiNet excels in both adaptability and performance under resource-constrained conditions.
Low-rank adaptation (LoRA) reduces the computational and memory demands of fine-tuning large language models (LLMs) by approximating updates with low-rank matrices. However, low-rank approximation in two-dimensional space fails to capture high-dimensional structures within the target matrix. Recently, tensor decomposition methods have been explored for fine-tuning LLMs, leveraging their ability to extract structured information. Yet, these approaches primarily rely on random initialization, and the impact of initialization on tensor adaptation remains underexplored. In this paper, we reveal that random initialization significantly diverges from the validation loss achieved by full fine-tuning. To address this, we propose Weight-Decomposed Tensor Adaptation (DoTA), which leverages the Matrix Product Operator (MPO) decomposition of pre-trained weights for effective initialization in fine-tuning LLMs. Additionally, we introduce QDoTA, a quantized version of DoTA designed for 4-bit quantization. Experiments on commonsense and arithmetic reasoning tasks show that DoTA outperforms random initialization methods with fewer parameters. QDoTA further reduces memory consumption and achieves comparable performance to DoTA on commonsense reasoning tasks. We will release our code to support future research.
Federated Learning (FL) enables clients to share learning parameters instead of local data, reducing communication overhead. Traditional wireless networks face latency challenges with FL. In contrast, Cell-Free Massive MIMO (CFmMIMO) can serve multiple clients on shared resources, boosting spectral efficiency and reducing latency for large-scale FL. However, clients' communication resource limitations can hinder the completion of the FL training. To address this challenge, we propose an energy-efficient, low-latency FL framework featuring optimized uplink power allocation for seamless client-server collaboration. Our framework employs an adaptive quantization scheme, dynamically adjusting bit allocation for local gradient updates to reduce communication costs. We formulate a joint optimization problem covering FL model updates, local iterations, and power allocation, solved using sequential quadratic programming (SQP) to balance energy and latency. Additionally, clients use the AdaDelta method for local FL model updates, enhancing local model convergence compared to standard SGD, and we provide a comprehensive analysis of FL convergence with AdaDelta local updates. Numerical results show that, within the same energy and latency budgets, our power allocation scheme outperforms the Dinkelbach and max-sum rate methods by increasing the test accuracy up to $7$\% and $19$\%, respectively. Moreover, for the three power allocation methods, our proposed quantization scheme outperforms AQUILA and LAQ by increasing test accuracy by up to $36$\% and $35$\%, respectively.
Processing long temporal sequences is a key challenge in deep learning. In recent years, Transformers have become state-of-the-art for this task, but suffer from excessive memory requirements due to the need to explicitly store the sequences. To address this issue, structured state-space sequential (S4) models recently emerged, offering a fixed memory state while still enabling the processing of very long sequence contexts. The recurrent linear update of the state in these models makes them highly efficient on modern graphics processing units (GPU) by unrolling the recurrence into a convolution. However, this approach demands significant memory and massively parallel computation, which is only available on the latest GPUs. In this work, we aim to bring the power of S4 models to edge hardware by significantly reducing the size and computational demand of an S4D model through quantization-aware training, even achieving ternary weights for a simple real-world task. To this end, we extend conventional quantization-aware training to tailor it for analog in-memory compute hardware. We then demonstrate the deployment of recurrent S4D kernels on memrisitve crossbar arrays, enabling their computation in an in-memory compute fashion. To our knowledge, this is the first implementation of S4 kernels on in-memory compute hardware.
Quantization of Large Language Models (LLMs) has recently gained popularity, particularly for on-device settings with limited hardware resources. While efficient, quantization inevitably degrades model quality, especially in aggressive low-bit settings such as 3-bit and 4-bit precision. In this paper, we propose QDEC, an inference scheme that improves the quality of low-bit LLMs while preserving the key benefits of quantization: GPU memory savings and inference latency reduction. QDEC stores the residual matrix -- the difference between full-precision and quantized weights -- in CPU, and dynamically fetches the residuals for only a small portion of the weights. This portion corresponds to the salient channels, marked by activation outliers, with the fetched residuals helping to correct quantization errors in these channels. Salient channels are identified dynamically at each decoding step by analyzing the input activations -- this allows for the adaptation to the dynamic nature of activation distribution, and thus maximizes the effectiveness of error compensation. We demonstrate the effectiveness of QDEC by augmenting state-of-the-art quantization methods. For example, QDEC reduces the perplexity of a 3-bit Llama-3-8B-Instruct model from 10.15 to 9.12 -- outperforming its 3.5-bit counterpart -- while adding less than 0.0003\% to GPU memory usage and incurring only a 1.7\% inference slowdown on NVIDIA RTX 4050 Mobile GPU. The code will be publicly available soon.
Visual Mamba is an approach that extends the selective space state model, Mamba, to vision tasks. It processes image tokens sequentially in a fixed order, accumulating information to generate outputs. Despite its growing popularity for delivering high-quality outputs at a low computational cost across various tasks, Visual Mamba is highly susceptible to quantization, which makes further performance improvements challenging. Our analysis reveals that the fixed token access order in Visual Mamba introduces unique quantization challenges, which we categorize into three main issues: 1) token-wise variance, 2) channel-wise outliers, and 3) a long tail of activations. To address these challenges, we propose Post-Training Quantization for Visual Mamba (PTQ4VM), which introduces two key strategies: Per-Token Static (PTS) quantization and Joint Learning of Smoothing Scale and Step Size (JLSS). To the our best knowledge, this is the first quantization study on Visual Mamba. PTQ4VM can be applied to various Visual Mamba backbones, converting the pretrained model to a quantized format in under 15 minutes without notable quality degradation. Extensive experiments on large-scale classification and regression tasks demonstrate its effectiveness, achieving up to 1.83x speedup on GPUs with negligible accuracy loss compared to FP16. Our code is available at https://github.com/YoungHyun197/ptq4vm.
Despite the revolutionary breakthroughs of large-scale text-to-image diffusion models for complex vision and downstream tasks, their extremely high computational and storage costs limit their usability. Quantization of diffusion models has been explored in recent works to reduce compute costs and memory bandwidth usage. To further improve inference time, fast convolution algorithms such as Winograd can be used for convolution layers, which account for a significant portion of computations in diffusion models. However, the significant quality loss of fully quantized Winograd using existing coarser-grained post-training quantization methods, combined with the complexity and cost of finetuning the Winograd transformation matrices for such large models to recover quality, makes them unsuitable for large-scale foundation models. Motivated by the presence of a large range of values in them, we investigate the impact of finer-grained group-wise quantization in quantizing diffusion models. While group-wise quantization can largely handle the fully quantized Winograd convolution, it struggles to deal with the large distribution imbalance in a sizable portion of the Winograd domain computation. To reduce range differences in the Winograd domain, we propose finetuning only the scale parameters of the Winograd transform matrices without using any domain-specific training data. Because our method does not depend on any training data, the generalization performance of quantized diffusion models is safely guaranteed. For text-to-image generation task, the 8-bit fully-quantized diffusion model with Winograd provides near-lossless quality (FID and CLIP scores) in comparison to the full-precision model. For image classification, our method outperforms the state-of-the-art Winograd PTQ method by 1.62% and 2.56% in top-1 ImageNet accuracy on ResNet18 and ResNet-34, respectively, with Winograd F(6, 3).
As cutting-edge large language models (LLMs) continue to transform various industries, their fast-growing model size and sequence length have led to memory traffic and capacity challenges. Recently, AMD, Arm, Intel, Meta, Microsoft, NVIDIA, and Qualcomm have proposed a Microscaling standard (Mx), which augments block floating-point with microexponents to achieve promising perplexity-to-footprint trade-offs. However, the Microscaling suffers from significant perplexity degradation on modern LLMs with less than six bits. This paper profiles modern LLMs and identifies three main challenges of low-bit Microscaling format, i.e., inaccurate tracking of outliers, vacant quantization levels, and wasted binary code. In response, Nanoscaling (NxFP) proposes three techniques, i.e., NanoMantissa, Adaptive Microexponent, and Code Recycling to enable better accuracy and smaller memory footprint than state-of-the-art MxFP. Experimental results on direct-cast inference across various modern LLMs demonstrate that our proposed methods outperform state-of-the-art MxFP by up to 0.64 in perplexity and by up to 30% in accuracy on MMLU benchmarks. Furthermore, NxFP reduces memory footprint by up to 16% while achieving comparable perplexity as MxFP.
Recent research in the domain of multimodal unified representations predominantly employs codebook as representation forms, utilizing Vector Quantization(VQ) for quantization, yet there has been insufficient exploration of other quantization representation forms. Our work explores more precise quantization methods and introduces a new framework, Semantic Residual Cross-modal Information Disentanglement (SRCID), inspired by the numerical residual concept inherent to Residual Vector Quantization (RVQ). SRCID employs semantic residual-based information disentanglement for multimodal data to better handle the inherent discrepancies between different modalities. Our method enhances the capabilities of unified multimodal representations and demonstrates exceptional performance in cross-modal generalization and cross-modal zero-shot retrieval. Its average results significantly surpass existing state-of-the-art models, as well as previous attempts with RVQ and Finite Scalar Quantization (FSQ) based on these modals.
Real-time speech enhancement (SE) is essential to online speech communication. Causal SE models use only the previous context while predicting future information, such as phoneme continuation, may help performing causal SE. The phonetic information is often represented by quantizing latent features of self-supervised learning (SSL) models. This work is the first to incorporate SSL features with causality into an SE model. The causal SSL features are encoded and combined with spectrogram features using feature-wise linear modulation to estimate a mask for enhancing the noisy input speech. Simultaneously, we quantize the causal SSL features using vector quantization to represent phonetic characteristics as semantic tokens. The model not only encodes SSL features but also predicts the future semantic tokens in multi-task learning (MTL). The experimental results using VoiceBank + DEMAND dataset show that our proposed method achieves 2.88 in PESQ, especially with semantic prediction MTL, in which we confirm that the semantic prediction played an important role in causal SE.