Large language models (LLMs) have been widely applied but face challenges in efficient inference. While quantization methods reduce computational demands, ultra-low bit quantization with arbitrary precision is hindered by limited GPU Tensor Core support and inefficient memory management, leading to suboptimal acceleration. To address these challenges, we propose a comprehensive acceleration scheme for arbitrary precision LLMs. At its core, we introduce a novel bipolar-INT data format that facilitates parallel computing and supports symmetric quantization, effectively reducing data redundancy. Building on this, we implement an arbitrary precision matrix multiplication scheme that decomposes and recovers matrices at the bit level, enabling flexible precision while maximizing GPU Tensor Core utilization. Furthermore, we develop an efficient matrix preprocessing method that optimizes data layout for subsequent computations. Finally, we design a data recovery-oriented memory management system that strategically utilizes fast shared memory, significantly enhancing kernel execution speed and minimizing memory access latency. Experimental results demonstrate our approach's effectiveness, with up to 2.4\times speedup in matrix multiplication compared to NVIDIA's CUTLASS. When integrated into LLMs, we achieve up to 6.7\times inference acceleration. These improvements significantly enhance LLM inference efficiency, enabling broader and more responsive applications of LLMs.
Despite the widespread use of statistical prior models in various fields, such models for neural network gradients have long been overlooked. The inherent challenge stems from their high-dimensional structures and complex interdependencies, which complicate effective modeling. In this work, we demonstrate the potential of large language models (LLMs) to act as gradient priors in a zero-shot setting. We examine the property by considering lossless gradient compression -- a critical application in distributed learning -- that depends heavily on precise probability modeling. To achieve this, we introduce LM-GC, a novel method that integrates LLMs with arithmetic coding. Our technique converts plain gradients into text-like formats, enhancing token efficiency by up to 38 times compared to their plain representations. We ensure that this data conversion maintains a close alignment with the structure of plain gradients and the symbols commonly recognized by LLMs. Our experiments indicate that LM-GC surpasses existing state-of-the-art lossless compression methods, improving compression rates by 10% up to 17.2% across various datasets and architectures. Additionally, our approach shows promising compatibility with lossy compression techniques such as quantization and sparsification. These findings highlight the significant potential of LLMs as a model for effectively handling gradients. Code is available at https://github.com/hui-po-wang/LM-GC.
Model quantization has become a crucial technique to address the issues of large memory consumption and long inference times associated with LLMs. Mixed-precision quantization, which distinguishes between important and unimportant parameters, stands out among numerous quantization schemes as it achieves a balance between precision and compression rate. However, existing approaches can only identify important parameters through qualitative analysis and manual experiments without quantitatively analyzing how their importance is determined. We propose a new criterion, so-called 'precision alignment', to build a quantitative framework to holistically evaluate the importance of parameters in mixed-precision quantization. Our observations on floating point addition under various real-world scenarios suggest that two addends should have identical precision, otherwise the information in the higher-precision number will be wasted. Such an observation offers an essential principle to determine the precision of each parameter in matrix multiplication operation. As the first step towards applying the above discovery to large model inference, we develop a dynamic KV-Cache quantization technique to effectively reduce memory access latency. Different from existing quantization approaches that focus on memory saving, this work directly aims to accelerate LLM inference through quantifying floating numbers. The proposed technique attains a 25% saving of memory access and delivers up to 1.3x speedup in the computation of attention in the decoding phase of LLM, with almost no loss of precision.
Federated learning (FL) is a popular machine learning technique that enables multiple users to collaboratively train a model while maintaining the user data privacy. A significant challenge in FL is the communication bottleneck in the upload direction, and thus the corresponding energy consumption of the devices, attributed to the increasing size of the model/gradient. In this paper, we address this issue by proposing a zero-order (ZO) optimization method that requires the upload of a quantized single scalar per iteration by each device instead of the whole gradient vector. We prove its theoretical convergence and find an upper bound on its convergence rate in the non-convex setting, and we discuss its implementation in practical scenarios. Our FL method and the corresponding convergence analysis take into account the impact of quantization and packet dropping due to wireless errors. We show also the superiority of our method, in terms of communication overhead and energy consumption, as compared to standard gradient-based FL methods.
In Implicit Neural Representations (INRs) a discrete signal is parameterized by a neural network that maps coordinates to the signal samples. INRs were successfully employed for encoding and compression, but such approaches are in their early stage and are still overcome by traditional codecs and autoencoders. Despite this, they have recently gained the attention of the research community due to their promising results as novel representation strategies for encoding visual content. In this paper, we propose Neural Imaging Format (NIF), an open-source INR-based image compression codec which takes advantage of a novel neural architecture which consists of two modules: a Genesis network, for mapping coordinates to pixels through bottleneck layers with sinusoidal activation units, and a Modulation network, for varying the period of the sinusoidal activations. Additionally, a final weights quantization step leads to an improvement in the compression ratio. Our proposal (NIF) consistently outperforms state-of-art INR-based compressors in terms of PSNR, by achieving comparable or better results with an outstanding up to X26 encoding speed. We also show that NIF reduces the gap between INR-based methods with respect to traditional approaches. Interestingly, our approach outperforms established codecs such as JPEG and WebP when one encodes high-resolution images at low-bitrate regimes. Extensive experiments on different datasets, a visual comparison, and an ablation study, prove the validity of the proposed approach.
Deep neural networks (DNNs) are powerful for cognitive tasks such as image classification, object detection, and scene segmentation. One drawback however is the significant high computational complexity and memory consumption, which makes them unfeasible to run real-time on embedded platforms because of the limited hardware resources. Block floating point (BFP) quantization is one of the representative compression approaches for reducing the memory and computational burden owing to their capability to effectively capture the broad data distribution of DNN models. Unfortunately, prior works on BFP-based quantization empirically choose the block size and the precision that preserve accuracy. In this paper, we develop a BFP-based bitwidth-aware analytical modeling framework (called ``BitQ'') for the best BFP implementation of DNN inference on embedded platforms. We formulate and resolve an optimization problem to identify the optimal BFP block size and bitwidth distribution by the trade-off of both accuracy and performance loss. Experimental results show that compared with an equal bitwidth setting, the BFP DNNs with optimized bitwidth allocation provide efficient computation, preserving accuracy on famous benchmarks. The source code and data are available at https://github.com/Cheliosoops/BitQ.
Referring Image Segmentation (RIS), aims to segment the object referred by a given sentence in an image by understanding both visual and linguistic information. However, existing RIS methods tend to explore top-performance models, disregarding considerations for practical applications on resources-limited edge devices. This oversight poses a significant challenge for on-device RIS inference. To this end, we propose an effective and efficient post-training quantization framework termed PTQ4RIS. Specifically, we first conduct an in-depth analysis of the root causes of performance degradation in RIS model quantization and propose dual-region quantization (DRQ) and reorder-based outlier-retained quantization (RORQ) to address the quantization difficulties in visual and text encoders. Extensive experiments on three benchmarks with different bits settings (from 8 to 4 bits) demonstrates its superior performance. Importantly, we are the first PTQ method specifically designed for the RIS task, highlighting the feasibility of PTQ in RIS applications. Code will be available at {https://github.com/gugu511yy/PTQ4RIS}.
Large Language Models (LLMs) often struggle with tasks requiring mathematical reasoning, particularly multiple-choice questions (MCQs). To address this issue, we developed LLaMa-SciQ, an educational chatbot designed to assist college students in solving and understanding MCQs in STEM fields. We begin by fine-tuning and aligning the models to human preferences. After comparing the performance of Mistral-7B and LLaMa-8B, we selected the latter as the base model due to its higher evaluation accuracy. To further enhance accuracy, we implement Retrieval-Augmented Generation (RAG) and apply quantization to compress the model, reducing inference time and increasing accessibility for students. For mathematical reasoning, LLaMa-SciQ achieved 74.5% accuracy on the GSM8k dataset and 30% on the MATH dataset. However, RAG does not improve performance and even reduces it, likely due to retriever issues or the model's unfamiliarity with context. Despite this, the quantized model shows only a 5% loss in performance, demonstrating significant efficiency improvements.
Large language models (LLMs) have achieved remarkable advancements in natural language processing, showcasing exceptional performance across various tasks. However, the expensive memory and computational requirements present significant challenges for their practical deployment. Low-bit quantization has emerged as a critical approach to mitigate these challenges by reducing the bit-width of model parameters, activations, and gradients, thus decreasing memory usage and computational demands. This paper presents a comprehensive survey of low-bit quantization methods tailored for LLMs, covering the fundamental principles, system implementations, and algorithmic strategies. An overview of basic concepts and new data formats specific to low-bit LLMs is first introduced, followed by a review of frameworks and systems that facilitate low-bit LLMs across various hardware platforms. Then, we categorize and analyze techniques and toolkits for efficient low-bit training and inference of LLMs. Finally, we conclude with a discussion of future trends and potential advancements of low-bit LLMs. Our systematic overview from basic, system, and algorithm perspectives can offer valuable insights and guidelines for future works to enhance the efficiency and applicability of LLMs through low-bit quantization.
Several recent studies have investigated low-precision accumulation, reporting improvements in throughput, power, and area across various platforms. However, the accompanying proposals have only considered the quantization-aware training (QAT) paradigm, in which models are fine-tuned or trained from scratch with quantization in the loop. As models continue to grow in size, QAT techniques become increasingly more expensive, which has motivated the recent surge in post-training quantization (PTQ) research. To the best of our knowledge, ours marks the first formal study of accumulator-aware quantization in the PTQ setting. To bridge this gap, we introduce AXE, a practical framework of accumulator-aware extensions designed to endow overflow avoidance guarantees to existing layer-wise PTQ algorithms. We theoretically motivate AXE and demonstrate its flexibility by implementing it on top of two state-of-the-art PTQ algorithms: GPFQ and OPTQ. We further generalize AXE to support multi-stage accumulation for the first time, opening the door for full datapath optimization and scaling to large language models (LLMs). We evaluate AXE across image classification and language generation models, and observe significant improvements in the trade-off between accumulator bit width and model accuracy over baseline methods.
Scaling model size significantly challenges the deployment and inference of Large Language Models (LLMs). Due to the redundancy in LLM weights, recent research has focused on pushing weight-only quantization to extremely low-bit (even down to 2 bits). It reduces memory requirements, optimizes storage costs, and decreases memory bandwidth needs during inference. However, due to numerical representation limitations, traditional scalar-based weight quantization struggles to achieve such extreme low-bit. Recent research on Vector Quantization (VQ) for LLMs has demonstrated the potential for extremely low-bit model quantization by compressing vectors into indices using lookup tables. In this paper, we introduce Vector Post-Training Quantization (VPTQ) for extremely low-bit quantization of LLMs. We use Second-Order Optimization to formulate the LLM VQ problem and guide our quantization algorithm design by solving the optimization. We further refine the weights using Channel-Independent Second-Order Optimization for a granular VQ. In addition, by decomposing the optimization problem, we propose a brief and effective codebook initialization algorithm. We also extend VPTQ to support residual and outlier quantization, which enhances model accuracy and further compresses the model. Our experimental results show that VPTQ reduces model quantization perplexity by $0.01$-$0.34$ on LLaMA-2, $0.38$-$0.68$ on Mistral-7B, $4.41$-$7.34$ on LLaMA-3 over SOTA at 2-bit, with an average accuracy improvement of $0.79$-$1.5\%$ on LLaMA-2, $1\%$ on Mistral-7B, $11$-$22\%$ on LLaMA-3 on QA tasks on average. We only utilize $10.4$-$18.6\%$ of the quantization algorithm execution time, resulting in a $1.6$-$1.8\times$ increase in inference throughput compared to SOTA.
As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.
The proliferation of Artificial Neural Networks (ANNs) has led to increased energy consumption, raising concerns about their sustainability. Spiking Neural Networks (SNNs), which are inspired by biological neural systems and operate using sparse, event-driven spikes to communicate information between neurons, offer a potential solution due to their lower energy requirements. An alternative technique for reducing a neural network's footprint is quantization, which compresses weight representations to decrease memory usage and energy consumption. In this study, we present Twin Network Augmentation (TNA), a novel training framework aimed at improving the performance of SNNs while also facilitating an enhanced compression through low-precision quantization of weights. TNA involves co-training an SNN with a twin network, optimizing both networks to minimize their cross-entropy losses and the mean squared error between their output logits. We demonstrate that TNA significantly enhances classification performance across various vision datasets and in addition is particularly effective when applied when reducing SNNs to ternary weight precision. Notably, during inference , only the ternary SNN is retained, significantly reducing the network in number of neurons, connectivity and weight size representation. Our results show that TNA outperforms traditional knowledge distillation methods and achieves state-of-the-art performance for the evaluated network architecture on benchmark datasets, including CIFAR-10, CIFAR-100, and CIFAR-10-DVS. This paper underscores the effectiveness of TNA in bridging the performance gap between SNNs and ANNs and suggests further exploration into the application of TNA in different network architectures and datasets.
The ability to process signals in digital form depends on analog-to-digital converters (ADCs). Traditionally, ADCs are designed to ensure that the digital representation closely matches the analog signal. However, recent studies have shown that significant power and memory savings can be achieved through task-based acquisition, where the acquisition process is tailored to the downstream processing task. An emerging technology for task-based acquisition involves the use of memristors, which are considered key enablers for neuromorphic computing. Memristors can implement ADCs with tunable mappings, allowing adaptation to specific system tasks or power constraints. In this work, we study task-based acquisition for a generic classification task using memristive ADCs. We consider the unique characteristics of this such neuromorphic ADCs, including their power consumption and noisy read-write behavior, and propose a physically compliant model based on resistive successive approximation register ADCs integrated with memristor components, enabling the adjustment of quantization regions. To optimize performance, we introduce a data-driven algorithm that jointly tunes task-based memristive ADCs alongside both digital and analog processing. Our design addresses the inherent stochasticity of memristors through power-aware distillation, complemented by a specialized learning algorithm that adapts to their unique analog-to-digital mapping. The proposed approach is shown to enhance accuracy by up to 27% and reduce power consumption by up to 66% compared to uniform ADCs. Even under noisy conditions, our method achieves substantial gains, with accuracy improvements of up to 19% and power reductions of up to 57%. These results highlight the effectiveness of our power-aware neuromorphic ADCs in improving system performance across diverse tasks.
For reasons such as privacy, there are use cases for language models at the edge. This has given rise to small language models (SLMs) targeted for deployment in resource-constrained devices where energy efficiency is a significant concern. Spiking neural networks (SNNs) offer a promising solution due to their energy efficiency, and there are already works on realizing transformer-based models on SNNs. However, key operations like softmax and layer normalization (LN) are difficult to implement on neuromorphic hardware, and many of these early works sidestepped them. To address these challenges, we introduce Sorbet, a transformer-based spiking language model that is more neuromorphic hardware-compatible. Sorbet incorporates a novel shifting-based softmax called PTsoftmax and a power normalization method using bit-shifting (BSPN), both designed to replace the respective energy-intensive operations. By leveraging knowledge distillation and model quantization, Sorbet achieved a highly compressed binary weight model that maintains competitive performance while significantly reducing energy consumption. We validate Sorbet's effectiveness through extensive testing on the GLUE benchmark and a series of ablation studies, demonstrating its potential as an energy-efficient solution for language model inference.
Image vectorization converts raster images into vector graphics composed of regions separated by curves. Typical vectorization methods first define the regions by grouping similar colored regions via color quantization, then approximate their boundaries by Bezier curves. In that way, the raster input is converted into an SVG format parameterizing the regions' colors and the Bezier control points. This compact representation has many graphical applications thanks to its universality and resolution-independence. In this paper, we remark that image vectorization is nothing but an image segmentation, and that it can be built by fine to coarse region merging. Our analysis of the problem leads us to propose a vectorization method alternating region merging and curve smoothing. We formalize the method by alternate operations on the dual and primal graph induced from any domain partition. In that way, we address a limitation of current vectorization methods, which separate the update of regional information from curve approximation. We formalize region merging methods by associating them with various gain functionals, including the classic Beaulieu-Goldberg and Mumford-Shah functionals. More generally, we introduce and compare region merging criteria involving region number, scale, area, and internal standard deviation. We also show that the curve smoothing, implicit in all vectorization methods, can be performed by the shape-preserving affine scale space. We extend this flow to a network of curves and give a sufficient condition for the topological preservation of the segmentation. The general vectorization method that follows from this analysis shows explainable behaviors, explicitly controlled by a few intuitive parameters. It is experimentally compared to state-of-the-art software and proved to have comparable or superior fidelity and cost efficiency.
Tensor Networks (TNs) are a computational paradigm used for representing quantum many-body systems. Recent works have shown how TNs can also be applied to perform Machine Learning (ML) tasks, yielding comparable results to standard supervised learning techniques. In this work, we study the use of Tree Tensor Networks (TTNs) in high-frequency real-time applications by exploiting the low-latency hardware of the Field-Programmable Gate Array (FPGA) technology. We present different implementations of TTN classifiers, capable of performing inference on classical ML datasets as well as on complex physics data. A preparatory analysis of bond dimensions and weight quantization is realized in the training phase, together with entanglement entropy and correlation measurements, that help setting the choice of the TTN architecture. The generated TTNs are then deployed on a hardware accelerator; using an FPGA integrated into a server, the inference of the TTN is completely offloaded. Eventually, a classifier for High Energy Physics (HEP) applications is implemented and executed fully pipelined with sub-microsecond latency.
Zero-shot singing voice synthesis (SVS) with style transfer and style control aims to generate high-quality singing voices with unseen timbres and styles (including singing method, emotion, rhythm, technique, and pronunciation) from audio and text prompts. However, the multifaceted nature of singing styles poses a significant challenge for effective modeling, transfer, and control. Furthermore, current SVS models often fail to generate singing voices rich in stylistic nuances for unseen singers. To address these challenges, we introduce TCSinger, the first zero-shot SVS model for style transfer across cross-lingual speech and singing styles, along with multi-level style control. Specifically, TCSinger proposes three primary modules: 1) the clustering style encoder employs a clustering vector quantization model to stably condense style information into a compact latent space; 2) the Style and Duration Language Model (S\&D-LM) concurrently predicts style information and phoneme duration, which benefits both; 3) the style adaptive decoder uses a novel mel-style adaptive normalization method to generate singing voices with enhanced details. Experimental results show that TCSinger outperforms all baseline models in synthesis quality, singer similarity, and style controllability across various tasks, including zero-shot style transfer, multi-level style control, cross-lingual style transfer, and speech-to-singing style transfer. Singing voice samples can be accessed at https://aaronz345.github.io/TCSingerDemo/.
Dynamic quantization has attracted rising attention in image super-resolution (SR) as it expands the potential of heavy SR models onto mobile devices while preserving competitive performance. Existing methods explore layer-to-bit configuration upon varying local regions, adaptively allocating the bit to each layer and patch. Despite the benefits, they still fall short in the trade-off of SR accuracy and quantization efficiency. Apart from this, adapting the quantization level for each layer individually can disturb the original inter-layer relationships, thus diminishing the representation capability of quantized models. In this work, we propose Granular-DQ, which capitalizes on the intrinsic characteristics of images while dispensing with the previous consideration for layer sensitivity in quantization. Granular-DQ conducts a multi-granularity analysis of local patches with further exploration of their information densities, achieving a distinctive patch-wise and layer-invariant dynamic quantization paradigm. Specifically, Granular-DQ initiates by developing a granularity-bit controller (GBC) to apprehend the coarse-to-fine granular representations of different patches, matching their proportional contribution to the entire image to determine the proper bit-width allocation. On this premise, we investigate the relation between bit-width and information density, devising an entropy-to-bit (E2B) mechanism that enables further fine-grained dynamic bit adaption of high-bit patches. Extensive experiments validate the superiority and generalization ability of Granular-DQ over recent state-of-the-art methods on various SR models. Code and supplementary statement can be found at \url{https://github.com/MmmingS/Granular-DQ.git}.
Disentangled representation learning aims to represent the underlying generative factors of a dataset in a latent representation independently of one another. In our work, we propose a discrete variational autoencoder (VAE) based model where the ground truth information about the generative factors are not provided to the model. We demonstrate the advantages of learning discrete representations over learning continuous representations in facilitating disentanglement. Furthermore, we propose incorporating an inductive bias into the model to further enhance disentanglement. Precisely, we propose scalar quantization of the latent variables in a latent representation with scalar values from a global codebook, and we add a total correlation term to the optimization as an inductive bias. Our method called FactorQVAE combines optimization based disentanglement approaches with discrete representation learning, and it outperforms the former disentanglement methods in terms of two disentanglement metrics (DCI and InfoMEC) while improving the reconstruction performance. Our code can be found at https://github.com/ituvisionlab/FactorQVAE.