We present 1.58-bit FLUX, the first successful approach to quantizing the state-of-the-art text-to-image generation model, FLUX.1-dev, using 1.58-bit weights (i.e., values in {-1, 0, +1}) while maintaining comparable performance for generating 1024 x 1024 images. Notably, our quantization method operates without access to image data, relying solely on self-supervision from the FLUX.1-dev model. Additionally, we develop a custom kernel optimized for 1.58-bit operations, achieving a 7.7x reduction in model storage, a 5.1x reduction in inference memory, and improved inference latency. Extensive evaluations on the GenEval and T2I Compbench benchmarks demonstrate the effectiveness of 1.58-bit FLUX in maintaining generation quality while significantly enhancing computational efficiency.
Vision-Language Models (VLMs) have enabled a variety of real-world applications. The large parameter size of VLMs brings large memory and computation overhead which poses significant challenges for deployment. Post-Training Quantization (PTQ) is an effective technique to reduce the memory and computation overhead. Existing PTQ methods mainly focus on large language models (LLMs), without considering the differences across other modalities. In this paper, we discover that there is a significant difference in sensitivity between language and vision tokens in large VLMs. Therefore, treating tokens from different modalities equally, as in existing PTQ methods, may over-emphasize the insensitive modalities, leading to significant accuracy loss. To deal with the above issue, we propose a simple yet effective method, Modality-Balanced Quantization (MBQ), for large VLMs. Specifically, MBQ incorporates the different sensitivities across modalities during the calibration process to minimize the reconstruction loss for better quantization parameters. Extensive experiments show that MBQ can significantly improve task accuracy by up to 4.4% and 11.6% under W3 and W4A8 quantization for 7B to 70B VLMs, compared to SOTA baselines. Additionally, we implement a W3 GPU kernel that fuses the dequantization and GEMV operators, achieving a 1.4x speedup on LLaVA-onevision-7B on the RTX 4090. The code is available at https://github.com/thu-nics/MBQ.
Large Language Models (LLMs) have revolutionized a wide range of domains such as natural language processing, computer vision, and multi-modal tasks due to their ability to comprehend context and perform logical reasoning. However, the computational and memory demands of LLMs, particularly during inference, pose significant challenges when scaling them to real-world, long-context, and real-time applications. Key-Value (KV) cache management has emerged as a critical optimization technique for accelerating LLM inference by reducing redundant computations and improving memory utilization. This survey provides a comprehensive overview of KV cache management strategies for LLM acceleration, categorizing them into token-level, model-level, and system-level optimizations. Token-level strategies include KV cache selection, budget allocation, merging, quantization, and low-rank decomposition, while model-level optimizations focus on architectural innovations and attention mechanisms to enhance KV reuse. System-level approaches address memory management, scheduling, and hardware-aware designs to improve efficiency across diverse computing environments. Additionally, the survey provides an overview of both text and multimodal datasets and benchmarks used to evaluate these strategies. By presenting detailed taxonomies and comparative analyses, this work aims to offer useful insights for researchers and practitioners to support the development of efficient and scalable KV cache management techniques, contributing to the practical deployment of LLMs in real-world applications. The curated paper list for KV cache management is in: \href{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}{https://github.com/TreeAI-Lab/Awesome-KV-Cache-Management}.
The availability of pre-trained models (PTMs) has enabled faster deployment of machine learning across applications by reducing the need for extensive training. Techniques like quantization and distillation have further expanded PTM applicability to resource-constrained IoT hardware. Given the many PTM options for any given task, engineers often find it too costly to evaluate each model's suitability. Approaches such as LogME, LEEP, and ModelSpider help streamline model selection by estimating task relevance without exhaustive tuning. However, these methods largely leave hardware constraints as future work-a significant limitation in IoT settings. In this paper, we identify the limitations of current model recommendation approaches regarding hardware constraints and introduce a novel, hardware-aware method for PTM selection. We also propose a research agenda to guide the development of effective, hardware-conscious model recommendation systems for IoT applications.
We introduce AKT (Advanced Knowledge Transfer), a novel method to enhance the training ability of low-bit quantized (Q) models in the field of zero-shot quantization (ZSQ). Existing research in ZSQ has focused on generating high-quality data from full-precision (FP) models. However, these approaches struggle with reduced learning ability in low-bit quantization due to its limited information capacity. To overcome this limitation, we propose effective training strategy compared to data generation. Particularly, we analyzed that refining feature maps in the feature distillation process is an effective way to transfer knowledge to the Q model. Based on this analysis, AKT efficiently transfer core information from the FP model to the Q model. AKT is the first approach to utilize both spatial and channel attention information in feature distillation in ZSQ. Our method addresses the fundamental gradient exploding problem in low-bit Q models. Experiments on CIFAR-10 and CIFAR-100 datasets demonstrated the effectiveness of the AKT. Our method led to significant performance enhancement in existing generative models. Notably, AKT achieved significant accuracy improvements in low-bit Q models, achieving state-of-the-art in the 3,5bit scenarios on CIFAR-10. The code is available at https://github.com/Inpyo-Hong/AKT-Advanced-knowledge-Transfer.
The need to digitize signals with intricate spectral characteristics often challenges traditional analog-to-digital converters (ADCs). The recently proposed modulo-ADC architecture offers a promising alternative by leveraging inherent features of the input signals. This approach can dramatically reduce the number of bits required for the conversion while maintaining the desired fidelity. However, the core algorithm of this architecture, which utilizes a prediction filter, functions properly only when the respective prediction error is bounded. In practice, this assumption may not always hold, leading to considerable instability and performance degradation. To address this limitation, we propose an enhanced modulo-unfolding solution without this assumption. We develop a reliable detector to successfully unfold the signals, yielding a robust solution. Consequently, the reinforced system maintains proper operation in scenarios where the original approach fails, while also reducing the quantization noise. We present simulation results that demonstrate the superior performance of our approach in a representative setting.
As Large Language Models (LLMs) demonstrate exceptional performance across various domains, deploying LLMs on edge devices has emerged as a new trend. Quantization techniques, which reduce the size and memory requirements of LLMs, are effective for deploying LLMs on resource-limited edge devices. However, existing one-size-fits-all quantization methods often fail to dynamically adjust the memory requirements of LLMs, limiting their applications to practical edge devices with various computation resources. To tackle this issue, we propose Layer-Specific Adaptive Quantization (LSAQ), a system for adaptive quantization and dynamic deployment of LLMs based on layer importance. Specifically, LSAQ evaluates the importance of LLMs' neural layers by constructing top-k token sets from the inputs and outputs of each layer and calculating their Jaccard similarity. Based on layer importance, our system adaptively adjusts quantization strategies in real time according to the computation resource of edge devices, which applies higher quantization precision to layers with higher importance, and vice versa. {Experimental results show that LSAQ consistently outperforms the selected quantization baselines in terms of perplexity and zero-shot tasks. Additionally, it can devise appropriate quantization schemes for different usage scenarios to facilitate the deployment of LLMs.
Quantization and pruning are two essential techniques for compressing neural networks, yet they are often treated independently, with limited theoretical analysis connecting them. This paper introduces a unified framework for post-training quantization and pruning using stochastic path-following algorithms. Our approach builds on the Stochastic Path Following Quantization (SPFQ) method, extending its applicability to pruning and low-bit quantization, including challenging 1-bit regimes. By incorporating a scaling parameter and generalizing the stochastic operator, the proposed method achieves robust error correction and yields rigorous theoretical error bounds for both quantization and pruning as well as their combination.
Graph neural networks (GNNs) have emerged as state-of-the-art methods to learn from graph-structured data for recommendation. However, most existing GNN-based recommendation methods focus on the optimization of model structures and learning strategies based on pre-defined graphs, neglecting the importance of the graph construction stage. Earlier works for graph construction usually rely on speciffic rules or crowdsourcing, which are either too simplistic or too labor-intensive. Recent works start to utilize large language models (LLMs) to automate the graph construction, in view of their abundant open-world knowledge and remarkable reasoning capabilities. Nevertheless, they generally suffer from two limitations: (1) invisibility of global view (e.g., overlooking contextual information) and (2) construction inefficiency. To this end, we introduce AutoGraph, an automatic graph construction framework based on LLMs for recommendation. Specifically, we first use LLMs to infer the user preference and item knowledge, which is encoded as semantic vectors. Next, we employ vector quantization to extract the latent factors from the semantic vectors. The latent factors are then incorporated as extra nodes to link the user/item nodes, resulting in a graph with in-depth global-view semantics. We further design metapath-based message aggregation to effectively aggregate the semantic and collaborative information. The framework is model-agnostic and compatible with different backbone models. Extensive experiments on three real-world datasets demonstrate the efficacy and efffciency of AutoGraph compared to existing baseline methods. We have deployed AutoGraph in Huawei advertising platform, and gain a 2.69% improvement on RPM and a 7.31% improvement on eCPM in the online A/B test. Currently AutoGraph has been used as the main trafffc model, serving hundreds of millions of people.
Detecting hallucinations in large language models (LLMs) is critical for enhancing their reliability and trustworthiness. Most research focuses on hallucinations as deviations from information seen during training. However, the opaque nature of an LLM's parametric knowledge complicates the understanding of why generated texts appear ungrounded: The LLM might not have picked up the necessary knowledge from large and often inaccessible datasets, or the information might have been changed or contradicted during further training. Our focus is on hallucinations involving information not used in training, which we determine by using recency to ensure the information emerged after a cut-off date. This study investigates these hallucinations by detecting them at sentence level using different internal states of various LLMs. We present HalluRAG, a dataset designed to train classifiers on these hallucinations. Depending on the model and quantization, MLPs trained on HalluRAG detect hallucinations with test accuracies ranging up to 75 %, with Mistral-7B-Instruct-v0.1 achieving the highest test accuracies. Our results show that IAVs detect hallucinations as effectively as CEVs and reveal that answerable and unanswerable prompts are encoded differently as separate classifiers for these categories improved accuracy. However, HalluRAG showed some limited generalizability, advocating for more diversity in datasets on hallucinations.
With the rapid growth in the scale and complexity of large language models (LLMs), the costs of training and inference have risen substantially. Model compression has emerged as a mainstream solution to reduce memory usage and computational overhead. This paper presents Group Quantization and Sparse Acceleration (\textbf{GQSA}), a novel compression technique tailored for LLMs. Traditional methods typically focus exclusively on either quantization or sparsification, but relying on a single strategy often results in significant performance loss at high compression rates. In contrast, GQSA integrates quantization and sparsification in a tightly coupled manner, leveraging GPU-friendly structured group sparsity and quantization for efficient acceleration. The proposed method consists of three key steps. First, GQSA applies group structured pruning to adhere to GPU-friendly sparse pattern constraints. Second, a two-stage sparsity-aware training process is employed to maximize performance retention after compression. Finally, the framework adopts the Block Sparse Row (BSR) format to enable practical deployment and efficient execution. Experimental results on the LLaMA model family show that GQSA achieves an excellent balance between model speed and accuracy. Furthermore, on the latest LLaMA-3 and LLaMA-3.1 models, GQSA outperforms existing LLM compression techniques significantly.
Diffusion models have achieved remarkable success in the image and video generation tasks. Nevertheless, they often require a large amount of memory and time overhead during inference, due to the complex network architecture and considerable number of timesteps for iterative diffusion. Recently, the post-training quantization (PTQ) technique has proved a promising way to reduce the inference cost by quantizing the float-point operations to low-bit ones. However, most of them fail to tackle with the large variations in the distribution of activations across distinct channels and timesteps, as well as the inconsistent of input between quantization and inference on diffusion models, thus leaving much room for improvement. To address the above issues, we propose a novel method dubbed Timestep-Channel Adaptive Quantization for Diffusion Models (TCAQ-DM). Specifically, we develop a timestep-channel joint reparameterization (TCR) module to balance the activation range along both the timesteps and channels, facilitating the successive reconstruction procedure. Subsequently, we employ a dynamically adaptive quantization (DAQ) module that mitigate the quantization error by selecting an optimal quantizer for each post-Softmax layers according to their specific types of distributions. Moreover, we present a progressively aligned reconstruction (PAR) strategy to mitigate the bias caused by the input mismatch. Extensive experiments on various benchmarks and distinct diffusion models demonstrate that the proposed method substantially outperforms the state-of-the-art approaches in most cases, especially yielding comparable FID metrics to the full precision model on CIFAR-10 in the W6A6 setting, while enabling generating available images in the W4A4 settings.
Data-free quantization (DFQ), which facilitates model quantization without real data to address increasing concerns about data security, has garnered significant attention within the model compression community. Recently, the unique architecture of vision transformers (ViTs) has driven the development of specialized DFQ techniques. However, we observe that the synthetic images from existing methods suffer from the deficient semantics issue compared to real images, thereby compromising performance. Motivated by this, we propose SPDFQ, a Semantics Prompting Data-Free Quantization method for ViTs. First, SPDFQ incorporates Attention Priors Alignment (APA), which uses randomly generated attention priors to enhance the semantics of synthetic images. Second, SPDFQ introduces Multi-Semantic Reinforcement (MSR), which utilizes localized patch optimization to prompt efficient parameterization and diverse semantics in synthetic images. Finally, SPDFQ employs Softlabel Learning (SL), where soft learning targets are adapted to encourage more complex semantics and accommodate images augmented by MSR. Experimental results demonstrate that SPDFQ significantly outperforms existing methods. For instance, SPDFQ achieves a 15.52% increase in top-1 accuracy on ImageNet for W4A4 ViT-B
In this work, we address unsupervised temporal action segmentation, which segments a set of long, untrimmed videos into semantically meaningful segments that are consistent across videos. While recent approaches combine representation learning and clustering in a single step for this task, they do not cope with large variations within temporal segments of the same class. To address this limitation, we propose a novel method, termed Hierarchical Vector Quantization (HVQ), that consists of two subsequent vector quantization modules. This results in a hierarchical clustering where the additional subclusters cover the variations within a cluster. We demonstrate that our approach captures the distribution of segment lengths much better than the state of the art. To this end, we introduce a new metric based on the Jensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We evaluate our approach on three public datasets, namely Breakfast, YouTube Instructional and IKEA ASM. Our approach outperforms the state of the art in terms of F1 score, recall and JSD.
Contemporary deep learning, characterized by the training of cumbersome neural networks on massive datasets, confronts substantial computational hurdles. To alleviate heavy data storage burdens on limited hardware resources, numerous dataset compression methods such as dataset distillation (DD) and coreset selection have emerged to obtain a compact but informative dataset through synthesis or selection for efficient training. However, DD involves an expensive optimization procedure and exhibits limited generalization across unseen architectures, while coreset selection is limited by its low data keep ratio and reliance on heuristics, hindering its practicality and feasibility. To address these limitations, we introduce a newly versatile framework for dataset compression, namely Adaptive Dataset Quantization (ADQ). Specifically, we first identify the sub-optimal performance of naive Dataset Quantization (DQ), which relies on uniform sampling and overlooks the varying importance of each generated bin. Subsequently, we propose a novel adaptive sampling strategy through the evaluation of generated bins' representativeness score, diversity score and importance score, where the former two scores are quantified by the texture level and contrastive learning-based techniques, respectively. Extensive experiments demonstrate that our method not only exhibits superior generalization capability across different architectures, but also attains state-of-the-art results, surpassing DQ by average 3\% on various datasets.
Clustering is a key task in machine learning, with $k$-means being widely used for its simplicity and effectiveness. While 1D clustering is common, existing methods often fail to exploit the structure of 1D data, leading to inefficiencies. This thesis introduces optimized algorithms for $k$-means++ initialization and Lloyd's algorithm, leveraging sorted data, prefix sums, and binary search for improved computational performance. The main contributions are: (1) an optimized $k$-cluster algorithm achieving $O(l \cdot k^2 \cdot \log n)$ complexity for greedy $k$-means++ initialization and $O(i \cdot k \cdot \log n)$ for Lloyd's algorithm, where $l$ is the number of greedy $k$-means++ local trials, and $i$ is the number of Lloyd's algorithm iterations, and (2) a binary search-based two-cluster algorithm, achieving $O(\log n)$ runtime with deterministic convergence to a Lloyd's algorithm local minimum. Benchmarks demonstrate over a 4500x speedup compared to scikit-learn for large datasets while maintaining clustering quality measured by within-cluster sum of squares (WCSS). Additionally, the algorithms achieve a 300x speedup in an LLM quantization task, highlighting their utility in emerging applications. This thesis bridges theory and practice for 1D $k$-means clustering, delivering efficient and sound algorithms implemented in a JIT-optimized open-source Python library.
Quantization-aware training (QAT) is a common paradigm for network quantization, in which the training phase incorporates the simulation of the low-precision computation to optimize the quantization parameters in alignment with the task goals. However, direct training of low-precision networks generally faces two obstacles: 1. The low-precision model exhibits limited representation capabilities and cannot directly replicate full-precision calculations, which constitutes a deficiency compared to full-precision alternatives; 2. Non-ideal deviations during gradient propagation are a common consequence of employing pseudo-gradients as approximations in derived quantized functions. In this paper, we propose a general QAT framework for alleviating the aforementioned concerns by permitting the forward and backward processes of the low-precision network to be guided by the full-precision partner during training. In conjunction with the direct training of the quantization model, intermediate mixed-precision models are generated through the block-by-block replacement on the full-precision model and working simultaneously with the low-precision backbone, which enables the integration of quantized low-precision blocks into full-precision networks throughout the training phase. Consequently, each quantized block is capable of: 1. simulating full-precision representation during forward passes; 2. obtaining gradients with improved estimation during backward passes. We demonstrate that the proposed method achieves state-of-the-art results for 4-, 3-, and 2-bit quantization on ImageNet and CIFAR-10. The proposed framework provides a compatible extension for most QAT methods and only requires a concise wrapper for existing codes.
Quantization has become one of the most effective methodologies to compress LLMs into smaller size. However, the existing quantization solutions still show limitations of either non-negligible accuracy drop or system inefficiency. In this paper, we make a comprehensive analysis of the general quantization principles on their effect to the triangle of accuracy, memory consumption and system efficiency. We propose MixLLM that explores the new optimization space of mixed-precision quantization between output features based on the insight that different output features matter differently in the model. MixLLM identifies the output features with high salience in the global view rather than within each single layer, effectively assigning the larger bit-width to output features that need it most to achieve good accuracy with low memory consumption. We present the sweet spot of quantization configuration of algorithm-system co-design that leads to high accuracy and system efficiency. To address the system challenge, we design the two-step dequantization to make use of the int8 Tensor Core easily and fast data type conversion to reduce dequantization overhead significantly, and present the software pipeline to overlap the memory access, dequantization and the MatMul to the best. Extensive experiments show that with only 10% more bits, the PPL increasement can be reduced from about 0.5 in SOTA to within 0.2 for Llama 3.1 70B, while on average MMLU-Pro improves by 0.93 over the SOTA of three popular models. In addition to its superior accuracy, MixLLM also achieves state-of-the-art system efficiency.
Post-training quantization (PTQ) of large language models (LLMs) holds the promise in reducing the prohibitive computational cost at inference time. Quantization of all weight, activation and key-value (KV) cache tensors to 4-bit without significantly degrading generalizability is challenging, due to the high quantization error caused by extreme outliers in activations. To tackle this problem, we propose ResQ, a PTQ method that pushes further the state-of-the-art. By means of principal component analysis (PCA), it identifies a low-rank subspace (in practice 1/8 of the hidden dimension) in which activation variances are highest, and keep the coefficients within this subspace in high precision, e.g. 8-bit, while quantizing the rest to 4-bit. Within each subspace, invariant random rotation is applied to further suppress outliers. We show that this is a provably optimal mixed precision quantization scheme that minimizes error. With the Llama families of models, we demonstrate that ResQ outperforms recent uniform and mixed precision PTQ methods on a variety of benchmarks, achieving up to 33% lower perplexity on Wikitext than the next best method SpinQuant, and a 2.4x speedup over 16-bit baseline. Code is available at https://github.com/utkarsh-dmx/project-resq.
Diffusion Models (DM) have democratized AI image generation through an iterative denoising process. Quantization is a major technique to alleviate the inference cost and reduce the size of DM denoiser networks. However, as denoisers evolve from variants of convolutional U-Nets toward newer Transformer architectures, it is of growing importance to understand the quantization sensitivity of different weight layers, operations and architecture types to performance. In this work, we address this challenge with Qua$^2$SeDiMo, a mixed-precision Post-Training Quantization framework that generates explainable insights on the cost-effectiveness of various model weight quantization methods for different denoiser operation types and block structures. We leverage these insights to make high-quality mixed-precision quantization decisions for a myriad of diffusion models ranging from foundational U-Nets to state-of-the-art Transformers. As a result, Qua$^2$SeDiMo can construct 3.4-bit, 3.9-bit, 3.65-bit and 3.7-bit weight quantization on PixArt-${\alpha}$, PixArt-${\Sigma}$, Hunyuan-DiT and SDXL, respectively. We further pair our weight-quantization configurations with 6-bit activation quantization and outperform existing approaches in terms of quantitative metrics and generative image quality.