Large Language Models (LLMs) excel in diverse applications but suffer inefficiency due to massive scale. While quantization reduces computational costs, existing methods degrade accuracy in medium-sized LLMs (e.g., Llama-3-8B) due to activation outliers. To address this, we propose QUAD (Quantization with Activation Decomposition), a framework leveraging Singular Value Decomposition (SVD) to suppress activation outliers for effective 4-bit quantization. QUAD estimates activation singular vectors offline using calibration data to construct an orthogonal transformation matrix P, shifting outliers to additional dimensions in full precision while quantizing rest components to 4-bit. Additionally, QUAD enables parameter-efficient fine-tuning via adaptable full-precision outlier weights, narrowing the accuracy gap between quantized and full-precision models. Experiments demonstrate that QUAD achieves 94% ~ 96% accuracy under W4A4 quantization and 98% accuracy with W4A4/A8 and parameter-efficient fine-tuning for Llama-3 and Qwen-2.5 models. Our code is available at \href{https://github.com/hyx1999/Quad}{repository}.
Video anomaly detection (VAD) methods are mostly CNN-based or Transformer-based, achieving impressive results, but the focus on detection accuracy often comes at the expense of inference speed. The emergence of state space models in computer vision, exemplified by the Mamba model, demonstrates improved computational efficiency through selective scans and showcases the great potential for long-range modeling. Our study pioneers the application of Mamba to VAD, dubbed VADMamba, which is based on multi-task learning for frame prediction and optical flow reconstruction. Specifically, we propose the VQ-Mamba Unet (VQ-MaU) framework, which incorporates a Vector Quantization (VQ) layer and Mamba-based Non-negative Visual State Space (NVSS) block. Furthermore, two individual VQ-MaU networks separately predict frames and reconstruct corresponding optical flows, further boosting accuracy through a clip-level fusion evaluation strategy. Experimental results validate the efficacy of the proposed VADMamba across three benchmark datasets, demonstrating superior performance in inference speed compared to previous work. Code is available at https://github.com/jLooo/VADMamba.
We introduce LogQuant, a groundbreaking 2-bit quantization technique for KV Cache in large language model (LLM) inference, delivering substantial memory savings while preserving superior performance. Previous methods either assume that later tokens are more important or attempt to predict important tokens based on earlier attention patterns. Both approaches, however, can result in performance bottlenecks or frequent mispredictions. LogQuant takes a different approach. By applying a log-based filtering mechanism, it selectively compresses the KV Cache across the entire context, achieving better performance with the same or even reduced memory footprint compared to existing methods. In benchmark tests, it enhances throughput by 25% and boosts batch size by 60% without increasing memory consumption. For challenging tasks such as Math and Code Completion, LogQuant improves accuracy by 40% to 200% at the same compression ratio, outperforming comparable techniques.LogQuant integrates effortlessly with popular inference frameworks like Python's transformers library. Implementation can be available in https://github.com/Concyclics/LogQuantKV.
The growing adoption of long-context Large Language Models (LLMs) has introduced significant memory and computational challenges in autoregressive decoding due to the expanding Key-Value (KV) cache. KV cache quantization has emerged as a promising solution, with prior work showing that 4-bit or even 2-bit quantization can maintain model accuracy while reducing memory costs. However, despite these benefits, preliminary implementations for the low-bit KV cache struggle to deliver the expected speedup due to quantization and dequantization overheads and the lack of Tensor Cores utilization. In this work, we propose BitDecoding, a GPU-optimized framework that unlocks Tensor Cores for efficient decoding with low-bit KV cache. Efficiently leveraging Tensor Cores for low-bit KV cache is challenging due to the dynamic nature of KV cache generation at each decoding step. BitDecoding addresses these challenges with a Tensor Cores-Centric BitFusion Scheme that ensures data layout compatibility to enable high utilization of Tensor Cores. Additionally, BitDecoding incorporates a warp-efficient parallel decoding kernel and a fine-grained asynchronous pipeline, minimizing dequantization overhead and improving computational efficiency. Experiments show that BitDecoding achieves up to 7.5x speedup on RTX 4090, 4.8x on A100, and 8.9x on H100, compared to FP16 FlashDecoding-v2. It also outperforms the state-of-the-art low-bit KV cache implementation (QServe) by up to 4.3x. On LLaMA-3.1-8B with a 128K sequence length, BitDecoding reduces single-batch decoding latency by 3x, demonstrating its effectiveness in long-context generation scenarios. The code is available at https://github.com/DD-DuDa/BitDecoding.
The outcome of Large Language Model (LLM) pre-training strongly depends on weight initialization and variance control strategies. Although the importance of initial variance control has been well documented in neural networks in general, the literature on initialization and management of its growth during LLM pre-training, specifically, is somewhat sparse. In this paper, we introduce the Layer Index Rescaling (LIR) weight initialization scheme, and the Target Variance Rescaling (TVR) variance control strategy. Experiments on a 1B parameter LLaMA model demonstrate that better variance management using these techniques yields substantial improvements in downstream task performance (up to 4.6% on common pre-training benchmarks) and reduces extreme activation values, thus mitigating challenges associated with quantization and low-precision training. Our code is available at: https://github.com/bluorion-com/weight_rescaling.
Diffusion Models (DM) have revolutionized the text-to-image visual generation process. However, the large computational cost and model footprint of DMs hinders practical deployment, especially on edge devices. Post-training quantization (PTQ) is a lightweight method to alleviate these burdens without the need for training or fine-tuning. While recent DM PTQ methods achieve W4A8 on integer-based PTQ, two key limitations remain: First, while most existing DM PTQ methods evaluate on classical DMs like Stable Diffusion XL, 1.5 or earlier, which use convolutional U-Nets, newer Diffusion Transformer (DiT) models like the PixArt series, Hunyuan and others adopt fundamentally different transformer backbones to achieve superior image synthesis. Second, integer (INT) quantization is prevailing in DM PTQ but doesn't align well with the network weight and activation distribution, while Floating-Point Quantization (FPQ) is still under-investigated, yet it holds the potential to better align the weight and activation distributions in low-bit settings for DiT. In response, we introduce FP4DiT, a PTQ method that leverages FPQ to achieve W4A6 quantization. Specifically, we extend and generalize the Adaptive Rounding PTQ technique to adequately calibrate weight quantization for FPQ and demonstrate that DiT activations depend on input patch data, necessitating robust online activation quantization techniques. Experimental results demonstrate that FP4DiT outperforms integer-based PTQ at W4A6 and W4A8 precision and generates convincing visual content on PixArt-$\alpha$, PixArt-$\Sigma$ and Hunyuan in terms of several T2I metrics such as HPSv2 and CLIP.
We quantify the upper bound on the size of the implicit neural representation (INR) model from a digital perspective. The upper bound of the model size increases exponentially as the required bit-precision increases. To this end, we present a bit-plane decomposition method that makes INR predict bit-planes, producing the same effect as reducing the upper bound of the model size. We validate our hypothesis that reducing the upper bound leads to faster convergence with constant model size. Our method achieves lossless representation in 2D image and audio fitting, even for high bit-depth signals, such as 16-bit, which was previously unachievable. We pioneered the presence of bit bias, which INR prioritizes as the most significant bit (MSB). We expand the application of the INR task to bit depth expansion, lossless image compression, and extreme network quantization. Our source code is available at https://github.com/WooKyoungHan/LosslessINR
Current transformer-based imitation learning approaches introduce discrete action representations and train an autoregressive transformer decoder on the resulting latent code. However, the initial quantization breaks the continuous structure of the action space thereby limiting the capabilities of the generative model. We propose a quantization-free method instead that leverages Generative Infinite-Vocabulary Transformers (GIVT) as a direct, continuous policy parametrization for autoregressive transformers. This simplifies the imitation learning pipeline while achieving state-of-the-art performance on a variety of popular simulated robotics tasks. We enhance our policy roll-outs by carefully studying sampling algorithms, further improving the results.
Automated speech recognition (ASR) models have gained prominence for applications such as captioning, speech translation, and live transcription. This paper studies Whisper and two model variants: one optimized for live speech streaming and another for offline transcription. Notably, these models have been found to generate hallucinated content, reducing transcription reliability. Furthermore, larger model variants exhibit increased latency and pose challenges for deployment on resource-constrained devices. This study analyzes the similarities and differences between three Whisper models, qualitatively examining their distinct capabilities. Next, this study quantifies the impact of model quantization on latency and evaluates its viability for edge deployment. Using the open source LibriSpeech dataset, this paper evaluates the word error rate (WER) along with latency analysis of whispercpp using 3 quantization methods (INT4, INT5, INT8). Results show that quantization reduces latency by 19\% and model size by 45\%, while preserving transcription accuracy. These findings provide insights into the optimal use cases of different Whisper models and edge device deployment possibilities. All code, datasets, and implementation details are available in a public GitHub repository: https://github.com/allisonandreyev/WhisperQuantization.git
3D Gaussian Splatting (3DGS) achieves impressive rendering fidelity and speed for novel view synthesis. However, its substantial data size poses a significant challenge for practical applications. While many compression techniques have been proposed, they fail to efficiently utilize existing bitstreams in on-demand applications due to their lack of progressivity, leading to a waste of resource. To address this issue, we propose PCGS (Progressive Compression of 3D Gaussian Splatting), which adaptively controls both the quantity and quality of Gaussians (or anchors) to enable effective progressivity for on-demand applications. Specifically, for quantity, we introduce a progressive masking strategy that incrementally incorporates new anchors while refining existing ones to enhance fidelity. For quality, we propose a progressive quantization approach that gradually reduces quantization step sizes to achieve finer modeling of Gaussian attributes. Furthermore, to compact the incremental bitstreams, we leverage existing quantization results to refine probability prediction, improving entropy coding efficiency across progressive levels. Overall, PCGS achieves progressivity while maintaining compression performance comparable to SoTA non-progressive methods. Code available at: github.com/YihangChen-ee/PCGS.
Despite recent advancements in federated learning (FL), the integration of generative models into FL has been limited due to challenges such as high communication costs and unstable training in heterogeneous data environments. To address these issues, we propose PRISM, a FL framework tailored for generative models that ensures (i) stable performance in heterogeneous data distributions and (ii) resource efficiency in terms of communication cost and final model size. The key of our method is to search for an optimal stochastic binary mask for a random network rather than updating the model weights, identifying a sparse subnetwork with high generative performance; i.e., a ``strong lottery ticket''. By communicating binary masks in a stochastic manner, PRISM minimizes communication overhead. This approach, combined with the utilization of maximum mean discrepancy (MMD) loss and a mask-aware dynamic moving average aggregation method (MADA) on the server side, facilitates stable and strong generative capabilities by mitigating local divergence in FL scenarios. Moreover, thanks to its sparsifying characteristic, PRISM yields a lightweight model without extra pruning or quantization, making it ideal for environments such as edge devices. Experiments on MNIST, FMNIST, CelebA, and CIFAR10 demonstrate that PRISM outperforms existing methods, while maintaining privacy with minimal communication costs. PRISM is the first to successfully generate images under challenging non-IID and privacy-preserving FL environments on complex datasets, where previous methods have struggled.
Implicit Neural Representations (INRs) have emerged as a powerful approach for video representation, offering versatility across tasks such as compression and inpainting. However, their implicit formulation limits both interpretability and efficacy, undermining their practicality as a comprehensive solution. We propose a novel video representation based on deformable 2D Gaussian splatting, dubbed D2GV, which aims to achieve three key objectives: 1) improved efficiency while delivering superior quality; 2) enhanced scalability and interpretability; and 3) increased friendliness for downstream tasks. Specifically, we initially divide the video sequence into fixed-length Groups of Pictures (GoP) to allow parallel training and linear scalability with video length. For each GoP, D2GV represents video frames by applying differentiable rasterization to 2D Gaussians, which are deformed from a canonical space into their corresponding timestamps. Notably, leveraging efficient CUDA-based rasterization, D2GV converges fast and decodes at speeds exceeding 400 FPS, while delivering quality that matches or surpasses state-of-the-art INRs. Moreover, we incorporate a learnable pruning and quantization strategy to streamline D2GV into a more compact representation. We demonstrate D2GV's versatility in tasks including video interpolation, inpainting and denoising, underscoring its potential as a promising solution for video representation. Code is available at: https://github.com/Evan-sudo/D2GV.
One-step diffusion-based image super-resolution (OSDSR) models are showing increasingly superior performance nowadays. However, although their denoising steps are reduced to one and they can be quantized to 8-bit to reduce the costs further, there is still significant potential for OSDSR to quantize to lower bits. To explore more possibilities of quantized OSDSR, we propose an efficient method, Quantization via reverse-module and timestep-retraining for OSDSR, named QArtSR. Firstly, we investigate the influence of timestep value on the performance of quantized models. Then, we propose Timestep Retraining Quantization (TRQ) and Reversed Per-module Quantization (RPQ) strategies to calibrate the quantized model. Meanwhile, we adopt the module and image losses to update all quantized modules. We only update the parameters in quantization finetuning components, excluding the original weights. To ensure that all modules are fully finetuned, we add extended end-to-end training after per-module stage. Our 4-bit and 2-bit quantization experimental results indicate that QArtSR obtains superior effects against the recent leading comparison methods. The performance of 4-bit QArtSR is close to the full-precision one. Our code will be released at https://github.com/libozhu03/QArtSR.
Recently, Diffusion Transformers (DiTs) have emerged as a dominant architecture in video generation, surpassing U-Net-based models in terms of performance. However, the enhanced capabilities of DiTs come with significant drawbacks, including increased computational and memory costs, which hinder their deployment on resource-constrained devices. Current acceleration techniques, such as quantization and cache mechanism, offer limited speedup and are often applied in isolation, failing to fully address the complexities of DiT architectures. In this paper, we propose QuantCache, a novel training-free inference acceleration framework that jointly optimizes hierarchical latent caching, adaptive importance-guided quantization, and structural redundancy-aware pruning. QuantCache achieves an end-to-end latency speedup of 6.72$\times$ on Open-Sora with minimal loss in generation quality. Extensive experiments across multiple video generation benchmarks demonstrate the effectiveness of our method, setting a new standard for efficient DiT inference. The code and models will be available at https://github.com/JunyiWuCode/QuantCache.
In this work, we propose an extreme compression technique for Large Multimodal Models (LMMs). While previous studies have explored quantization as an efficient post-training compression method for Large Language Models (LLMs), low-bit compression for multimodal models remains under-explored. The redundant nature of inputs in multimodal models results in a highly sparse attention matrix. We theoretically and experimentally demonstrate that the attention matrix's sparsity bounds the compression error of the Query and Key weight matrices. Based on this, we introduce CASP, a model compression technique for LMMs. Our approach performs a data-aware low-rank decomposition on the Query and Key weight matrix, followed by quantization across all layers based on an optimal bit allocation process. CASP is compatible with any quantization technique and enhances state-of-the-art 2-bit quantization methods (AQLM and QuIP#) by an average of 21% on image- and video-language benchmarks.
Full 3D human pose reconstruction is a critical enabler for extended reality (XR) applications in future sixth generation (6G) networks, supporting immersive interactions in gaming, virtual meetings, and remote collaboration. However, achieving accurate pose reconstruction over wireless networks remains challenging due to channel impairments, bit errors, and quantization effects. Existing approaches often assume error-free transmission in indoor settings, limiting their applicability to real-world scenarios. To address these challenges, we propose a novel deep learning-based framework for human pose reconstruction over orthogonal frequency-division multiplexing (OFDM) systems. The framework introduces a two-stage deep learning receiver: the first stage jointly estimates the wireless channel and decodes OFDM symbols, and the second stage maps the received sensor signals to full 3D body poses. Simulation results demonstrate that the proposed neural receiver reduces bit error rate (BER), thus gaining a 5 dB gap at $10^{-4}$ BER, compared to the baseline method that employs separate signal detection steps, i.e., least squares channel estimation and linear minimum mean square error equalization. Additionally, our empirical findings show that 8-bit quantization is sufficient for accurate pose reconstruction, achieving a mean squared error of $5\times10^{-4}$ for reconstructed sensor signals, and reducing joint angular error by 37\% for the reconstructed human poses compared to the baseline.
Model binarization has made significant progress in enabling real-time and energy-efficient computation for convolutional neural networks (CNN), offering a potential solution to the deployment challenges faced by Vision Transformers (ViTs) on edge devices. However, due to the structural differences between CNN and Transformer architectures, simply applying binary CNN strategies to the ViT models will lead to a significant performance drop. To tackle this challenge, we propose BHViT, a binarization-friendly hybrid ViT architecture and its full binarization model with the guidance of three important observations. Initially, BHViT utilizes the local information interaction and hierarchical feature aggregation technique from coarse to fine levels to address redundant computations stemming from excessive tokens. Then, a novel module based on shift operations is proposed to enhance the performance of the binary Multilayer Perceptron (MLP) module without significantly increasing computational overhead. In addition, an innovative attention matrix binarization method based on quantization decomposition is proposed to evaluate the token's importance in the binarized attention matrix. Finally, we propose a regularization loss to address the inadequate optimization caused by the incompatibility between the weight oscillation in the binary layers and the Adam Optimizer. Extensive experimental results demonstrate that our proposed algorithm achieves SOTA performance among binary ViT methods.
We introduce a novel class of regularization functions, called Cauchy-Schwarz (CS) regularizers, which can be designed to induce a wide range of properties in solution vectors of optimization problems. To demonstrate the versatility of CS regularizers, we derive regularization functions that promote discrete-valued vectors, eigenvectors of a given matrix, and orthogonal matrices. The resulting CS regularizers are simple, differentiable, and can be free of spurious stationary points, making them suitable for gradient-based solvers and large-scale optimization problems. In addition, CS regularizers automatically adapt to the appropriate scale, which is, for example, beneficial when discretizing the weights of neural networks. To demonstrate the efficacy of CS regularizers, we provide results for solving underdetermined systems of linear equations and weight quantization in neural networks. Furthermore, we discuss specializations, variations, and generalizations, which lead to an even broader class of new and possibly more powerful regularizers.
Reducing radiation doses benefits patients, however, the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) shows promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to address these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures being scanned, we design a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM), respectively. Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying the unseen scanning code as one of the codes in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on public datasets.
Layer-wise quantization is a key technique for efficiently compressing large models without expensive retraining. Previous methods typically quantize the weights of each layer by "uniformly" optimizing the layer reconstruction loss across all output tokens. However, in this paper, we demonstrate that better-quantized models can be obtained by prioritizing learning from important tokens (e.g. which have large attention scores). Building on this finding, we propose RSQ (Rotate, Scale, then Quantize), which (1) applies rotations (orthogonal transformation) to the model to mitigate outliers (those with exceptionally large magnitude), (2) scales the token feature based on its importance, and (3) quantizes the model using the GPTQ framework with the second-order statistics computed by scaled tokens. To compute token importance, we explore both heuristic and dynamic strategies. Based on a thorough analysis of all approaches, we adopt attention concentration, which uses attention scores of each token as its importance, as the best approach. We demonstrate that RSQ consistently outperforms baseline methods across multiple downstream tasks and three model families: LLaMA3, Mistral, and Qwen2.5. Additionally, models quantized with RSQ achieve superior performance on long-context tasks, further highlighting its effectiveness. Lastly, RSQ demonstrates generalizability across various setups, including different model sizes, calibration datasets, bit precisions, and quantization methods.