Zero-shot quantization (ZSQ) enables neural network compression without original training data, making it a promising solution for restricted data access scenarios. To compensate for the lack of data, recent ZSQ methods typically rely on synthetic inputs generated from the full-precision model. However, these synthetic inputs often lead to activation distortion, especially under low-bit settings. As a result, existing methods struggle to mitigate this issue due to coarse activation scaling. To address this issue, we propose GranQ, a novel activation quantization framework that efficiently applies per-channel scaling through vectorized computation. In contrast to conventional channel-wise methods, which apply vectorization only to the quantization step, GranQ improves efficiency by vectorizing the scaling operation. This design allows GranQ to maintain fine-grained quantization granularity with minimal computational overhead, even in low-bit environments. Extensive experiments under quantization-aware training (QAT) settings demonstrate that GranQ consistently outperforms state-of-the-art ZSQ methods across CIFAR and ImageNet. In particular, our method achieves up to 5.45% higher accuracy in the 3-bit setting on CIFAR-100 and even surpasses the full-precision baseline on CIFAR-10. Furthermore, GranQ achieves significant speedup in quantization latency over conventional per-channel methods, demonstrating improved efficiency. With these findings, we anticipate that GranQ will inspire future research beyond conventional ZSQ approaches centered on data generation and model fine-tuning.
3D Gaussian Splatting (3DGS) has substantial potential for enabling photorealistic Free-Viewpoint Video (FVV) experiences. However, the vast number of Gaussians and their associated attributes poses significant challenges for storage and transmission. Existing methods typically handle dynamic 3DGS representation and compression separately, neglecting motion information and the rate-distortion (RD) trade-off during training, leading to performance degradation and increased model redundancy. To address this gap, we propose 4DGC, a novel rate-aware 4D Gaussian compression framework that significantly reduces storage size while maintaining superior RD performance for FVV. Specifically, 4DGC introduces a motion-aware dynamic Gaussian representation that utilizes a compact motion grid combined with sparse compensated Gaussians to exploit inter-frame similarities. This representation effectively handles large motions, preserving quality and reducing temporal redundancy. Furthermore, we present an end-to-end compression scheme that employs differentiable quantization and a tiny implicit entropy model to compress the motion grid and compensated Gaussians efficiently. The entire framework is jointly optimized using a rate-distortion trade-off. Extensive experiments demonstrate that 4DGC supports variable bitrates and consistently outperforms existing methods in RD performance across multiple datasets.
Advanced Large Language Models (LLMs) have revolutionized various fields, including communication networks, sparking an innovation wave that has led to new applications and services, and significantly enhanced solution schemes. Despite all these impressive developments, most LLMs typically require huge computational resources, resulting in terribly high energy consumption. Thus, this research study proposes an end-to-end pipeline that investigates the trade-off between energy efficiency and model performance for an LLM during fault ticket analysis in communication networks. It further evaluates the pipeline performance using two real-world datasets for the tasks of root cause analysis and response feedback in a communication network. Our results show that an appropriate combination of quantization and pruning techniques is able to reduce energy consumption while significantly improving model performance.
With an increasing demand for assistive technologies that promote the independence and mobility of visually impaired people, this study suggests an innovative real-time system that gives audio descriptions of a user's surroundings to improve situational awareness. The system acquires live video input and processes it with a quantized and fine-tuned Florence-2 big model, adjusted to 4-bit accuracy for efficient operation on low-power edge devices such as the NVIDIA Jetson Orin Nano. By transforming the video signal into frames with a 5-frame latency, the model provides rapid and contextually pertinent descriptions of objects, pedestrians, and barriers, together with their estimated distances. The system employs Parler TTS Mini, a lightweight and adaptable Text-to-Speech (TTS) solution, for efficient audio feedback. It accommodates 34 distinct speaker types and enables customization of speech tone, pace, and style to suit user requirements. This study examines the quantization and fine-tuning techniques utilized to modify the Florence-2 model for this application, illustrating how the integration of a compact model architecture with a versatile TTS component improves real-time performance and user experience. The proposed system is assessed based on its accuracy, efficiency, and usefulness, providing a viable option to aid vision-impaired users in navigating their surroundings securely and successfully.
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.
Modern Large Language Model serving system batches multiple requests to achieve high throughput, while batching attention operations is challenging, rendering memory bandwidth a critical bottleneck. The community relies on high-end GPUs with multiple high-bandwidth memory channels. Unfortunately, HBM's high bandwidth often comes at the expense of limited memory capacity, which reduces core utilization and increases costs. Recent advancements enabling longer contexts for LLMs have substantially increased the key-value cache size, further intensifying the pressures on memory capacity. The literature has explored KV cache quantization techniques, which commonly use low bitwidth for most values, selectively using higher bitwidth for outlier values. While this approach helps achieve high accuracy and low bitwidth simultaneously, it comes with the limitation that cost for online outlier detection is excessively high, negating the advantages. We propose Oaken, an acceleration solution that achieves high accuracy and high performance simultaneously through co-designing algorithm and hardware. To effectively find a sweet spot in the accuracy-performance trade-off space of KV cache quantization, Oaken employs an online-offline hybrid approach, setting outlier thresholds offline, which are then used to determine the quantization scale online. To translate the proposed algorithmic technique into tangible performance gains, Oaken also comes with custom quantization engines and memory management units that can be integrated with any LLM accelerators. We built an Oaken accelerator on top of an LLM accelerator, LPU, and conducted a comprehensive evaluation. Our experiments show that for a batch size of 256, Oaken achieves up to 1.58x throughput improvement over NVIDIA A100 GPU, incurring a minimal accuracy loss of only 0.54\% on average, compared to state-of-the-art KV cache quantization techniques.
The size of a model has been a strong predictor of its quality, as well as its cost. As such, the trade-off between model cost and quality has been well-studied. Post-training optimizations like quantization and pruning have typically focused on reducing the overall volume of pre-trained models to reduce inference costs while maintaining model quality. However, recent advancements have introduced optimization techniques that, interestingly, expand models post-training, increasing model size to improve quality when reducing volume. For instance, to enable 4-bit weight and activation quantization, incoherence processing often necessitates inserting online Hadamard rotations in the compute graph, and preserving highly sensitive weights often calls for additional higher precision computations. However, if application requirements cannot be met, the prevailing solution is to relax quantization constraints. In contrast, we demonstrate post-training model expansion is a viable strategy to improve model quality within a quantization co-design space, and provide theoretical justification. We show it is possible to progressively and selectively expand the size of a pre-trained large language model (LLM) to improve model quality without end-to-end retraining. In particular, when quantizing the weights and activations to 4 bits for Llama3 1B, we reduce the zero-shot accuracy gap to full precision by an average of 3% relative to both QuaRot and SpinQuant with only 5% more parameters, which is still a 3.8% reduction in volume relative to a BF16 reference model.
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.
As large language models (LLMs) scale, model compression is crucial for edge deployment and accessibility. Weight-only quantization reduces model size but suffers from performance degradation at lower bit widths. Moreover, standard finetuning is incompatible with quantized models, and alternative methods often fall short of full finetuning. In this paper, we propose ClusComp, a simple yet effective compression paradigm that clusters weight matrices into codebooks and finetunes them block-by-block. ClusComp (1) achieves superior performance in 2-4 bit quantization, (2) pushes compression to 1-bit while outperforming ultra-low-bit methods with minimal finetuning, and (3) enables efficient finetuning, even surpassing existing quantization-based approaches and rivaling full FP16 finetuning. Notably, ClusComp supports compression and finetuning of 70B LLMs on a single A6000-48GB GPU.
Autoregressive visual generation models typically rely on tokenizers to compress images into tokens that can be predicted sequentially. A fundamental dilemma exists in token representation: discrete tokens enable straightforward modeling with standard cross-entropy loss, but suffer from information loss and tokenizer training instability; continuous tokens better preserve visual details, but require complex distribution modeling, complicating the generation pipeline. In this paper, we propose TokenBridge, which bridges this gap by maintaining the strong representation capacity of continuous tokens while preserving the modeling simplicity of discrete tokens. To achieve this, we decouple discretization from the tokenizer training process through post-training quantization that directly obtains discrete tokens from continuous representations. Specifically, we introduce a dimension-wise quantization strategy that independently discretizes each feature dimension, paired with a lightweight autoregressive prediction mechanism that efficiently model the resulting large token space. Extensive experiments show that our approach achieves reconstruction and generation quality on par with continuous methods while using standard categorical prediction. This work demonstrates that bridging discrete and continuous paradigms can effectively harness the strengths of both approaches, providing a promising direction for high-quality visual generation with simple autoregressive modeling. Project page: https://yuqingwang1029.github.io/TokenBridge.
Video large language models (VideoLLMs) have demonstrated the capability to process longer video inputs and enable complex reasoning and analysis. However, due to the thousands of visual tokens from the video frames, key-value (KV) cache can significantly increase memory requirements, becoming a bottleneck for inference speed and memory usage. KV cache quantization is a widely used approach to address this problem. In this paper, we find that 2-bit KV quantization of VideoLLMs can hardly hurt the model performance, while the limit of KV cache quantization in even lower bits has not been investigated. To bridge this gap, we introduce VidKV, a plug-and-play KV cache quantization method to compress the KV cache to lower than 2 bits. Specifically, (1) for key, we propose a mixed-precision quantization strategy in the channel dimension, where we perform 2-bit quantization for anomalous channels and 1-bit quantization combined with FFT for normal channels; (2) for value, we implement 1.58-bit quantization while selectively filtering semantically salient visual tokens for targeted preservation, for a better trade-off between precision and model performance. Importantly, our findings suggest that the value cache of VideoLLMs should be quantized in a per-channel fashion instead of the per-token fashion proposed by prior KV cache quantization works for LLMs. Empirically, extensive results with LLaVA-OV-7B and Qwen2.5-VL-7B on six benchmarks show that VidKV effectively compresses the KV cache to 1.5-bit and 1.58-bit precision with almost no performance drop compared to the FP16 counterparts.
Autoregressive models have shown remarkable success in image generation by adapting sequential prediction techniques from language modeling. However, applying these approaches to images requires discretizing continuous pixel data through vector quantization methods like VQ-VAE. To alleviate the quantization errors that existed in VQ-VAE, recent works tend to use larger codebooks. However, this will accordingly expand vocabulary size, complicating the autoregressive modeling task. This paper aims to find a way to enjoy the benefits of large codebooks without making autoregressive modeling more difficult. Through empirical investigation, we discover that tokens with similar codeword representations produce similar effects on the final generated image, revealing significant redundancy in large codebooks. Based on this insight, we propose to predict tokens from coarse to fine (CTF), realized by assigning the same coarse label for similar tokens. Our framework consists of two stages: (1) an autoregressive model that sequentially predicts coarse labels for each token in the sequence, and (2) an auxiliary model that simultaneously predicts fine-grained labels for all tokens conditioned on their coarse labels. Experiments on ImageNet demonstrate our method's superior performance, achieving an average improvement of 59 points in Inception Score compared to baselines. Notably, despite adding an inference step, our approach achieves faster sampling speeds.
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
In this paper, we explore the algebra of quantum idempotents and the quantization of fermions which gives rise to a Hilbert space equal to the Grassmann algebra associated with the Lie algebra. Since idempotents carry representations of the algebra under consideration, they form algebraic varieties and smooth manifolds in the natural topology. In addition to the motivation of linking up mathematical physics with machine learning, it is also shown that by using idempotents and invariant subspace of the corresponding algebras, these representations encode and perhaps provide a probabilistic interpretation of reasoning and relational paths in geometrical terms.
Transformer-based large language models (LLMs) have already achieved remarkable results on long-text tasks, but the limited GPU memory (VRAM) resources struggle to accommodate the linearly growing demand for key-value (KV) cache as the sequence length increases, which has become a bottleneck for the application of LLMs on long sequences. Existing KV cache compression methods include eviction, merging, or quantization of the KV cache to reduce its size. However, compression results in irreversible information forgetting, potentially affecting the accuracy of subsequent decoding. In this paper, we propose SpeCache, which takes full advantage of the large and easily expandable CPU memory to offload the complete KV cache, and dynamically fetches KV pairs back in each decoding step based on their importance measured by low-bit KV cache copy in VRAM. To avoid inference latency caused by CPU-GPU communication, SpeCache speculatively predicts the KV pairs that the next token might attend to, allowing us to prefetch them before the next decoding step which enables parallelization of prefetching and computation. Experiments on LongBench and Needle-in-a-Haystack benchmarks verify that SpeCache effectively reduces VRAM usage while avoiding information forgetting for long sequences without re-training, even with a 10x high KV cache compression ratio.
While there are many advantages to deploying machine learning models on edge devices, the resource constraints of mobile platforms, the dynamic nature of the environment, and differences between the distribution of training versus in-the-wild data make such deployments challenging. Current test-time adaptation methods are often memory-intensive and not designed to be quantization-compatible or deployed on low-resource devices. To address these challenges, we present LeanTTA, a novel backpropagation-free and stateless framework for quantized test-time adaptation tailored to edge devices. Our approach minimizes computational costs by dynamically updating normalization statistics without backpropagation, which frees LeanTTA from the common pitfall of relying on large batches and historical data, making our method robust to realistic deployment scenarios. Our approach is the first to enable further computational gains by combining partial adaptation with quantized module fusion. We validate our framework across sensor modalities, demonstrating significant improvements over state-of-the-art TTA methods, including a 15.7% error reduction, peak memory usage of only 11.2MB for ResNet18, and fast adaptation within an order-of-magnitude of normal inference speeds on-device. LeanTTA provides a robust solution for achieving the right trade offs between accuracy and system efficiency in edge deployments, addressing the unique challenges posed by limited data and varied operational conditions.
This study investigates the problem of learning linear block codes optimized for Belief-Propagation decoders significantly improving performance compared to the state-of-the-art. Our previous research is extended with an enhanced system design that facilitates a more effective learning process for the parity check matrix. We simplify the input dataset, restrict the number of parameters to learn and improve the gradient back-propagation within the model. We also introduce novel optimizers specifically designed for discrete-valued weights. Based on conventional gradient computation, these optimizers provide discrete weights updates, enabling finer control and improving explainability of the learning process. Through these changes, we consistently achieve improved code performance, provided appropriately chosen hyper-parameters. To rigorously evaluate the performance of learned codes in the context of short to medium block lengths, we propose a comprehensive code performance assessment framework. This framework enables a fair comparison between our learning methodology and random search approaches, ensuring statistical significance in our results. The proposed model pave the way for a new approach to the efficient learning of linear block codes tailored to specific decoder structures.
We develop a principled method for quantization-aware training (QAT) of large-scale machine learning models. Specifically, we show that convex, piecewise-affine regularization (PAR) can effectively induce the model parameters to cluster towards discrete values. We minimize PAR-regularized loss functions using an aggregate proximal stochastic gradient method (AProx) and prove that it has last-iterate convergence. Our approach provides an interpretation of the straight-through estimator (STE), a widely used heuristic for QAT, as the asymptotic form of PARQ. We conduct experiments to demonstrate that PARQ obtains competitive performance on convolution- and transformer-based vision tasks.
We propose a natural quantization of a standard neural network, where the neurons correspond to qubits and the activation functions are implemented via quantum gates and measurements. The simplest quantized neural network corresponds to applying single-qubit rotations, with the rotation angles being dependent on the weights and measurement outcomes of the previous layer. This realization has the advantage of being smoothly tunable from the purely classical limit with no quantum uncertainty (thereby reproducing the classical neural network exactly) to a quantum case, where superpositions introduce an intrinsic uncertainty in the network. We benchmark this architecture on a subset of the standard MNIST dataset and find a regime of "quantum advantage," where the validation error rate in the quantum realization is smaller than that in the classical model. We also consider another approach where quantumness is introduced via weak measurements of ancilla qubits entangled with the neuron qubits. This quantum neural network also allows for smooth tuning of the degree of quantumness by controlling an entanglement angle, $g$, with $g=\frac\pi 2$ replicating the classical regime. We find that validation error is also minimized within the quantum regime in this approach. We also observe a quantum transition, with sharp loss of the quantum network's ability to learn at a critical point $g_c$. The proposed quantum neural networks are readily realizable in present-day quantum computers on commercial datasets.