Neural Network Quantization

This work proposes a unified three-stage framework that produces a quantized DNN with balanced fault and attack robustness. The first stage improves attack resilience via fine-tuning that desensitizes feature representations to small input perturbations. The second stage reinforces fault resilience through fault-aware fine-tuning under simulated bit-flip faults. Finally, a lightweight post-training adjustment integrates quantization to enhance efficiency and further mitigate fault sensitivity without degrading attack resilience. Experiments on ResNet18, VGG16, EfficientNet, and Swin-Tiny in CIFAR-10, CIFAR-100, and GTSRB show consistent gains of up to 10.35% in attack resilience and 12.47% in fault resilience, while maintaining competitive accuracy in quantized networks. The results also highlight an asymmetric interaction in which improvements in fault resilience generally increase resilience to adversarial attacks, whereas enhanced adversarial resilience does not necessarily lead to higher fault resilience.

The KV cache in self-attention has emerged as a major bottleneck in long-context and large-batch inference for LLMs. Existing approaches often treat sparsity prediction and compression as separate modules, relying on auxiliary index structures to select relevant tokens, and on complex quantization schemes to reduce memory usage. This fragmented design introduces redundant overhead and limits scalability. In this paper, we propose a novel paradigm: treating the compressed key representation not merely as storage, but as a self-indexing structure that directly enables efficient sparse attention. By designing a sign-based 1-bit vector quantization (VQ) scheme, our method unifies compression and retrieval in a single, hardware-friendly format. This approach eliminates the need for external indices or learning-based predictors, offering a lightweight yet robust solution for memory-constrained inference. All components are designed to be hardware-efficient and easy to implement. By implementing custom CUDA kernels, our method integrates seamlessly with FlashAttention, minimizing additional runtime and memory overhead. Experimental results demonstrate that our approach delivers both effectiveness and efficiency.

Low-precision neural network training has emerged as a promising direction for reducing computational costs and democratizing access to deep learning research. However, existing 4-bit quantization methods either rely on expensive GPU infrastructure or suffer from significant accuracy degradation. In this work, we present a practical method for training convolutional neural networks at true 4-bit precision using standard PyTorch operations on commodity CPUs. We introduce a novel tanh-based soft weight clipping technique that, combined with symmetric quantization, dynamic per-layer scaling, and straight-through estimators, achieves stable convergence and competitive accuracy. Training a VGG-style architecture with 3.25 million parameters from scratch on CIFAR-10, our method achieves 92.34% test accuracy on Google Colab's free CPU tier -- matching full-precision baseline performance (92.5%) with only a 0.16% gap. We further validate on CIFAR-100, achieving 70.94% test accuracy across 100 classes with the same architecture and training procedure, demonstrating that 4-bit training from scratch generalizes to harder classification tasks. Both experiments achieve 8x memory compression over FP32 while maintaining exactly 15 unique weight values per layer throughout training. We additionally validate hardware independence by demonstrating rapid convergence on a consumer mobile device (OnePlus 9R), achieving 83.16% accuracy in only 6 epochs. To the best of our knowledge, no prior work has demonstrated 4-bit quantization-aware training achieving full-precision parity on standard CPU hardware without specialized kernels or post-training quantization.

Monotone operator equilibrium networks are implicit-layer models whose output is the unique equilibrium of a monotone operator, guaranteeing existence, uniqueness, and convergence. When deployed on low-precision hardware, weights are quantized, potentially destroying these guarantees. We analyze weight quantization as a spectral perturbation of the underlying monotone inclusion. Convergence of the quantized solver is guaranteed whenever the spectral-norm weight perturbation is smaller than the monotonicity margin; the displacement between quantized and full-precision equilibria is bounded in terms of the perturbation size and margin; and a condition number characterizing the ratio of the operator norm to the margin links quantization precision to forward error. MNIST experiments confirm a phase transition at the predicted threshold: three- and four-bit post-training quantization diverge, while five-bit and above converge. The backward-pass guarantee enables quantization-aware training, which recovers provable convergence at four bits.

To tackle the huge computational demand of large foundation models, activation-aware compression techniques without retraining have been introduced. However, since these methods highly rely on calibration data, domain shift issues may arise for unseen downstream tasks. We propose a test-time quantization (TTQ) framework which compresses large models on the fly at inference time to resolve this issue. With an efficient online calibration, instant activation-aware quantization can adapt every prompt regardless of the downstream tasks, yet achieving inference speedup. Several experiments demonstrate that TTQ can improve the quantization performance over state-of-the-art baselines.

Vision-language-action (VLA) models achieve favorable task performance, yet runtime errors in closed-loop execution evolve with alternating updates of actions and observations. Existing VLA quantization methods mainly trade off inference speed and model performance while rarely investigating how quantization alters runtime errors. By comparing error distributions between full-precision and quantized models, we identify distinct evolutionary patterns for the two types of errors: quantization enlarges the variance of translation error distributions and increases their dispersion, whereas the distribution center of rotation errors gradually shifts across execution steps, demonstrating cumulative drift. Motivated by this observation, we rethink the optimal quantization strategy for VLA models and propose DyQ-VLA, a runtime-error-aware quantization framework, which dynamically selects activation precision according to execution steps and tracks as well as compensates rotation errors via accumulated quantization residuals. Corresponding operators and runtime adaptation strategies are devised within the framework to enable dynamic-precision execution. Experiments show that DyQ-VLA achieves a 1.88 to 1.93 inference speedups while maintaining comparable or higher average task success rates. Moreover, it reduces mean execution steps by 17.7% to 19.8%. Our code is here: https://anonymous.4open.science/r/DyQ-VLA-7F51/.

Equivariant Graph Neural Networks (GNNs) are essential for physically consistent molecular simulations but suffer from high computational costs and memory bottlenecks, especially with high-order representations. While low-bit quantization offers a solution, applying it naively to rotation-sensitive features destroys the SO(3)-equivariant structure, leading to significant errors and violations of conservation laws. To address this issue, in this work, we propose a Geometric-Aware Quantization (GAQ) framework that compresses and accelerates equivariant models while rigorously preserving continuous symmetry in discrete spaces. Our approach introduces three key contributions: (1) a Magnitude-Direction Decoupled Quantization (MDDQ) scheme that separates invariant lengths from equivariant orientations to maintain geometric fidelity; (2) a symmetry-aware training strategy that treats scalar and vector features with distinct quantization schedules; and (3) a robust attention normalization mechanism to stabilize gradients in low-bit regimes. Experiments on the rMD17 benchmark demonstrate that our W4A8 models match the accuracy of FP32 baselines (9.31 meV vs. 23.20 meV) while reducing Local Equivariance Error (LEE) by over 30x compared to naive quantization. On consumer hardware, GAQ achieves 2.39x inference speedup and 4x memory reduction, enabling stable, energy-conserving molecular dynamics simulations for nanosecond timescales.

Quantization addresses the high resource demand for large language models (LLMs) by alleviating memory pressure and bandwidth congestion and providing significantly scaled compute power with a tolerable impact on accuracy. Four-bit floating point (FP4), the lowest-precision format that preserves essential numerical properties such as exponent and sign, has begun to be adopted in cutting-edge architectures, including Blackwell and AMD CDNA, to support LLM quantization and reduce deployment costs. Although aggressive quantization can yield efficiency gains, the quantization sensitivity of within-transformer layers and whether these sensitivities generalize across existing FP4 formats and model scales remain underexplored. To elucidate quantization sensitivity, this study conducts a systematic analysis of two FP4 formats, MXFP4 and NVFP4, across three Qwen2.5 model scales (0.5B, 7B, and 14B), using controlled component-wise and block-wise isolation methodologies. We observe that MLP up- and down-projection layers consistently dominate in terms of sensitivity, while gate and attention projections are moderately and substantially less sensitive to FP4 quantization, respectively. We further find that sensitivity does not universally localize to the final blocks, but early blocks can be highly sensitive, particularly under MXFP4. Our results provide a diagnostic characterization of the inference behavior of FP4 across components, depths, and FP4 formats.

The growing demand for machine learning applications in the context of the Internet of Things calls for new approaches to optimize the use of limited compute and memory resources. Despite significant progress that has been made w.r.t. reducing model sizes and improving efficiency, many applications still require remote servers to provide the required resources. However, such approaches rely on transmitting data from edge devices to remote servers, which may not always be feasible due to bandwidth, latency, or energy constraints. We propose a task-specific, trainable feature quantization layer that compresses the input features of a neural network. This can significantly reduce the amount of data that needs to be transferred from the device to a remote server. In particular, the layer allows each input feature to be quantized to a user-defined number of bits, enabling a simple on-device compression at the time of data collection. The layer is designed to approximate step functions with sigmoids, enabling trainable quantization thresholds. By concatenating outputs from multiple sigmoids, introduced as bitwise soft quantization, it achieves trainable quantized values when integrated with a neural network. We compare our method to full-precision inference as well as to several quantization baselines. Experiments show that our approach outperforms standard quantization methods, while maintaining accuracy levels close to those of full-precision models. In particular, depending on the dataset, compression factors of $5\times$ to $16\times$ can be achieved compared to $32$-bit input without significant performance loss.

This paper considers the problem of converting a given dense linear layer to low precision. The tradeoff between compressed length and output discrepancy is analyzed information theoretically (IT). It is shown that a popular GPTQ algorithm may have an arbitrarily large gap to the IT limit. To alleviate this problem, a novel algorithm, termed ``WaterSIC'', is proposed and is shown to be within a rate gap of 0.255 bits to the IT limit, uniformly over all possible covariance matrices of input activations. The key innovation of WaterSIC's is to allocate different quantization rates to different columns (in-features) of the weight matrix, mimicking the classical IT solution known as "waterfilling". Applying WaterSIC to the Llama and Qwen family of LLMs establishes new state-of-the-art performance for all quantization rates from 1 to 4 bits. Our code is available at https://github.com/egorlifar/watersic.

Post-training quantization (PTQ) with computational invariance for Large Language Models~(LLMs) have demonstrated remarkable advances, however, their application to Multimodal Large Language Models~(MLLMs) presents substantial challenges. In this paper, we analyze SmoothQuant as a case study and identify two critical issues: Smoothing Misalignment and Cross-Modal Computational Invariance. To address these issues, we propose Modality-Aware Smoothing Quantization (MASQuant), a novel framework that introduces (1) Modality-Aware Smoothing (MAS), which learns separate, modality-specific smoothing factors to prevent Smoothing Misalignment, and (2) Cross-Modal Compensation (CMC), which addresses Cross-modal Computational Invariance by using SVD whitening to transform multi-modal activation differences into low-rank forms, enabling unified quantization across modalities. MASQuant demonstrates stable quantization performance across both dual-modal and tri-modal MLLMs. Experimental results show that MASQuant is competitive among the state-of-the-art PTQ algorithms. Source code: https://github.com/alibaba/EfficientAI.

Deep neural networks have achieved strong performance in image classification tasks due to their ability to learn complex patterns from high-dimensional data. However, their large computational and memory requirements often limit deployment on resource-constrained platforms such as remote sensing devices and edge systems. Network compression techniques have therefore been proposed to reduce model size and computational cost while maintaining predictive performance. In this study, we conduct a systematic evaluation of neural network compression methods for a remote sensing application, namely hyperspectral land cover classification. Specifically, we examine three widely used compression strategies for convolutional neural networks: pruning, quantization, and knowledge distillation. Experiments are conducted on two benchmark hyperspectral datasets, considering classification accuracy, memory consumption, and inference efficiency. Our results demonstrate that compressed models can significantly reduce model size and computational cost while maintaining competitive classification performance. These findings provide insights into the trade-offs between compression ratio, efficiency, and accuracy, and highlight the potential of compression techniques for enabling efficient deep learning deployment in remote sensing applications.

Quantization can drastically increase the efficiency of large language and vision models, but typically incurs an accuracy drop. Recently, function-preserving transforms (e.g. rotations, Hadamard transform, channel-wise scaling) have been successfully applied to reduce post-training quantization error, yet a principled explanation remains elusive. We analyze linear-layer quantization via the signal-to-quantization-noise ratio (SQNR), showing that for uniform integer quantization at a fixed bit width, SQNR decomposes into (i) the concentration of weights and activations (capturing spread and outliers), and (ii) the alignment of their dominant variation directions. This reveals an actionable insight: beyond concentration - the focus of most prior transforms (e.g. rotations or Hadamard) - improving alignment between weight and activation can further reduce quantization error. Motivated by this, we introduce block Concentration-Alignment Transforms (CAT), a lightweight linear transformation that uses a covariance estimate from a small calibration set to jointly improve concentration and alignment, approximately maximizing SQNR. Experiments across several LLMs show that CAT consistently matches or outperforms prior transform-based quantization methods at 4-bit precision, confirming the insights gained in our framework.

Diffusion Transformers (DiTs) achieve state-of-the-art video generation quality, but their substantial memory and computational footprints hinder edge deployment. Quantization can reduce these costs, yet existing methods often degrade video quality due to high activation variation and the difficulty of preserving semantic and temporal coherence. We propose SemanticDialect, which advances block-wise mixed-format quantization. In this framework, each block selects an optimal format (dialect) from a candidate set (formatbook), which is augmented with lookup tables that store quantization errors and quantized indices, enabling efficient per-block format selection and quantization with minimal online overhead. We further introduce attention-guided activation decomposition, which reduces quantization error via residual quantization, and semantic-aware dialect assignment (SeDA), which reduces cross-token quantization inconsistency by enforcing format uniformity among semantically correlated tokens. Experiments demonstrate that SemanticDialect outperforms prior quantization methods and block-wise formats (MXFP4, NVFP4) while approaching FP16 quality on Open-Sora 2.0. We also validate hardware deployability through RTL design and GPU kernel implementation.

Quantization-Aware Pre-Training (QAPT) is an effective technique to reduce the compute and memory overhead of Deep Neural Networks while improving their energy efficiency on edge devices. Existing QAPT methods produce models stored in compute-efficient data types (e.g. integers) that are not information theoretically optimal (ITO). On the other hand, existing ITO data types (e.g. Quantile/NormalFloat Quantization) are not compute-efficient. We propose BBQ, the first ITO quantization method that is also compute-efficient. BBQ builds on our key insight that since learning is domain-agnostic, the output of a quantizer does not need to reside in the same domain as its input. BBQ performs ITO quantization in its input domain, and returns its output in a compute-efficient domain where ITO data types are mapped to compute-efficient data types. Without sacrificing compute efficiency, BBQ outperforms prior SOTA QAPT methods by a perplexity reduction of up to 2 points for 4-bit models, up to 4 points for 3-bit models, up to 5 points for 2-bit models, and up to 18 points for 1-bit models. Code is available at https://github.com/1733116199/bbq.

Post-Training Quantization (PTQ) has emerged as an effective technique for alleviating the substantial computational and memory overheads of Vision-Language Models (VLMs) by compressing both weights and activations without retraining the full model. Existing PTQ methods primarily rely on static identification and global compensation of sensitive or outlier channels, yet they often overlook the distributional differences of these important channels across inputs, leading to unsatisfactory quantization. In this work, we observe that the distributions and occurrence frequencies of important channels vary significantly both across modalities and among tokens, even within the same modality. Accordingly, we propose Quant Experts (QE), a token-aware adaptive error compensation with mixture-of-experts for VLMs quantization. QE divides the important channels into token-independent and token-dependent groups. For the former, a shared expert is designed for most tokens to compensate for global quantization error using a low-rank adapter. For the latter, routed experts including multiple routed low-rank adapters are elaborated to compensate for local quantization error related to specific tokens. Extensive experiments demonstrate that QE consistently enhances task accuracy across various quantization settings and model scales, ranging from 2B to 70B parameters, while maintaining performance comparable to full-precision models.

Standard mixed-precision training of neural networks requires many bytes of accelerator memory for each model parameter. These bytes reflect not just the parameter itself, but also its gradient and one or more optimizer state variables. With each of these values typically requiring 4 bytes, training even a 7 billion parameter model can be impractical for researchers with less than 100GB of accelerator memory. We introduce FlashOptim, a suite of optimizations that reduces per-parameter memory by over 50% while preserving model quality and API compatibility. Our approach introduces two key techniques. First, we improve master weight splitting by finding and exploiting a tight bound on its quantization error. Second, we design companding functions that greatly reduce the error in 8-bit optimizer state quantization. Together with 16-bit gradients, these techniques reduce AdamW memory from 16 bytes to 7 bytes per parameter, or 5 bytes with gradient release. They also cut model checkpoint sizes by more than half. Experiments with FlashOptim applied to SGD, AdamW, and Lion show no measurable quality degradation on any task from a collection of standard vision and language benchmarks, including Llama-3.1-8B finetuning.

Neural network accelerators have been widely applied to edge devices for complex tasks like object tracking, image recognition, etc. Previous works have explored the quantization technologies in related lightweight accelerator designs to reduce hardware resource consumption. However, low precision leads to high accuracy loss in inference. Therefore, mixed-precision quantization becomes an alternative solution by applying different precision in different layers to trade off resource consumption and accuracy. Because regular designs for multiplication on hardware cannot support the precision reconfiguration for a multi-precision Quantized Neural Network (QNN) model in runtime, we propose a runtime reconfigurable multi-precision multi-channel bitwise systolic array design for QNN accelerators. We have implemented and evaluated our work on the Ultra96 FPGA platform. Results show that our work can achieve 1.3185 to 3.5671 times speedup in inferring mixed-precision models and has less critical path delay, supporting a higher clock frequency (250MHz).

When transformer-based language models are deployed for text generation, most of the inference time is spent in the decoding stage, where output tokens are generated sequentially. Reducing the hardware cost of each decoding step is therefore critical for efficient long-context generation. A major bottleneck is the key-value (KV) cache, whose size grows with sequence length and often dominates the model's memory footprint. Prior work has proposed quantization methods to compress the KV cache while minimizing its loss of precision. We present InnerQ, a hardware-aware KV cache quantization scheme that reduces decode latency without compromising evaluation performance. InnerQ performs group-wise quantization by grouping cache matrices along their inner dimension. This grouping strategy aligns dequantization with vector-matrix multiplication and increases data reuse across GPU compute units. As a result, InnerQ reduces memory access and accelerates dequantization, achieving an average $1.3\times$ speedup over prior KV cache quantization methods and $2.7\times$ over the non-quantized baseline. To maintain fidelity under aggressive compression, InnerQ incorporates three techniques: (i) hybrid quantization, which chooses symmetric or asymmetric quantization for each group based on local statistics; (ii) high-precision windows for both recent tokens and attention sink tokens to mitigate outlier leakage; and (iii) per-channel normalization of the key cache, computed once during prefill and folded into the model parameters to eliminate runtime overhead. Beyond reducing latency, experiments on Llama and Mistral models show that InnerQ also improves few-shot evaluation scores relative to prior KV cache quantization methods.

Quantization-Aware Training from scratch has emerged as a promising approach for building efficient large language models (LLMs) with extremely low-bit weights (sub 2-bit), which can offer substantial advantages for edge deployment. However, existing methods still fail to achieve satisfactory accuracy and scalability. In this work, we identify a parameter democratization effect as a key bottleneck: the sensitivity of all parameters becomes homogenized, severely limiting expressivity. To address this, we propose pQuant, a method that decouples parameters by splitting linear layers into two specialized branches: a dominant 1-bit branch for efficient computation and a compact high-precision branch dedicated to preserving the most sensitive parameters. Through tailored feature scaling, we explicitly guide the model to allocate sensitive parameters to the high-precision branch. Furthermore, we extend this branch into multiple, sparsely-activated experts, enabling efficient capacity scaling. Extensive experiments indicate our pQuant achieves state-of-the-art performance in extremely low-bit quantization.

Deep neural networks (DNNs) are essential for performing advanced tasks on edge or mobile devices, yet their deployment is often hindered by severe resource constraints, including limited memory, energy, and computational power. While uniform quantization provides a straightforward approach to compress model and reduce hardware requirement, it fails to fully leverage the varying robustness across layers, and often lead to accuracy degradation or suboptimal resource usage, particularly at low bitwidths. In contrast, heterogeneous quantization, which allocates different bitwidths to individual layers, can mitigate these drawbacks. Nonetheless, current heterogeneous quantization methods either needs huge brute-force design space search or lacks the adaptability to meet different hardware conditions, such as memory size, energy budget, and latency requirement. Filling these gaps, this work introduces SigmaQuant, an adaptive layer-wise heterogeneous quantization framework designed to efficiently balance accuracy and resource usage for varied edge environments without exhaustive search.

This paper presents a compression framework for Reservoir Computing that enables systematic design-space exploration of trade-offs among quantization levels, pruning rates, model accuracy, and hardware efficiency. The proposed approach leverages a sensitivity-based pruning mechanism to identify and remove less critical quantized weights with minimal impact on model accuracy, thereby reducing computational overhead while preserving accuracy. We perform an extensive trade-off analysis to validate the effectiveness of the proposed framework and the impact of pruning and quantization on model performance and hardware parameters. For this evaluation, we employ three time-series datasets, including both classification and regression tasks. Experimental results across selected benchmarks demonstrate that our proposed approach maintains high accuracy while substantially improving computational and resource efficiency in FPGA-based implementations, with variations observed across different configurations and time series applications. For instance, for the MELBOEN dataset, an accelerator quantized to 4-bit at a 15% pruning rate reduces resource utilization by 1.2% and the Power Delay Product (PDP) by 50.8% compared to an unpruned model, without any noticeable degradation in accuracy.

Dynamic runtime latency and memory constraints necessitate flexible large language model (LLM) deployment, where an LLM can be inferred with various quantization precisions based on available computational resources. Recent work on such any-precision quantization either relies on hardware-inefficient vector quantization or induces additional scaling factors when switching between bit-widths. Meanwhile, existing post-training quantization (PTQ) methods calibrated for a fixed low precision show poor generalizability under runtime precision change. In this work, we attribute the source of poor generalization across bit-widths to a precision-dependent outlier migration phenomenon where the distribution of PTQ-sensitive tokens changes across precisions. Motivated by this observation, we propose MoBiQuant, a novel any-precision Mixture-of-Bits quantization framework that adjusts weight precision for flexible LLM inference based on token sensitivity. Specifically, we propose a many-in-one recursive residual quantization that can iteratively reconstruct higher-precision weights at runtime and mitigates outlier migration with a token-aware router to dynamically select the optimal inference precision of each token.Extensive experiments show that MoBiQuant matches or surpasses frontier single-precision PTQ while exhibiting strong elasticity, achieving significant memory savings and throughput gains of up to $1.34\times$ over state-of-the-art any-precision methods.

We present a framework for end-to-end joint quantization of Vision Transformers trained on ImageNet for the purpose of image classification. Unlike prior post-training or block-wise reconstruction methods, we jointly optimize over the entire set of all layers and inter-block dependencies without any labeled data, scaling effectively with the number of samples and completing in just one hour on a single GPU for ViT-small. We achieve state-of-the-art W4A4 and W3A3 accuracies on ImageNet and, to the best of our knowledge, the first PTQ results that maintain strong accuracy on ViT, DeiT, and Swin-T models under extremely low-bit settings (W1.58A8), demonstrating the potential for efficient edge deployment. Furthermore, we introduce a data-free calibration strategy that synthesizes diverse, label-free samples using Stable Diffusion Turbo guided by learned multi-mode prompts. By encouraging diversity in both the learned prompt embeddings and the generated image features, our data-free approach achieves performance on par with real-data ImageNet calibration and surpasses simple text-prompt baselines such as "a <adjective> photo of <adjective> <cls>".

Neuromorphic accelerators promise unparalleled energy efficiency and computational density for spiking neural networks, especially in wearable biomedical devices and neural prosthetics where power constraints are stringent. However, most existing platforms exhibit rigid architectures with limited configurability, restricting their adaptability to heterogeneous biological signals and diverse design objectives. To address these limitations, we present Flexi-NeurA--a flexible neuromorphic accelerator that unifies configurability and efficiency. Flexi-NeurA allows users to customize neuron models, network structures, and precision settings at design time. By pairing these design-time configurability features with a time-multiplexed and event-driven processing approach, Flexi-NeurA reduces the required hardware resources and total power while preserving high efficiency and low inference latency. Complementing this, we introduce Flex-plorer, a design-space exploration tool that determines cost-effective fixed-point precisions for critical parameters--such as decay factors, synaptic weights, and membrane potentials--based on user-defined trade-offs between accuracy and resource usage. Based on the configuration selected through the Flex-plorer process, RTL code is configured to match the specified design. Comprehensive evaluations across distinct domains--biomedical auditory processing, dynamic vision sensor gesture recognition, and standard vision classification--demonstrate that the hardware/software co-framework successfully balances accuracy and power budgets for diverse applications. A 3-layer 256-128-10 fully connected network with LIF neurons mapped onto two processing cores achieves 96.23% accuracy on MNIST with 1.1 ms inference latency, utilizing only 1,623 logic cells, 7 BRAMs, and 111 mW of total power--demonstrating superior resource efficiency compared to SoTA hardware baselines.

The rapid proliferation of Large Language Models (LLMs) has revolutionized Natural Language Processing (NLP) but has simultaneously created a "resource divide." State-of-the-art legal intelligence systems typically rely on massive parameter counts (7B+) and cloud-based inference, rendering them inaccessible to practitioners in resource-constrained environments and posing significant data sovereignty risks. This paper introduces Quecto-V1, a domain-specific Small Language Model (SLM) engineered to democratize access to Indian legal intelligence. Built upon a custom configuration of the GPT-2 architecture (124 million parameters), Quecto-V1 was trained from scratch exclusively on a corpus of Indian statutes, including the Indian Penal Code (IPC), the Code of Criminal Procedure (CrPC), and the Constitution of India. Unlike generalist models, which prioritize broad world knowledge, our approach maximizes "lexical density" within the legal domain. Furthermore, we address the deployment bottleneck by applying post-training 8-bit quantization (GGUF format), compressing the model to a memory footprint of under 150 MB. Our empirical analysis demonstrates that Quecto-V1 achieves high fidelity in retrieving statutory definitions and penal provisions, outperforming general-purpose SLMs in domain-specific exact match tasks while running entirely offline on consumer-grade CPUs. We further present an ablation study showing that 8-bit quantization yields a 74% reduction in model size with less than 3.5% degradation in retrieval accuracy compared to full-precision baselines. These findings suggest that for specialized, high-stakes domains like law, domain-specific training coupled with aggressive quantization offers a viable, privacy-preserving alternative to monolithic cloud models.

This paper presents an end-to-end demonstration of a viable, ultra-fast, radiation-hard machine learning (ML) application on FPGAs, which could be used in future high-energy physics experiments. We present a three-fold contribution, with the PicoCal calorimeter, planned for the LHCb Upgrade II experiment, used as a test case. First, we develop a lightweight autoencoder to compress a 32-sample timing readout, representative of that of the PicoCal, into a two-dimensional latent space. Second, we introduce a systematic, hardware-aware quantization strategy and show that the model can be reduced to 10-bit weights with minimal performance loss. Third, as a barrier to the adoption of on-detector ML is the lack of support for radiation-hard FPGAs in the High-Energy Physics community's standard ML synthesis tool, hls4ml, we develop a new backend for this library. This new back-end enables the automatic translation of ML models into High-Level Synthesis (HLS) projects for the Microchip PolarFire family of FPGAs, one of the few commercially available and radiation hard FPGAs. We present the synthesis of the autoencoder on a target PolarFire FPGA, which indicates that a latency of 25 ns can be achieved. We show that the resources utilized are low enough that the model can be placed within the inherently protected logic of the FPGA. Our extension to hls4ml is a significant contribution, paving the way for broader adoption of ML on FPGAs in high-radiation environments.

Multi-hop all-reduce is the de facto backbone of large model training. As the training scale increases, the network often becomes a bottleneck, motivating the reduction of the volume of transmitted data. Accordingly, recent systems have demonstrated significant acceleration of the training process using gradient quantization. However, these systems are not optimized for multi-hop aggregation, where entries are partially summed multiple times along their aggregation topology. We present DynamiQ, a quantization framework that bridges the gap between quantization best practices and multi-hop aggregation. DynamiQ introduces novel techniques to better represent partial sums, codesigned with a decompress accumulate recompress fused kernel to facilitate fast execution. We extend PyTorch DDP to support DynamiQ over NCCL P2P, and across different LLMs, tasks, and scales, we demonstrate consistent improvement of up to 34.2% over the best among state-of-the-art methods such as Omni-Reduce, THC, and emerging standards such as MXFP4, MXFP6, and MXFP8. Further, DynamiQ is the only evaluated method that consistently reaches near-baseline accuracy (e.g., 99.9% of the BF16 baseline) and does so while significantly accelerating the training.

Post-training quantization reduces the memory needed to run large language models but alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced bias flipping, in which quantization causes models to change responses from biased to unbiased and vice versa, up to 21% of the time, despite no change in aggregate bias scores. These flips are strongly associated with model uncertainty, where the responses with high uncertainty are 3-11x more likely to change than the confident ones. Quantization strength amplifies this effect, with 4-bit quantized models exhibiting 4-6x more behavioral changes than 8-bit quantized models. Critically, these changes create asymmetric impacts across demographic groups, where bias can worsen by up to 18.6% for some groups while improving by 14.1% for others, yielding misleadingly neutral aggregate outcomes. Larger models show no consistent robustness advantage, and group-specific shifts vary unpredictably across model families. Our findings demonstrate that compression fundamentally alters bias patterns, requiring crucial post-quantization evaluation and interventions to ensure reliability in practice.

Matryoshka Quantization (MatQuant), Any-Precision-LLM (AP) and AnyBCQ (AB) are recent quantization approaches showing that a single integer-quantized model can be served across multiple precisions. In this paradigm, lower-precision models are extracted from a higher-precision model by simply reading fewer bits of the weights. This enables a single checkpoint to cover a wide range of memory and latency budgets, but makes both quantization and efficient execution substantially harder. Existing methods rely on expensive quantization-aware training (QAT) or gradient-based post-training quantization (PTQ) rather than fast one-shot PTQ, and offer limited system support: dedicated kernels are either missing or restricted to single- or small-batch decoding. We address these limitations with Post-Training Matryoshka Quantization (MatGPTQ), an end-to-end pipeline for nested-model quantization and inference. MatGPTQ casts Matryoshka quantization as multi-precision error compensation, producing a single "sliceable" parent model jointly optimized for multiple target precisions in one pass over a small calibration set. We further refine the MatQuant representation so that an $r$-bit model reads exactly $r$ bits, and introduce the first dedicated inference kernels for this format, supporting batch sizes beyond one and integrated into vLLM. Across standard LLMs and benchmarks, MatGPTQ outperforms MatQuant while remaining competitive with AP and AB at the smallest checkpoint size, and our kernels achieve end-to-end speedups of up to 3.5$\times$ over BF16 at the low-bit regime. Overall, MatGPTQ makes nested quantized models practical to serve from a single, compact checkpoint. Code is available at https://github.com/IST-DASLab/MatGPTQ.