Artificial intelligence and machine learning models deployed on edge devices, e.g., for quality control in Additive Manufacturing (AM), are frequently small in size. Such models usually have to deliver highly accurate results within a short time frame. Methods that are commonly employed in literature start out with larger trained models and try to reduce their memory and latency footprint by structural pruning, knowledge distillation, or quantization. It is, however, also possible to leverage hardware-aware Neural Architecture Search (NAS), an approach that seeks to systematically explore the architecture space to find optimized configurations. In this study, a hardware-aware NAS workflow is introduced that couples an edge device located in Belgium with a powerful High-Performance Computing system in Germany, to train possible architecture candidates as fast as possible while performing real-time latency measurements on the target hardware. The approach is verified on a use case in the AM domain, based on the open RAISE-LPBF dataset, achieving ~8.8 times faster inference speed while simultaneously enhancing model quality by a factor of ~1.35, compared to a human-designed baseline.
Large language models (LLMs) now support context windows exceeding 128K tokens, but this comes with significant memory requirements and high inference latency. Quantization can mitigate these costs, but may degrade performance. In this work, we present the first systematic evaluation of quantized LLMs on tasks with long-inputs (>64K tokens) and long-form outputs. Our evaluation spans 9.7K test examples, five quantization methods (FP8, GPTQ-int8, AWQ-int4, GPTQ-int4, BNB-nf4), and five models (Llama-3.1 8B and 70B; Qwen-2.5 7B, 32B, and 72B). We find that, on average, 8-bit quantization preserves accuracy (~0.8% drop), whereas 4-bit methods lead to substantial losses, especially for tasks involving long context inputs (drops of up to 59%). This degradation tends to worsen when the input is in a language other than English. Crucially, the effects of quantization depend heavily on the quantization method, model, and task. For instance, while Qwen-2.5 72B remains robust under BNB-nf4, Llama-3.1 70B experiences a 32% performance drop on the same task. These findings highlight the importance of a careful, task-specific evaluation before deploying quantized LLMs, particularly in long-context scenarios and with languages other than English.
State Space Models (SSMs), like recent Mamba2, have achieved remarkable performance and received extensive attention. However, deploying Mamba2 on resource-constrained edge devices encounters many problems: severe outliers within the linear layer challenging the quantization, diverse and irregular element-wise tensor operations, and hardware-unfriendly nonlinear functions in the SSM block. To address these issues, this paper presents FastMamba, a dedicated accelerator on FPGA with hardware-algorithm co-design to promote the deployment efficiency of Mamba2. Specifically, we successfully achieve 8-bit quantization for linear layers through Hadamard transformation to eliminate outliers. Moreover, a hardware-friendly and fine-grained power-of-two quantization framework is presented for the SSM block and convolution layer, and a first-order linear approximation is developed to optimize the nonlinear functions. Based on the accurate algorithm quantization, we propose an accelerator that integrates parallel vector processing units, pipelined execution dataflow, and an efficient SSM Nonlinear Approximation Unit, which enhances computational efficiency and reduces hardware complexity. Finally, we evaluate FastMamba on Xilinx VC709 FPGA. For the input prefill task on Mamba2-130M, FastMamba achieves 68.80\times and 8.90\times speedup over Intel Xeon 4210R CPU and NVIDIA RTX 3090 GPU, respectively. In the output decode experiment with Mamba2-2.7B, FastMamba attains 6\times higher energy efficiency than RTX 3090 GPU.
Recently, significant progress has been made in developing reasoning-capable Large Language Models (LLMs) through long Chain-of-Thought (CoT) techniques. However, this long-CoT reasoning process imposes substantial memory overhead due to the large Key-Value (KV) Cache memory overhead. Post-training KV Cache quantization has emerged as a promising compression technique and has been extensively studied in short-context scenarios. However, directly applying existing methods to long-CoT LLMs causes significant performance degradation due to the following two reasons: (1) Large cumulative error: Existing methods fail to adequately leverage available memory, and they directly quantize the KV Cache during each decoding step, leading to large cumulative quantization error. (2) Short-context calibration: Due to Rotary Positional Embedding (RoPE), the use of short-context data during calibration fails to account for the distribution of less frequent channels in the Key Cache, resulting in performance loss. We propose Progressive Mixed-Precision KV Cache Quantization (PM-KVQ) for long-CoT LLMs to address the above issues in two folds: (1) To reduce cumulative error, we design a progressive quantization strategy to gradually lower the bit-width of KV Cache in each block. Then, we propose block-wise memory allocation to assign a higher bit-width to more sensitive transformer blocks. (2) To increase the calibration length without additional overhead, we propose a new calibration strategy with positional interpolation that leverages short calibration data with positional interpolation to approximate the data distribution of long-context data. Extensive experiments on 7B-70B long-CoT LLMs show that PM-KVQ improves reasoning benchmark performance by up to 8% over SOTA baselines under the same memory budget. Our code is available at https://github.com/thu-nics/PM-KVQ.
The emergence of multi-codebook neutral audio codecs such as Residual Vector Quantization (RVQ) and Group Vector Quantization (GVQ) has significantly advanced Large-Language-Model (LLM) based Text-to-Speech (TTS) systems. These codecs are crucial in separating semantic and acoustic information while efficiently harnessing semantic priors. However, since semantic and acoustic information cannot be fully aligned, a significant drawback of these methods when applied to LLM-based TTS is that large language models may have limited access to comprehensive audio information. To address this limitation, we propose DistilCodec and UniTTS, which collectively offer the following advantages: 1) This method can distill a multi-codebook audio codec into a single-codebook audio codec with 32,768 codes while achieving a near 100\% utilization. 2) As DistilCodec does not employ a semantic alignment scheme, a large amount of high-quality unlabeled audio (such as audiobooks with sound effects, songs, etc.) can be incorporated during training, further expanding data diversity and broadening its applicability. 3) Leveraging the comprehensive audio information modeling of DistilCodec, we integrated three key tasks into UniTTS's pre-training framework: audio modality autoregression, text modality autoregression, and speech-text cross-modal autoregression. This allows UniTTS to accept interleaved text and speech/audio prompts while substantially preserving LLM's text capabilities. 4) UniTTS employs a three-stage training process: Pre-Training, Supervised Fine-Tuning (SFT), and Alignment. Source code and model checkpoints are publicly available at https://github.com/IDEA-Emdoor-Lab/UniTTS and https://github.com/IDEA-Emdoor-Lab/DistilCodec.
Quantizing large language models (LLMs) to 1-bit precision significantly reduces computational costs, but existing quantization techniques suffer from noticeable performance degradation when using weight and activation precisions below 4 bits (W4A4). In this paper, we propose a post-training quantization framework with W(1+1)A(1*4) configuration, where weights are quantized to 1 bit with an additional 1 bit for fine-grain grouping and activations are quantized to 1 bit with a 4-fold increase in the number of channels. For weight quantization, we propose utilizing Hessian-aware fine-grained grouping along with an EM-based quantization scheme. For activation quantization, we decompose INT4-quantized activations into a 4 * INT1 format equivalently and simultaneously smooth the scaling factors based on quantization errors, which further reduces the quantization errors in activations. Our method surpasses state-of-the-art (SOTA) LLM quantization baselines on W2A4 across multiple tasks, pushing the boundaries of existing LLM quantization methods toward fully binarized models.
Visual autoregressive (VAR) modeling has marked a paradigm shift in image generation from next-token prediction to next-scale prediction. VAR predicts a set of tokens at each step from coarse to fine scale, leading to better image quality and faster inference speed compared to existing diffusion models. However, the large parameter size and computation cost hinder its deployment on edge devices. To reduce the memory and computation cost, we propose FPQVAR, an efficient post-training floating-point (FP) quantization framework for VAR featuring algorithm and hardware co-design. At the algorithm level, we first identify the challenges of quantizing VAR. To address them, we propose Dual Format Quantization for the highly imbalanced input activation. We further propose Group-wise Hadamard Transformation and GHT-Aware Learnable Transformation to address the time-varying outlier channels. At the hardware level, we design the first low-bit FP quantizer and multiplier with lookup tables on FPGA and propose the first FPGA-based VAR accelerator featuring low-bit FP computation and an elaborate two-level pipeline. Extensive experiments show that compared to the state-of-the-art quantization method, our proposed FPQVAR significantly improves Fr\'echet Inception Distance (FID) from 10.83 to 3.58, Inception Score (IS) from 175.9 to 241.5 under 4-bit quantization. FPQVAR also significantly improves the performance of 6-bit quantized VAR, bringing it on par with the FP16 model. Our accelerator on AMD-Xilinx VCK190 FPGA achieves a throughput of 1.1 image/s, which is 3.1x higher than the integer-based accelerator. It also demonstrates 3.6x and 2.8x higher energy efficiency compared to the integer-based accelerator and GPU baseline, respectively.
Large Language Models (LLMs) have demonstrated remarkable proficiency across a wide range of tasks. However, LLMs often require larger batch sizes to enhance throughput or longer context lengths to meet task demands, which significantly increases the memory resource consumption of the Key-Value (KV) cache during inference, becoming a major bottleneck in LLM deployment. To address this issue, quantization is a common and straightforward approach. Currently, quantization methods for activations are limited to 8-bit, and quantization to even lower bits can lead to substantial accuracy drops. To further save space by quantizing the KV cache to even lower bits, we analyzed the element distribution of the KV cache and designed the NQKV algorithm. Since the elements within each block of the KV cache follow a normal distribution, NQKV employs per-block quantile quantization to achieve information-theoretically optimal quantization error. Without significantly compromising model output quality, NQKV enables the OPT model to perform inference with an 2x larger batch size or a 4x longer context length, and it improves throughput by 9.3x compared to when the KV cache is not used.
Quantization became a necessary tool for serving ever-increasing Large Language Models (LLMs). RTN (Round-to-Nearest) is perhaps the simplest quantization technique that has been around well before LLMs surged to the forefront of machine learning (ML) research. Yet, it has been largely dismissed by recent and more advanced quantization methods that claim superiority over RTN in nearly every aspect of performance. This work aims to dispel this established point of view, showing that RTN is not only much cheaper to apply, but also its token generation throughput can be better than and accuracy can be similar to more advanced alternatives. In particular, we discuss our implementation of RTN based on the recent Marlin kernels and demonstrate how the accuracy of RTN can be gradually improved by selectively increasing the data precision format of certain model layers and modules. Based on our results, we argue that RTN presents a viable and practical choice for quantizing LLMs.
Quantum optimization is the most mature quantum computing technology to date, providing a promising approach towards efficiently solving complex combinatorial problems. Methods such as adiabatic quantum computing (AQC) have been employed in recent years on important optimization problems across various domains. In deep learning, deep neural networks (DNN) have reached immense sizes to support new predictive capabilities. Optimization of large-scale models is critical for sustainable deployment, but becomes increasingly challenging with ever-growing model sizes and complexity. While quantum optimization is suitable for solving complex problems, its application to DNN optimization is not straightforward, requiring thorough reformulation for compatibility with commercially available quantum devices. In this work, we explore the potential of adopting AQC for fine-grained pruning-quantization of convolutional neural networks. We rework established heuristics to formulate model compression as a quadratic unconstrained binary optimization (QUBO) problem, and assess the solution space offered by commercial quantum annealing devices. Through our exploratory efforts of reformulation, we demonstrate that AQC can achieve effective compression of practical DNN models. Experiments demonstrate that adiabatic quantum computing (AQC) not only outperforms classical algorithms like genetic algorithms and reinforcement learning in terms of time efficiency but also excels at identifying global optima.
Service providers must encode a large volume of noisy videos to meet the demand for user-generated content (UGC) in online video-sharing platforms. However, low-quality UGC challenges conventional codecs based on rate-distortion optimization (RDO) with full-reference metrics (FRMs). While effective for pristine videos, FRMs drive codecs to preserve artifacts when the input is degraded, resulting in suboptimal compression. A more suitable approach used to assess UGC quality is based on non-reference metrics (NRMs). However, RDO with NRMs as a measure of distortion requires an iterative workflow of encoding, decoding, and metric evaluation, which is computationally impractical. This paper overcomes this limitation by linearizing the NRM around the uncompressed video. The resulting cost function enables block-wise bit allocation in the transform domain by estimating the alignment of the quantization error with the gradient of the NRM. To avoid large deviations from the input, we add sum of squared errors (SSE) regularization. We derive expressions for both the SSE regularization parameter and the Lagrangian, akin to the relationship used for SSE-RDO. Experiments with images and videos show bitrate savings of more than 30\% over SSE-RDO using the target NRM, with no decoder complexity overhead and minimal encoder complexity increase.
Large language models (LLMs) are considered valuable Intellectual Properties (IP) for legitimate owners due to the enormous computational cost of training. It is crucial to protect the IP of LLMs from malicious stealing or unauthorized deployment. Despite existing efforts in watermarking and fingerprinting LLMs, these methods either impact the text generation process or are limited in white-box access to the suspect model, making them impractical. Hence, we propose DuFFin, a novel $\textbf{Du}$al-Level $\textbf{Fin}$gerprinting $\textbf{F}$ramework for black-box setting ownership verification. DuFFin extracts the trigger pattern and the knowledge-level fingerprints to identify the source of a suspect model. We conduct experiments on a variety of models collected from the open-source website, including four popular base models as protected LLMs and their fine-tuning, quantization, and safety alignment versions, which are released by large companies, start-ups, and individual users. Results show that our method can accurately verify the copyright of the base protected LLM on their model variants, achieving the IP-ROC metric greater than 0.95. Our code is available at https://github.com/yuliangyan0807/llm-fingerprint.
Large language models (LLMs) have made exciting achievements across various domains, yet their deployment on resource-constrained personal devices remains hindered by the prohibitive computational and memory demands of task-specific fine-tuning. While quantization offers a pathway to efficiency, existing methods struggle to balance performance and overhead, either incurring high computational/memory costs or failing to address activation outliers, a critical bottleneck in quantized fine-tuning. To address these challenges, we propose the Outlier Spatial Stability Hypothesis (OSSH): During fine-tuning, certain activation outlier channels retain stable spatial positions across training iterations. Building on OSSH, we propose Quaff, a Quantized parameter-efficient fine-tuning framework for LLMs, optimizing low-precision activation representations through targeted momentum scaling. Quaff dynamically suppresses outliers exclusively in invariant channels using lightweight operations, eliminating full-precision weight storage and global rescaling while reducing quantization errors. Extensive experiments across ten benchmarks validate OSSH and demonstrate Quaff's efficacy. Specifically, on the GPQA reasoning benchmark, Quaff achieves a 1.73x latency reduction and 30% memory savings over full-precision fine-tuning while improving accuracy by 0.6% on the Phi-3 model, reconciling the triple trade-off between efficiency, performance, and deployability. By enabling consumer-grade GPU fine-tuning (e.g., RTX 2080 Super) without sacrificing model utility, Quaff democratizes personalized LLM deployment. The code is available at https://github.com/Little0o0/Quaff.git.
Spoken Language Understanding (SLU) systems must balance performance and efficiency, particularly in resource-constrained environments. Existing methods apply distillation and quantization separately, leading to suboptimal compression as distillation ignores quantization constraints. We propose QUADS, a unified framework that optimizes both through multi-stage training with a pre-tuned model, enhancing adaptability to low-bit regimes while maintaining accuracy. QUADS achieves 71.13\% accuracy on SLURP and 99.20\% on FSC, with only minor degradations of up to 5.56\% compared to state-of-the-art models. Additionally, it reduces computational complexity by 60--73$\times$ (GMACs) and model size by 83--700$\times$, demonstrating strong robustness under extreme quantization. These results establish QUADS as a highly efficient solution for real-world, resource-constrained SLU applications.
Quantization in SSL speech models (e.g., HuBERT) improves compression and performance in tasks like language modeling, resynthesis, and text-to-speech but often discards prosodic and paralinguistic information (e.g., emotion, prominence). While increasing codebook size mitigates some loss, it inefficiently raises bitrates. We propose Segmentation-Variant Codebooks (SVCs), which quantize speech at distinct linguistic units (frame, phone, word, utterance), factorizing it into multiple streams of segment-specific discrete features. Our results show that SVCs are significantly more effective at preserving prosodic and paralinguistic information across probing tasks. Additionally, we find that pooling before rather than after discretization better retains segment-level information. Resynthesis experiments further confirm improved style realization and slightly improved quality while preserving intelligibility.
The increasing deployment of Large Language Models (LLMs) on edge devices, driven by model advancements and hardware improvements, offers significant privacy benefits. However, these on-device LLMs inherently face performance limitations due to reduced model capacity and necessary compression techniques. To address this, we introduce a systematic methodology -- encompassing model capability, development efficiency, and system resources -- for evaluating on-device LLMs. Our comprehensive evaluation, encompassing models from 0.5B to 14B parameters and seven post-training quantization (PTQ) methods on commodity laptops, yields several critical insights: 1) System-level metrics exhibit near-linear scaling with effective bits-per-weight (BPW). 2) A practical threshold exists around $\sim$3.5 effective BPW, larger models subjected to low-bit quantization consistently outperform smaller models utilizing higher bit-precision. 3) Quantization with low BPW incurs marginal accuracy loss but significant memory savings. 4) Determined by low-level implementation specifics power consumption on CPU, where computation-intensive operations spend more power than memory-intensive ones. These findings offer crucial insights and practical guidelines for the efficient deployment and optimized configuration of LLMs on resource-constrained edge devices. Our codebase is available at https://github.com/simmonssong/LLMOnDevice.
Integer quantization has emerged as a critical technique to facilitate deployment on resource-constrained devices. Although they do reduce the complexity of the learning models, their inference performance is often prone to quantization-induced errors. To this end, we introduce InTreeger: an end-to-end framework that takes a training dataset as input, and outputs an architecture-agnostic integer-only C implementation of tree-based machine learning model, without loss of precision. This framework enables anyone, even those without prior experience in machine learning, to generate a highly optimized integer-only classification model that can run on any hardware simply by providing an input dataset and target variable. We evaluated our generated implementations across three different architectures (ARM, x86, and RISC-V), resulting in significant improvements in inference latency. In addition, we show the energy efficiency compared to typical decision tree implementations that rely on floating-point arithmetic. The results underscore the advantages of integer-only inference, making it particularly suitable for energy- and area-constrained devices such as embedded systems and edge computing platforms, while also enabling the execution of decision trees on existing ultra-low power devices.
Federated Learning (FL) is a decentralized model training approach that preserves data privacy but struggles with low efficiency. Quantization, a powerful training optimization technique, has been widely explored for integration into FL. However, many studies fail to consider the distinct performance attribution between particular quantization strategies, such as post-training quantization (PTQ) or quantization-aware training (QAT). As a result, existing FL quantization methods rely solely on either PTQ or QAT, optimizing for speed or accuracy while compromising the other. To efficiently accelerate FL and maintain distributed convergence accuracy across various FL settings, this paper proposes a hybrid quantitation approach combining PTQ and QAT for FL systems. We conduct case studies to validate the effectiveness of using hybrid quantization in FL. To solve the difficulty of modeling speed and accuracy caused by device and data heterogeneity, we propose a hardware-related analysis and data-distribution-related analysis to help identify the trade-off boundaries for strategy selection. Based on these, we proposed a novel framework named FedHQ to automatically adopt optimal hybrid strategy allocation for FL systems. Specifically, FedHQ develops a coarse-grained global initialization and fine-grained ML-based adjustment to ensure efficiency and robustness. Experiments show that FedHQ achieves up to 2.47x times training acceleration and up to 11.15% accuracy improvement and negligible extra overhead.
Convolutional Neural Networks (CNNs) have gained high popularity as a tool for computer vision tasks and for that reason are used in various applications. There are many different concepts, like single shot detectors, that have been published for detecting objects in images or video streams. However, CNNs suffer from disadvantages regarding the deployment on embedded platforms such as re-configurable hardware like Field Programmable Gate Arrays (FPGAs). Due to the high computational intensity, memory requirements and arithmetic conditions, a variety of strategies for running CNNs on FPGAs have been developed. The following methods showcase our best practice approaches for a TinyYOLOv3 detector network on a XILINX Artix-7 FPGA using techniques like fusion of batch normalization, filter pruning and post training network quantization.
QLoRA effectively combines low-bit quantization and LoRA to achieve memory-friendly fine-tuning for large language models (LLM). Recently, methods based on SVD for continuous update iterations to initialize LoRA matrices to accommodate quantization errors have generally failed to consistently improve performance. Dynamic mixed precision is a natural idea for continuously improving the fine-tuning performance of quantized models, but previous methods often optimize low-rank subspaces or quantization components separately, without considering their synergy. To address this, we propose \textbf{QR-Adaptor}, a unified, gradient-free strategy that uses partial calibration data to jointly search the quantization components and the rank of low-rank spaces for each layer, thereby continuously improving model performance. QR-Adaptor does not minimize quantization error but treats precision and rank allocation as a discrete optimization problem guided by actual downstream performance and memory usage. Compared to state-of-the-art (SOTA) quantized LoRA fine-tuning methods, our approach achieves a 4.89\% accuracy improvement on GSM8K, and in some cases even outperforms the 16-bit fine-tuned model while maintaining the memory footprint of the 4-bit setting.