Vector quantization is a fundamental operation for data compression and vector search. To obtain high accuracy, multi-codebook methods represent each vector using codewords across several codebooks. Residual quantization (RQ) is one such method, which iteratively quantizes the error of the previous step. While the error distribution is dependent on previously-selected codewords, this dependency is not accounted for in conventional RQ as it uses a fixed codebook per quantization step. In this paper, we propose QINCo, a neural RQ variant that constructs specialized codebooks per step that depend on the approximation of the vector from previous steps. Experiments show that QINCo outperforms state-of-the-art methods by a large margin on several datasets and code sizes. For example, QINCo achieves better nearest-neighbor search accuracy using 12-byte codes than the state-of-the-art UNQ using 16 bytes on the BigANN1M and Deep1M datasets.
Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code is publicly available at https://github.com/usyd-fsalab/fp6_llm.
In disentangled representation learning, a model is asked to tease apart a dataset's underlying sources of variation and represent them independently of one another. Since the model is provided with no ground truth information about these sources, inductive biases take a paramount role in enabling disentanglement. In this work, we construct an inductive bias towards encoding to and decoding from an organized latent space. Concretely, we do this by (i) quantizing the latent space into discrete code vectors with a separate learnable scalar codebook per dimension and (ii) applying strong model regularization via an unusually high weight decay. Intuitively, the latent space design forces the encoder to combinatorially construct codes from a small number of distinct scalar values, which in turn enables the decoder to assign a consistent meaning to each value. Regularization then serves to drive the model towards this parsimonious strategy. We demonstrate the broad applicability of this approach by adding it to both basic data-reconstructing (vanilla autoencoder) and latent-reconstructing (InfoGAN) generative models. For reliable evaluation, we also propose InfoMEC, a new set of metrics for disentanglement that is cohesively grounded in information theory and fixes well-established shortcomings in previous metrics. Together with regularization, latent quantization dramatically improves the modularity and explicitness of learned representations on a representative suite of benchmark datasets. In particular, our quantized-latent autoencoder (QLAE) consistently outperforms strong methods from prior work in these key disentanglement properties without compromising data reconstruction.
Sleep Stage Classification (SSC) is a labor-intensive task, requiring experts to examine hours of electrophysiological recordings for manual classification. This is a limiting factor when it comes to leveraging sleep stages for therapeutic purposes. With increasing affordability and expansion of wearable devices, automating SSC may enable deployment of sleep-based therapies at scale. Deep Learning has gained increasing attention as a potential method to automate this process. Previous research has shown accuracy comparable to manual expert scores. However, previous approaches require sizable amount of memory and computational resources. This constrains the ability to classify in real time and deploy models on the edge. To address this gap, we aim to provide a model capable of predicting sleep stages in real-time, without requiring access to external computational sources (e.g., mobile phone, cloud). The algorithm is power efficient to enable use on embedded battery powered systems. Our compact sleep stage classifier can be deployed on most off-the-shelf microcontrollers (MCU) with constrained hardware settings. This is due to the memory footprint of our approach requiring significantly fewer operations. The model was tested on three publicly available data bases and achieved performance comparable to the state of the art, whilst reducing model complexity by orders of magnitude (up to 280 times smaller compared to state of the art). We further optimized the model with quantization of parameters to 8 bits with only an average drop of 0.95% in accuracy. When implemented in firmware, the quantized model achieves a latency of 1.6 seconds on an Arm CortexM4 processor, allowing its use for on-line SSC-based therapies.
Many common types of data can be represented as functions that map coordinates to signal values, such as pixel locations to RGB values in the case of an image. Based on this view, data can be compressed by overfitting a compact neural network to its functional representation and then encoding the network weights. However, most current solutions for this are inefficient, as quantization to low-bit precision substantially degrades the reconstruction quality. To address this issue, we propose overfitting variational Bayesian neural networks to the data and compressing an approximate posterior weight sample using relative entropy coding instead of quantizing and entropy coding it. This strategy enables direct optimization of the rate-distortion performance by minimizing the $\beta$-ELBO, and target different rate-distortion trade-offs for a given network architecture by adjusting $\beta$. Moreover, we introduce an iterative algorithm for learning prior weight distributions and employ a progressive refinement process for the variational posterior that significantly enhances performance. Experiments show that our method achieves strong performance on image and audio compression while retaining simplicity.
The development of deep learning algorithms has extensively empowered humanity's task automatization capacity. However, the huge improvement in the performance of these models is highly correlated with their increasing level of complexity, limiting their usefulness in human-oriented applications, which are usually deployed in resource-constrained devices. This led to the development of compression techniques that drastically reduce the computational and memory costs of deep learning models without significant performance degradation. This paper aims to systematize the current literature on this topic by presenting a comprehensive survey of model compression techniques in biometrics applications, namely quantization, knowledge distillation and pruning. We conduct a critical analysis of the comparative value of these techniques, focusing on their advantages and disadvantages and presenting suggestions for future work directions that can potentially improve the current methods. Additionally, we discuss and analyze the link between model bias and model compression, highlighting the need to direct compression research toward model fairness in future works.
Existing unsupervised deep product quantization methods primarily aim for the increased similarity between different views of the identical image, whereas the delicate multi-level semantic similarities preserved between images are overlooked. Moreover, these methods predominantly focus on the Euclidean space for computational convenience, compromising their ability to map the multi-level semantic relationships between images effectively. To mitigate these shortcomings, we propose a novel unsupervised product quantization method dubbed \textbf{Hi}erarchical \textbf{H}yperbolic \textbf{P}roduct \textbf{Q}uantization (HiHPQ), which learns quantized representations by incorporating hierarchical semantic similarity within hyperbolic geometry. Specifically, we propose a hyperbolic product quantizer, where the hyperbolic codebook attention mechanism and the quantized contrastive learning on the hyperbolic product manifold are introduced to expedite quantization. Furthermore, we propose a hierarchical semantics learning module, designed to enhance the distinction between similar and non-matching images for a query by utilizing the extracted hierarchical semantics as an additional training supervision. Experiments on benchmarks show that our proposed method outperforms state-of-the-art baselines.
We present Inferflow, an efficient and highly configurable inference engine for large language models (LLMs). With Inferflow, users can serve most of the common transformer models by simply modifying some lines in corresponding configuration files, without writing a single line of source code. Compared with most existing inference engines, Inferflow has some key features. First, by implementing a modular framework of atomic build-blocks and technologies, Inferflow is compositionally generalizable to new models. Second, 3.5-bit quantization is introduced in Inferflow as a tradeoff between 3-bit and 4-bit quantization. Third, hybrid model partitioning for multi-GPU inference is introduced in Inferflow to better balance inference speed and throughput than the existing partition-by-layer and partition-by-tensor strategies.
Large neural networks require enormous computational clusters of machines. Model-parallel training, when the model architecture is partitioned sequentially between workers, is a popular approach for training modern models. Information compression can be applied to decrease workers communication time, as it is often a bottleneck in such systems. This work explores how simultaneous compression of activations and gradients in model-parallel distributed training setup affects convergence. We analyze compression methods such as quantization and TopK compression, and also experiment with error compensation techniques. Moreover, we employ TopK with AQ-SGD per-batch error feedback approach. We conduct experiments on image classification and language model fine-tuning tasks. Our findings demonstrate that gradients require milder compression rates than activations. We observe that $K=10\%$ is the lowest TopK compression level, which does not harm model convergence severely. Experiments also show that models trained with TopK perform well only when compression is also applied during inference. We find that error feedback techniques do not improve model-parallel training compared to plain compression, but allow model inference without compression with almost no quality drop. Finally, when applied with the AQ-SGD approach, TopK stronger than with $ K=30\%$ worsens model performance significantly.
Early wildfire detection in remote and forest areas is crucial for minimizing devastation and preserving ecosystems. Autonomous drones offer agile access to remote, challenging terrains, equipped with advanced imaging technology that delivers both high-temporal and detailed spatial resolution, making them valuable assets in the early detection and monitoring of wildfires. However, the limited computation and battery resources of Unmanned Aerial Vehicles (UAVs) pose significant challenges in implementing robust and efficient image classification models. Current works in this domain often operate offline, emphasizing the need for solutions that can perform inference in real time, given the constraints of UAVs. To address these challenges, this paper aims to develop a real-time image classification and fire segmentation model. It presents a comprehensive investigation into hardware acceleration using the Jetson Nano P3450 and the implications of TensorRT, NVIDIA's high-performance deep-learning inference library, on fire classification accuracy and speed. The study includes implementations of Quantization Aware Training (QAT), Automatic Mixed Precision (AMP), and post-training mechanisms, comparing them against the latest baselines for fire segmentation and classification. All experiments utilize the FLAME dataset - an image dataset collected by low-altitude drones during a prescribed forest fire. This work contributes to the ongoing efforts to enable real-time, on-board wildfire detection capabilities for UAVs, addressing speed and the computational and energy constraints of these crucial monitoring systems. The results show a 13% increase in classification speed compared to similar models without hardware optimization. Comparatively, loss and accuracy are within 1.225% of the original values.
The emergence of accurate open large language models (LLMs) has led to a race towards performant quantization techniques which can enable their execution on end-user devices. In this paper, we revisit the problem of "extreme" LLM compression-defined as targeting extremely low bit counts, such as 2 to 3 bits per parameter-from the point of view of classic methods in Multi-Codebook Quantization (MCQ). Our algorithm, called AQLM, generalizes the classic Additive Quantization (AQ) approach for information retrieval to advance the state-of-the-art in LLM compression, via two innovations: 1) learned additive quantization of weight matrices in input-adaptive fashion, and 2) joint optimization of codebook parameters across each transformer blocks. Broadly, AQLM is the first scheme that is Pareto optimal in terms of accuracy-vs-model-size when compressing to less than 3 bits per parameter, and significantly improves upon all known schemes in the extreme compression (2bit) regime. In addition, AQLM is practical: we provide fast GPU and CPU implementations of AQLM for token generation, which enable us to match or outperform optimized FP16 implementations for speed, while executing in a much smaller memory footprint.
Diffusion models have achieved great success in image generation tasks through iterative noise estimation. However, the heavy denoising process and complex neural networks hinder their low-latency applications in real-world scenarios. Quantization can effectively reduce model complexity, and post-training quantization (PTQ), which does not require fine-tuning, is highly promising for compressing and accelerating diffusion models. Unfortunately, we find that due to the highly dynamic distribution of activations in different denoising steps, existing PTQ methods for diffusion models suffer from distribution mismatch issues at both calibration sample level and reconstruction output level, which makes the performance far from satisfactory, especially in low-bit cases. In this paper, we propose Enhanced Distribution Alignment for Post-Training Quantization of Diffusion Models (EDA-DM) to address the above issues. Specifically, at the calibration sample level, we select calibration samples based on the density and variety in the latent space, thus facilitating the alignment of their distribution with the overall samples; and at the reconstruction output level, we modify the loss of block reconstruction with the losses of layers, aligning the outputs of quantized model and full-precision model at different network granularity. Extensive experiments demonstrate that EDA-DM significantly outperforms the existing PTQ methods across various models (DDIM, LDM-4, LDM-8, Stable-Diffusion) and different datasets (CIFAR-10, LSUN-Bedroom, LSUN-Church, ImageNet, MS-COCO).
We investigate parameter-efficient fine-tuning (PEFT) methods that can provide good accuracy under limited computational and memory budgets in the context of large language models (LLMs). We present a new PEFT method called Robust Adaptation (RoSA) inspired by robust principal component analysis that jointly trains $\textit{low-rank}$ and $\textit{highly-sparse}$ components on top of a set of fixed pretrained weights to efficiently approximate the performance of a full-fine-tuning (FFT) solution. Across a series of challenging generative tasks such as grade-school math and SQL query generation, which require fine-tuning for good performance, we show that RoSA outperforms LoRA, pure sparse fine-tuning, and alternative hybrid methods at the same parameter budget, and can even recover the performance of FFT on some tasks. We provide system support for RoSA to complement the training algorithm, specifically in the form of sparse GPU kernels which enable memory- and computationally-efficient training, and show that it is also compatible with low-precision base weights, resulting in the first joint representation combining quantization, low-rank and sparse approximations. Our code is available at https://github.com/IST-DASLab/RoSA.
Recently, high-fidelity scene reconstruction with an optimized 3D Gaussian splat representation has been introduced for novel view synthesis from sparse image sets. Making such representations suitable for applications like network streaming and rendering on low-power devices requires significantly reduced memory consumption as well as improved rendering efficiency. We propose a compressed 3D Gaussian splat representation that utilizes sensitivity-aware vector clustering with quantization-aware training to compress directional colors and Gaussian parameters. The learned codebooks have low bitrates and achieve a compression rate of up to $31\times$ on real-world scenes with only minimal degradation of visual quality. We demonstrate that the compressed splat representation can be efficiently rendered with hardware rasterization on lightweight GPUs at up to $4\times$ higher framerates than reported via an optimized GPU compute pipeline. Extensive experiments across multiple datasets demonstrate the robustness and rendering speed of the proposed approach.
Quantization is of significance for compressing the over-parameterized deep neural models and deploying them on resource-limited devices. Fixed-precision quantization suffers from performance drop due to the limited numerical representation ability. Conversely, mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. In this paper, we devise a one-shot training-searching paradigm for mixed-precision model compression. Specifically, in the first stage, all potential bit-width configurations are coupled and thus optimized simultaneously within a set of shared weights. However, our observations reveal a previously unseen and severe bit-width interference phenomenon among highly coupled weights during optimization, leading to considerable performance degradation under a high compression ratio. To tackle this problem, we first design a bit-width scheduler to dynamically freeze the most turbulent bit-width of layers during training, to ensure the rest bit-widths converged properly. Then, taking inspiration from information theory, we present an information distortion mitigation technique to align the behavior of the bad-performing bit-widths to the well-performing ones. In the second stage, an inference-only greedy search scheme is devised to evaluate the goodness of configurations without introducing any additional training costs. Extensive experiments on three representative models and three datasets demonstrate the effectiveness of the proposed method. Code can be available on \href{https://www.github.com/1hunters/retraining-free-quantization}{https://github.com/1hunters/retraining-free-quantization}.
We present a framework for generating full-bodied photorealistic avatars that gesture according to the conversational dynamics of a dyadic interaction. Given speech audio, we output multiple possibilities of gestural motion for an individual, including face, body, and hands. The key behind our method is in combining the benefits of sample diversity from vector quantization with the high-frequency details obtained through diffusion to generate more dynamic, expressive motion. We visualize the generated motion using highly photorealistic avatars that can express crucial nuances in gestures (e.g. sneers and smirks). To facilitate this line of research, we introduce a first-of-its-kind multi-view conversational dataset that allows for photorealistic reconstruction. Experiments show our model generates appropriate and diverse gestures, outperforming both diffusion- and VQ-only methods. Furthermore, our perceptual evaluation highlights the importance of photorealism (vs. meshes) in accurately assessing subtle motion details in conversational gestures. Code and dataset available online.
Vector quantization-based image semantic communication systems have successfully boosted transmission efficiency, but face challenges with conflicting requirements between codebook design and digital constellation modulation. Traditional codebooks need wide index ranges, while modulation favors few discrete states. To address this, we propose a multilevel generative semantic communication system with a two-stage training framework. In the first stage, we train a high-quality codebook, using a multi-head octonary codebook (MOC) to compress the index range. In addition, a residual vector quantization (RVQ) mechanism is also integrated for effective multilevel communication. In the second stage, a noise reduction block (NRB) based on Swin Transformer is introduced, coupled with the multilevel codebook from the first stage, serving as a high-quality semantic knowledge base (SKB) for generative feature restoration. Finally, to simulate modern image transmission scenarios, we employ a diverse collection of high-resolution 2K images as the test set. The experimental results consistently demonstrate the superior performance of MOC-RVQ over conventional methods such as BPG or JPEG. Additionally, MOC-RVQ achieves comparable performance to an analog JSCC scheme, while needing only one-sixth of the channel bandwidth ratio (CBR) and being directly compatible with digital transmission systems.
Modern deep learning models have the ability to generate high-dimensional vectors whose similarity reflects semantic resemblance. Thus, similarity search, i.e., the operation of retrieving those vectors in a large collection that are similar to a given query, has become a critical component of a wide range of applications that demand highly accurate and timely answers. In this setting, the high vector dimensionality puts similarity search systems under compute and memory pressure, leading to subpar performance. Additionally, cross-modal retrieval tasks have become increasingly common, e.g., where a user inputs a text query to find the most relevant images for that query. However, these queries often have different distributions than the database embeddings, making it challenging to achieve high accuracy. In this work, we present LeanVec, a framework that combines linear dimensionality reduction with vector quantization to accelerate similarity search on high-dimensional vectors while maintaining accuracy. We present LeanVec variants for in-distribution (ID) and out-of-distribution (OOD) queries. LeanVec-ID yields accuracies on par with those from recently introduced deep learning alternatives whose computational overhead precludes their usage in practice. LeanVec-OOD uses two novel techniques for dimensionality reduction that consider the query and database distributions to simultaneously boost the accuracy and the performance of the framework even further (even presenting competitive results when the query and database distributions match). All in all, our extensive and varied experimental results show that LeanVec produces state-of-the-art results, with up to 3.7x improvement in search throughput and up to 4.9x faster index build time over the state of the art.
In recent years, multimodal large language models (MLLMs) such as GPT-4V have demonstrated remarkable advancements, excelling in a variety of vision-language tasks. Despite their prowess, the closed-source nature and computational demands of such models limit their accessibility and applicability. This study introduces TinyGPT-V, a novel open-source MLLM, designed for efficient training and inference across various vision-language tasks, including image captioning (IC) and visual question answering (VQA). Leveraging a compact yet powerful architecture, TinyGPT-V integrates the Phi-2 language model with pre-trained vision encoders, utilizing a unique mapping module for visual and linguistic information fusion. With a training regimen optimized for small backbones and employing a diverse dataset amalgam, TinyGPT-V requires significantly lower computational resources 24GB for training and as little as 8GB for inference without compromising on performance. Our experiments demonstrate that TinyGPT-V, with its language model 2.8 billion parameters, achieves comparable results in VQA and image inference tasks to its larger counterparts while being uniquely suited for deployment on resource-constrained devices through innovative quantization techniques. This work not only paves the way for more accessible and efficient MLLMs but also underscores the potential of smaller, optimized models in bridging the gap between high performance and computational efficiency in real-world applications. Additionally, this paper introduces a new approach to multimodal large language models using smaller backbones. Our code and training weights are available in the supplementary material.
With the widespread adoption of Large Language Models (LLMs), many deep learning practitioners are looking for strategies of running these models more efficiently. One such strategy is to use sparse Mixture-of-Experts (MoE) - a type of model architectures where only a fraction of model layers are active for any given input. This property allows MoE-based language models to generate tokens faster than their dense counterparts, but it also increases model size due to having multiple experts. Unfortunately, this makes state-of-the-art MoE language models difficult to run without high-end GPUs. In this work, we study the problem of running large MoE language models on consumer hardware with limited accelerator memory. We build upon parameter offloading algorithms and propose a novel strategy that accelerates offloading by taking advantage of innate properties of MoE LLMs. Using this strategy, we build can run Mixtral-8x7B with mixed quantization on desktop hardware and free-tier Google Colab instances.