Large language models (LLMs) with billions of parameters excel at predicting the next token in a sequence. Recent work computes non-vacuous compression-based generalization bounds for LLMs, but these bounds are vacuous for large models at the billion-parameter scale. Moreover, these bounds are obtained through restrictive compression techniques, bounding compressed models that generate low-quality text. Additionally, the tightness of these existing bounds depends on the number of IID documents in a training set rather than the much larger number of non-IID constituent tokens, leaving untapped potential for tighter bounds. In this work, we instead use properties of martingales to derive generalization bounds that benefit from the vast number of tokens in LLM training sets. Since a dataset contains far more tokens than documents, our generalization bounds not only tolerate but actually benefit from far less restrictive compression schemes. With Monarch matrices, Kronecker factorizations, and post-training quantization, we achieve non-vacuous generalization bounds for LLMs as large as LLaMA2-70B. Unlike previous approaches, our work achieves the first non-vacuous bounds for models that are deployed in practice and generate high-quality text.
Vision transformers (ViTs) have demonstrated their superior accuracy for computer vision tasks compared to convolutional neural networks (CNNs). However, ViT models are often computation-intensive for efficient deployment on resource-limited edge devices. This work proposes Quasar-ViT, a hardware-oriented quantization-aware architecture search framework for ViTs, to design efficient ViT models for hardware implementation while preserving the accuracy. First, Quasar-ViT trains a supernet using our row-wise flexible mixed-precision quantization scheme, mixed-precision weight entanglement, and supernet layer scaling techniques. Then, it applies an efficient hardware-oriented search algorithm, integrated with hardware latency and resource modeling, to determine a series of optimal subnets from supernet under different inference latency targets. Finally, we propose a series of model-adaptive designs on the FPGA platform to support the architecture search and mitigate the gap between the theoretical computation reduction and the practical inference speedup. Our searched models achieve 101.5, 159.6, and 251.6 frames-per-second (FPS) inference speed on the AMD/Xilinx ZCU102 FPGA with 80.4%, 78.6%, and 74.9% top-1 accuracy, respectively, for the ImageNet dataset, consistently outperforming prior works.
This thesis investigates the application of state-of-the-art advances in generative neural networks for fast simulation of the Zero Degree Calorimeter (ZDC) neutron detector in the ALICE experiment at CERN. Traditional simulation methods using the GEANT Monte Carlo toolkit, while accurate, are computationally demanding. With increasing computational needs at CERN, efficient simulation techniques are essential. The thesis provides a comprehensive literature review on the application of neural networks in computer vision, fast simulations using machine learning, and generative neural networks in high-energy physics. The theory of the analyzed models is also discussed, along with technical aspects and the challenges associated with a practical implementation. The experiments evaluate various neural network architectures, including convolutional neural networks, vision transformers, and MLP-Mixers, as well as generative frameworks such as autoencoders, generative adversarial networks, vector quantization models, and diffusion models. Key contributions include the implementation and evaluation of these models, a significant improvement in the Wasserstein metric compared to existing methods with a low generation time of 5 milliseconds per sample, and the formulation of a list of recommendations for developing models for fast ZDC simulation. Open-source code and detailed hyperparameter settings are provided for reproducibility. Additionally, the thesis outlines future research directions to further enhance simulation fidelity and efficiency.
Representing and quantifying Minimal Residual Disease (MRD) in Acute Myeloid Leukemia (AML), a type of cancer that affects the blood and bone marrow, is essential in the prognosis and follow-up of AML patients. As traditional cytological analysis cannot detect leukemia cells below 5\%, the analysis of flow cytometry dataset is expected to provide more reliable results. In this paper, we explore statistical learning methods based on optimal transport (OT) to achieve a relevant low-dimensional representation of multi-patient flow cytometry measurements (FCM) datasets considered as high-dimensional probability distributions. Using the framework of OT, we justify the use of the K-means algorithm for dimensionality reduction of multiple large-scale point clouds through mean measure quantization by merging all the data into a single point cloud. After this quantization step, the visualization of the intra and inter-patients FCM variability is carried out by embedding low-dimensional quantized probability measures into a linear space using either Wasserstein Principal Component Analysis (PCA) through linearized OT or log-ratio PCA of compositional data. Using a publicly available FCM dataset and a FCM dataset from Bordeaux University Hospital, we demonstrate the benefits of our approach over the popular kernel mean embedding technique for statistical learning from multiple high-dimensional probability distributions. We also highlight the usefulness of our methodology for low-dimensional projection and clustering patient measurements according to their level of MRD in AML from FCM. In particular, our OT-based approach allows a relevant and informative two-dimensional representation of the results of the FlowSom algorithm, a state-of-the-art method for the detection of MRD in AML using multi-patient FCM.
Large Language Models (LLMs) have demonstrated impressive performance across various domains. However, the enormous number of model parameters makes fine-tuning challenging, significantly limiting their application and deployment. Existing solutions combine parameter quantization with Low-Rank Adaptation (LoRA), greatly reducing memory usage but resulting in noticeable performance degradation. In this paper, we identify an imbalance in fine-tuning quantized pre-trained models: overly complex adapter inputs and outputs versus low effective trainability of the adaptation. We propose Quantized LLMs with Balanced-rank Adaptation (Q-BaRA), which simplifies the adapter inputs and outputs while increasing the adapter's rank to achieve a more suitable balance for fine-tuning quantized LLMs. Additionally, for scenarios where fine-tuned LLMs need to be deployed as low-precision inference models, we introduce Quantization-Aware Fine-tuning with Higher Rank Adaptation (QA-HiRA), which simplifies the adapter inputs and outputs to align with the pre-trained model's block-wise quantization while employing a single matrix to achieve a higher rank. Both Q-BaRA and QA-HiRA are easily implemented and offer the following optimizations: (i) Q-BaRA consistently achieves the highest accuracy compared to baselines and other variants, requiring the same number of trainable parameters and computational effort; (ii) QA-HiRA naturally merges adapter parameters into the block-wise quantized model after fine-tuning, achieving the highest accuracy compared to other methods. We apply our Q-BaRA and QA-HiRA to the LLaMA and LLaMA2 model families and validate their effectiveness across different fine-tuning datasets and downstream scenarios. Code will be made available at \href{https://github.com/xiaocaigou/qbaraqahira}{https://github.com/xiaocaigou/qbaraqahira}
Deep learning models have achieved tremendous success in most of the industries in recent years. The evolution of these models has also led to an increase in the model size and energy requirement, making it difficult to deploy in production on low compute devices. An increase in the number of connected devices around the world warrants compressed models that can be easily deployed at the local devices with low compute capacity and power accessibility. A wide range of solutions have been proposed by different researchers to reduce the size and complexity of such models, prominent among them are, Weight Quantization, Parameter Pruning, Network Pruning, low-rank representation, weights sharing, neural architecture search, knowledge distillation etc. In this research work, we investigate the performance impacts on various trained deep learning models, compressed using quantization and pruning techniques. We implemented both, quantization and pruning, compression techniques on popular deep learning models used in the image classification, object detection, language models and generative models-based problem statements. We also explored performance of various large language models (LLMs) after quantization and low rank adaptation. We used the standard evaluation metrics (model's size, accuracy, and inference time) for all the related problem statements and concluded this paper by discussing the challenges and future work.
By quantizing network weights and activations to low bitwidth, we can obtain hardware-friendly and energy-efficient networks. However, existing quantization techniques utilizing the straight-through estimator and piecewise constant functions face the issue of how to represent originally high-bit input data with low-bit values. To fully quantize deep neural networks, we propose pixel embedding, which replaces each float-valued input pixel with a vector of quantized values by using a lookup table. The lookup table or low-bit representation of pixels is differentiable and trainable by backpropagation. Such replacement of inputs with vectors is similar to word embedding in the natural language processing field. Experiments on ImageNet and CIFAR-100 show that pixel embedding reduces the top-5 error gap caused by quantizing the floating points at the first layer to only 1% for the ImageNet dataset, and the top-1 error gap caused by quantizing first and last layers to slightly over 1% for the CIFAR-100 dataset. The usefulness of pixel embedding is further demonstrated by inference time measurements, which demonstrate over 1.7 times speedup compared to floating point precision first layer.
Large language models (LLMs) are crucial in modern natural language processing and artificial intelligence. However, they face challenges in managing their significant memory requirements. Although quantization-aware training (QAT) offers a solution by reducing memory consumption through low-bit representations with minimal accuracy loss, it is impractical due to substantial training resources. To address this, we propose Efficient Quantization-Aware Training (EfficientQAT), a more feasible QAT algorithm. EfficientQAT involves two consecutive phases: Block-wise training of all parameters (Block-AP) and end-to-end training of quantization parameters (E2E-QP). To the best of our knowledge, Block-AP is the first method to enable direct training of all parameters in a block-wise manner, reducing accuracy loss in low-bit scenarios by enhancing the solution space during optimization. E2E-QP then trains only the quantization parameters (step sizes) end-to-end, further improving the performance of quantized models by considering interactions among all sub-modules. Extensive experiments demonstrate that EfficientQAT outperforms previous quantization methods across a range of models, including base LLMs, instruction-tuned LLMs, and multimodal LLMs, with scales from 7B to 70B parameters at various quantization bits. For instance, EfficientQAT obtains a 2-bit Llama-2-70B model on a single A100-80GB GPU in 41 hours, with less than 3 points accuracy degradation compared to the full precision (69.48 vs. 72.41). Code is available at https://github.com/OpenGVLab/EfficientQAT.
Post-Training Quantization (PTQ) has received significant attention because it requires only a small set of calibration data to quantize a full-precision model, which is more practical in real-world applications in which full access to a large training set is not available. However, it often leads to overfitting on the small calibration dataset. Several methods have been proposed to address this issue, yet they still rely on only the calibration set for the quantization and they do not validate the quantized model due to the lack of a validation set. In this work, we propose a novel meta-learning based approach to enhance the performance of post-training quantization. Specifically, to mitigate the overfitting problem, instead of only training the quantized model using the original calibration set without any validation during the learning process as in previous PTQ works, in our approach, we both train and validate the quantized model using two different sets of images. In particular, we propose a meta-learning based approach to jointly optimize a transformation network and a quantized model through bi-level optimization. The transformation network modifies the original calibration data and the modified data will be used as the training set to learn the quantized model with the objective that the quantized model achieves a good performance on the original calibration data. Extensive experiments on the widely used ImageNet dataset with different neural network architectures demonstrate that our approach outperforms the state-of-the-art PTQ methods.
Camera relocalization relies on 3D models of the scene with a large memory footprint that is incompatible with the memory budget of several applications. One solution to reduce the scene memory size is map compression by removing certain 3D points and descriptor quantization. This achieves high compression but leads to performance drop due to information loss. To address the memory performance trade-off, we train a light-weight scene-specific auto-encoder network that performs descriptor quantization-dequantization in an end-to-end differentiable manner updating both product quantization centroids and network parameters through back-propagation. In addition to optimizing the network for descriptor reconstruction, we encourage it to preserve the descriptor-matching performance with margin-based metric loss functions. Results show that for a local descriptor memory of only 1MB, the synergistic combination of the proposed network and map compression achieves the best performance on the Aachen Day-Night compared to existing compression methods.
The quantization of large language models (LLMs) has been a prominent research area aimed at enabling their lightweight deployment in practice. Existing research about LLM's quantization has mainly explored the interplay between weights and activations, or employing auxiliary components while neglecting the necessity of adjusting weights during quantization. Consequently, original weight distributions frequently fail to yield desired results after round-to-nearest (RTN) quantization. Even though incorporating techniques such as mixed precision and low-rank error approximation in LLM's quantization can yield improved results, they inevitably introduce additional computational overhead. On the other hand, traditional techniques for weight quantization, such as Generative Post-Training Quantization, rely on manually tweaking weight distributions to minimize local errors, but they fall short of achieving globally optimal outcomes. Although the recently proposed Learnable Singular-value Increment improves global weight quantization by modifying weight distributions, it disrupts the original distribution considerably. This introduces pronounced bias toward the training data and can degrade downstream task performance. In this paper, we introduce Singular-value Diagonal Expansion, a more nuanced approach to refining weight distributions to achieve better quantization alignment. Furthermore, we introduce Cross-layer Learning that improves overall quantization outcomes by distributing errors more evenly across layers. Our plug-and-play weight-quantization methods demonstrate substantial performance improvements over state-of-the-art approaches, including OmniQuant, DuQuant, and PrefixQuant.
This paper addresses the problem of uplink transmit power optimization in distributed massive multiple-input multiple-output systems, where remote radio heads (RRHs) are equipped with 1-bit analog-to-digital converters (ADCs). First, in a scenario where a single RRH serves a single user equipment (UE), the signal-to-noise-and-distortion ratio (SNDR) is shown to be a non-monotonic and unimodal function of the UE transmit power due to the quantization distortion (QD). Upon the introduction of multiple RRHs, adding properly tuned dithering at each RRH is shown to render the SNDR at the output of the joint receiver unimodal. In a scenario with multiple RRHs and UEs, considering the non-monotonic nature of the signal-to-interference-plus-noise-and-distortion ratio (SINDR), both the UE transmit powers and the RRH dithering levels are jointly optimized subject to the min-power and max-min-SINDR criteria, while employing Bussgang-based maximum ratio combining (BMRC) and minimum mean squared error (BMMSE) receivers. To this end, gradient and block coordinate descent methods are introduced to tune the UE transmit powers, whereas a line search coupled with gradient updates is used to adjust the RRH dithering levels. Numerical results demonstrate that jointly optimizing the UE transmit power and the RRH dithering levels can significantly enhance the system performance, thus facilitating joint reception from multiple RRHs across a range of scenarios. Comparing the BMMSE and BMRC receivers, the former offers a better interference and QD alleviation while the latter has a lower computational complexity.
We introduce FedDM, a novel training framework designed for the federated training of diffusion models. Our theoretical analysis establishes the convergence of diffusion models when trained in a federated setting, presenting the specific conditions under which this convergence is guaranteed. We propose a suite of training algorithms that leverage the U-Net architecture as the backbone for our diffusion models. These include a basic Federated Averaging variant, FedDM-vanilla, FedDM-prox to handle data heterogeneity among clients, and FedDM-quant, which incorporates a quantization module to reduce the model update size, thereby enhancing communication efficiency across the federated network. We evaluate our algorithms on FashionMNIST (28x28 resolution), CIFAR-10 (32x32 resolution), and CelebA (64x64 resolution) for DDPMs, as well as LSUN Church Outdoors (256x256 resolution) for LDMs, focusing exclusively on the imaging modality. Our evaluation results demonstrate that FedDM algorithms maintain high generation quality across image resolutions. At the same time, the use of quantized updates and proximal terms in the local training objective significantly enhances communication efficiency (up to 4x) and model convergence, particularly in non-IID data settings, at the cost of increased FID scores (up to 1.75x).
Recent advancements in quantization and mixed-precision approaches offers substantial opportunities to improve the speed and energy efficiency of Neural Networks (NN). Research has shown that individual parameters with varying low precision, can attain accuracies comparable to full-precision counterparts. However, modern embedded microprocessors provide very limited support for mixed-precision NNs regarding both Instruction Set Architecture (ISA) extensions and their hardware design for efficient execution of mixed-precision operations, i.e., introducing several performance bottlenecks due to numerous instructions for data packing and unpacking, arithmetic unit under-utilizations etc. In this work, we bring together, for the first time, ISA extensions tailored to mixed-precision hardware optimizations, targeting energy-efficient DNN inference on leading RISC-V CPU architectures. To this end, we introduce a hardware-software co-design framework that enables cooperative hardware design, mixed-precision quantization, ISA extensions and inference in cycle-accurate emulations. At hardware level, we firstly expand the ALU unit within our proof-of-concept micro-architecture to support configurable fine grained mixed-precision arithmetic operations. Subsequently, we implement multi-pumping to minimize execution latency, with an additional soft SIMD optimization applied for 2-bit operations. At the ISA level, three distinct MAC instructions are encoded extending the RISC-V ISA, and exposed up to the compiler level, each corresponding to a different mixed-precision operational mode. Our extensive experimental evaluation over widely used DNNs and datasets, such as CIFAR10 and ImageNet, demonstrates that our framework can achieve, on average, 15x energy reduction for less than 1% accuracy loss and outperforms the ISA-agnostic state-of-the-art RISC-V cores.
To fill the gap of traditional GS compression method, in this paper, we first propose a simple and effective GS data compression anchor called Graph-based GS Compression (GGSC). GGSC is inspired by graph signal processing theory and uses two branches to compress the primitive center and attributes. We split the whole GS sample via KDTree and clip the high-frequency components after the graph Fourier transform. Followed by quantization, G-PCC and adaptive arithmetic coding are used to compress the primitive center and attribute residual matrix to generate the bitrate file. GGSS is the first work to explore traditional GS compression, with advantages that can reveal the GS distortion characteristics corresponding to typical compression operation, such as high-frequency clipping and quantization. Second, based on GGSC, we create a GS Quality Assessment dataset (GSQA) with 120 samples. A subjective experiment is conducted in a laboratory environment to collect subjective scores after rendering GS into Processed Video Sequences (PVS). We analyze the characteristics of different GS distortions based on Mean Opinion Scores (MOS), demonstrating the sensitivity of different attributes distortion to visual quality. The GGSC code and the dataset, including GS samples, MOS, and PVS, are made publicly available at https://github.com/Qi-Yangsjtu/GGSC.
The increasing demand for deploying large Mixture-of-Experts (MoE) models in resource-constrained environments necessitates efficient approaches to address their high memory and computational requirements challenges. Moreover, given that tasks come in different user-defined constraints and the available resources change over time in multi-tenant environments, it is necessary to design an approach which provides a flexible configuration space. This paper presents an adaptive serving approach for the efficient deployment of MoE models, capitalizing on partial quantization of the experts. By dynamically determining the number of quantized experts and their distribution across CPU and GPU, our approach explores the Pareto frontier and offers a fine-grained range of configurations for tuning throughput and model quality. Our evaluation on an NVIDIA A100 GPU using a Mixtral 8x7B MoE model for three language modelling benchmarks demonstrates that the throughput of token generation can be adjusted from 0.63 to 13.00 token per second. This enhancement comes with a marginal perplexity increase of 3.81 to 4.00, 13.59 to 14.17, and 7.24 to 7.40 for WikiText2, PTB, and C4 datasets respectively under maximum quantization. These results highlight the practical applicability of our approach in dynamic and accuracy-sensitive applications where both memory usage and output quality are important.
Diffusion models have recently dominated image synthesis tasks. However, the iterative denoising process is expensive in computations at inference time, making diffusion models less practical for low-latency and scalable real-world applications. Post-training quantization (PTQ) of diffusion models can significantly reduce the model size and accelerate the sampling process without re-training. Nonetheless, applying existing PTQ methods directly to low-bit diffusion models can significantly impair the quality of generated samples. Specifically, for each denoising step, quantization noise leads to deviations in the estimated mean and mismatches with the predetermined variance schedule. As the sampling process proceeds, the quantization noise may accumulate, resulting in a low signal-to-noise ratio (SNR) during the later denoising steps. To address these challenges, we propose a unified formulation for the quantization noise and diffusion perturbed noise in the quantized denoising process. Specifically, we first disentangle the quantization noise into its correlated and residual uncorrelated parts regarding its full-precision counterpart. The correlated part can be easily corrected by estimating the correlation coefficient. For the uncorrelated part, we subtract the bias from the quantized results to correct the mean deviation and calibrate the denoising variance schedule to absorb the excess variance resulting from quantization. Moreover, we introduce a mixed-precision scheme for selecting the optimal bitwidth for each denoising step. Extensive experiments demonstrate that our method outperforms previous post-training quantized diffusion models, with only a 0.06 increase in FID score compared to full-precision LDM-4 on ImageNet 256x256, while saving 19.9x bit operations. Code is available at https://github.com/ziplab/PTQD.
The digital revolution and the advent of the world wide web have transformed human communication, notably through the emergence of memes. While memes are a popular and straightforward form of expression, they can also be used to spread misinformation and hate due to their anonymity and ease of use. In response to these challenges, this paper introduces a solution developed by team 'Baseline' for the AI Singapore Online Safety Prize Challenge. Focusing on computational efficiency and feature engineering, the solution achieved an AUROC of 0.76 and an accuracy of 0.69 on the test dataset. As key features, the solution leverages the inherent probabilistic capabilities of large Vision-Language Models (VLMs) to generate task-adapted feature encodings from text, and applies a distilled quantization tailored to the specific cultural nuances present in Singapore. This type of processing and fine-tuning can be adapted to various visual and textual understanding and classification tasks, and even applied on private VLMs such as OpenAI's GPT. Finally it can eliminate the need for extensive model training on large GPUs for resource constrained applications, also offering a solution when little or no data is available.
As the field of Large Language Models (LLMs) continues to evolve, the context length in inference is steadily growing. Key-Value Cache (KVCache), the intermediate representations of tokens within LLM inference, has now become the primary memory bottleneck due to limited GPU memory. Current methods selectively determine suitable keys and values for self-attention computation in LLMs to address the issue. However, they either fall short in maintaining model quality or result in high serving latency. Drawing inspiration from advanced embedding retrieval techniques prevalent in the data management community, we consider the storage and retrieval of KVCache as a typical embedding retrieval problem. We propose PQCache, which employs Product Quantization (PQ) to manage KVCache, maintaining model quality while ensuring low serving latency. During the prefilling phase, we apply PQ to tokens' keys for each LLM layer and head. During the autoregressive decoding phase, we use PQ codes and centroids to approximately identify important preceding tokens, then fetch the corresponding key-value pairs for self-attention computation. Through meticulous design of overlapping and caching, we minimize any additional computation and communication overhead during both phases. Extensive experiments demonstrate that PQCache achieves both effectiveness and efficiency, with 4.60% score improvement over existing methods on InfiniteBench and low system latency in both prefilling and decoding.
While Reconfigurable Intelligent Surfaces (RISs) constitute one of the most prominent enablers for the upcoming sixth Generation (6G) of wireless networks, the design of efficient RIS phase profiles remains a notorious challenge when large numbers of phase-quantized unit cells are involved, typically of a single bit, as implemented by a vast majority of existing metasurface prototypes. In this paper, we focus on the RIS phase configuration problem for the exemplary case of the Signal-to-Noise Ratio (SNR) maximization for an RIS-enabled single-input single-output system where the metasurface tunable elements admit a phase difference of $\pi$ radians. We present a novel closed-form configuration which serves as a lower bound guaranteeing at least half the SNR of the ideal continuous (upper bound) SNR gain, and whose mean performance is shown to be asymptotically optimal. The proposed sign alignment configuration can be further used as initialization to standard discrete optimization algorithms. A discussion on the reduced complexity hardware benefits via the presented configuration is also included. Our numerical results demonstrate the efficacy of the proposed RIS sign alignment scheme over iterative approaches as well as the commonplace continuous phase quantization treatment.