In the era of large-scale language models, the substantial parameter size poses significant challenges for deployment. Being a prevalent compression technique, quantization has emerged as the mainstream practice to tackle this issue, which is mainly centered on two recipes W8A8 and W4A16 (i.e. weights and activations in such bit widths). In this study, we propose a novel W4A8 post-training quantization method for the available open-sourced LLMs, which combines the advantages of both two recipes. Therefore, we can leverage the benefit in the I/O utilization of 4-bit weight quantization and the acceleration due to 8-bit matrix computation. Nevertheless, the W4A8 faces notorious performance degradation. As a remedy, we involve layerwise activation quantization strategies which feature a novel logarithmic equalization for most intractable layers, and we combine them with fine-grained weight quantization. Without whistles and bells, we eliminate the necessity for further fine-tuning and obtain the state-of-the-art W4A8 quantized performance on BLOOM, LLaMA, and LLaMA-2 on standard benchmarks. We confirm that the W4A8 quantization is achievable for the deployment of large language models, fostering their wide-spreading real-world applications.
Generative retrieval (GR) directly predicts the identifiers of relevant documents (i.e., docids) based on a parametric model. It has achieved solid performance on many ad-hoc retrieval tasks. So far, these tasks have assumed a static document collection. In many practical scenarios, however, document collections are dynamic, where new documents are continuously added to the corpus. The ability to incrementally index new documents while preserving the ability to answer queries with both previously and newly indexed relevant documents is vital to applying GR models. In this paper, we address this practical continual learning problem for GR. We put forward a novel Continual-LEarner for generatiVE Retrieval (CLEVER) model and make two major contributions to continual learning for GR: (i) To encode new documents into docids with low computational cost, we present Incremental Product Quantization, which updates a partial quantization codebook according to two adaptive thresholds; and (ii) To memorize new documents for querying without forgetting previous knowledge, we propose a memory-augmented learning mechanism, to form meaningful connections between old and new documents. Empirical results demonstrate the effectiveness and efficiency of the proposed model.
Graph Neural Network (GNN) training and inference involve significant challenges of scalability with respect to both model sizes and number of layers, resulting in degradation of efficiency and accuracy for large and deep GNNs. We present an end-to-end solution that aims to address these challenges for efficient GNNs in resource constrained environments while avoiding the oversmoothing problem in deep GNNs. We introduce a quantization based approach for all stages of GNNs, from message passing in training to node classification, compressing the model and enabling efficient processing. The proposed GNN quantizer learns quantization ranges and reduces the model size with comparable accuracy even under low-bit quantization. To scale with the number of layers, we devise a message propagation mechanism in training that controls layer-wise changes of similarities between neighboring nodes. This objective is incorporated into a Lagrangian function with constraints and a differential multiplier method is utilized to iteratively find optimal embeddings. This mitigates oversmoothing and suppresses the quantization error to a bound. Significant improvements are demonstrated over state-of-the-art quantization methods and deep GNN approaches in both full-precision and quantized models. The proposed quantizer demonstrates superior performance in INT2 configurations across all stages of GNN, achieving a notable level of accuracy. In contrast, existing quantization approaches fail to generate satisfactory accuracy levels. Finally, the inference with INT2 and INT4 representations exhibits a speedup of 5.11 $\times$ and 4.70 $\times$ compared to full precision counterparts, respectively.
Deep Neural Networks (DNNs) have been a large driver for AI breakthroughs in recent years. However, these models have been getting increasingly large as they become more accurate and safe. This means that their training becomes increasingly costly and time-consuming and typically yields a single model to fit all targets. Various techniques have been proposed in the literature to mitigate this, including pruning, sparsification, or quantization of model weights and updates. While achieving high compression rates, they often incur significant computational overheads at training or lead to non-negligible accuracy penalty. Alternatively, factorization methods have been leveraged for low-rank compression of DNNs. Similarly, such techniques (e.g., SVD) frequently rely on heavy iterative decompositions of layers and are potentially sub-optimal for non-linear models, such as DNNs. We take a further step in designing efficient low-rank models and propose Maestro, a framework for trainable low-rank layers. Instead of iteratively applying a priori decompositions, the low-rank structure is baked into the training process through LoD, a low-rank ordered decomposition. Not only is this the first time importance ordering via sampling is applied on the decomposed DNN structure, but it also allows selecting ranks at a layer granularity. Our theoretical analysis demonstrates that in special cases LoD recovers the SVD decomposition and PCA. Applied to DNNs, Maestro enables the extraction of lower footprint models that preserve performance. Simultaneously, it enables the graceful trade-off between accuracy-latency for deployment to even more constrained devices without retraining.
Retrieval augmentation is a powerful but expensive method to make language models more knowledgeable about the world. Memory-based methods like LUMEN pre-compute token representations for retrieved passages to drastically speed up inference. However, memory also leads to much greater storage requirements from storing pre-computed representations. We propose MEMORY-VQ, a new method to reduce storage requirements of memory-augmented models without sacrificing performance. Our method uses a vector quantization variational autoencoder (VQ-VAE) to compress token representations. We apply MEMORY-VQ to the LUMEN model to obtain LUMEN-VQ, a memory model that achieves a 16x compression rate with comparable performance on the KILT benchmark. LUMEN-VQ enables practical retrieval augmentation even for extremely large retrieval corpora.
Existing Continual Learning (CL) solutions only partially address the constraints on power, memory and computation of the deep learning models when deployed on low-power embedded CPUs. In this paper, we propose a CL solution that embraces the recent advancements in CL field and the efficiency of the Binary Neural Networks (BNN), that use 1-bit for weights and activations to efficiently execute deep learning models. We propose a hybrid quantization of CWR* (an effective CL approach) that considers differently forward and backward pass in order to retain more precision during gradient update step and at the same time minimizing the latency overhead. The choice of a binary network as backbone is essential to meet the constraints of low power devices and, to the best of authors' knowledge, this is the first attempt to prove on-device learning with BNN. The experimental validation carried out confirms the validity and the suitability of the proposed method.
In an age dominated by resource-intensive foundation models, the ability to efficiently adapt to downstream tasks is crucial. Visual Prompting (VP), drawing inspiration from the prompting techniques employed in Large Language Models (LLMs), has emerged as a pivotal method for transfer learning in the realm of computer vision. As the importance of efficiency continues to rise, research into model compression has become indispensable in alleviating the computational burdens associated with training and deploying over-parameterized neural networks. A primary objective in model compression is to develop sparse and/or quantized models capable of matching or even surpassing the performance of their over-parameterized, full-precision counterparts. Although previous studies have explored the effects of model compression on transfer learning, its impact on visual prompting-based transfer remains unclear. This study aims to bridge this gap, shedding light on the fact that model compression detrimentally impacts the performance of visual prompting-based transfer, particularly evident in scenarios with low data volume. Furthermore, our findings underscore the adverse influence of sparsity on the calibration of downstream visual-prompted models. However, intriguingly, we also illustrate that such negative effects on calibration are not present when models are compressed via quantization. This empirical investigation underscores the need for a nuanced understanding beyond mere accuracy in sparse and quantized settings, thereby paving the way for further exploration in Visual Prompting techniques tailored for sparse and quantized models.
Few-shot font generation is challenging, as it needs to capture the fine-grained stroke styles from a limited set of reference glyphs, and then transfer to other characters, which are expected to have similar styles. However, due to the diversity and complexity of Chinese font styles, the synthesized glyphs of existing methods usually exhibit visible artifacts, such as missing details and distorted strokes. In this paper, we propose a VQGAN-based framework (i.e., VQ-Font) to enhance glyph fidelity through token prior refinement and structure-aware enhancement. Specifically, we pre-train a VQGAN to encapsulate font token prior within a codebook. Subsequently, VQ-Font refines the synthesized glyphs with the codebook to eliminate the domain gap between synthesized and real-world strokes. Furthermore, our VQ-Font leverages the inherent design of Chinese characters, where structure components such as radicals and character components are combined in specific arrangements, to recalibrate fine-grained styles based on references. This process improves the matching and fusion of styles at the structure level. Both modules collaborate to enhance the fidelity of the generated fonts. Experiments on a collected font dataset show that our VQ-Font outperforms the competing methods both quantitatively and qualitatively, especially in generating challenging styles.
JPEG is one of the most popular image compression methods. It is beneficial to compress those existing JPEG files without introducing additional distortion. In this paper, we propose a deep learning based method to further compress JPEG images losslessly. Specifically, we propose a Multi-Level Parallel Conditional Modeling (ML-PCM) architecture, which enables parallel decoding in different granularities. First, luma and chroma are processed independently to allow parallel coding. Second, we propose pipeline parallel context model (PPCM) and compressed checkerboard context model (CCCM) for the effective conditional modeling and efficient decoding within luma and chroma components. Our method has much lower latency while achieves better compression ratio compared with previous SOTA. After proper software optimization, we can obtain a good throughput of 57 FPS for 1080P images on NVIDIA T4 GPU. Furthermore, combined with quantization, our approach can also act as a lossy JPEG codec which has obvious advantage over SOTA lossy compression methods in high bit rate (bpp$>0.9$).
Large language models (LLMs) have revolutionized natural language processing tasks. However, their practical deployment is hindered by their immense memory and computation requirements. Although recent post-training quantization (PTQ) methods are effective in reducing memory footprint and improving the computational efficiency of LLM, they hand-craft quantization parameters, leading to low performance, especially in extremely low-bit quantization. To tackle this issue, we introduce an Omnidirectionally calibrated Quantization (\textbf{OmniQuant}) technique for LLMs, which achieves good performance in diverse quantization settings while maintaining the computational efficiency of PTQ by efficiently optimizing various quantization parameters. OmniQuant comprises two innovative components including Learnable Weight Clipping (LWC) and Learnable Equivalent Transformation (LET). LWC modulates the extreme values of weights by optimizing the clipping threshold. Meanwhile, LET tackles activation outliers by shifting the challenge of quantization from activations to weights. Operating within a differentiable framework using block-wise error minimization, OmniQuant can optimize the quantization process efficiently for both weight-only and weight-activation quantization. For instance, the LLaMA-2 model family size 7-70B can be processed with OmniQuant on a single A100-40G GPU within 1-16 hours using 128 samples. Extensive experiments validate OmniQuant's superior performance across diverse quantization configurations such as W4A4 (4-bit weight, 4-bit activation), W6A6, W4A16, W3A16, and W2A16. Additionally, OmniQuant demonstrates effectiveness in instruction-tuned models and delivers notable improvements in inference speed and memory reduction on real devices. Codes are available at \url{https://github.com/OpenGVLab/OmniQuant}.
We present accumulator-aware quantization (A2Q), a novel weight quantization method designed to train quantized neural networks (QNNs) to avoid overflow when using low-precision accumulators during inference. A2Q introduces a unique formulation inspired by weight normalization that constrains the L1-norm of model weights according to accumulator bit width bounds that we derive. Thus, in training QNNs for low-precision accumulation, A2Q also inherently promotes unstructured weight sparsity to guarantee overflow avoidance. We apply our method to deep learning-based computer vision tasks to show that A2Q can train QNNs for low-precision accumulators while maintaining model accuracy competitive with a floating-point baseline. In our evaluations, we consider the impact of A2Q on both general-purpose platforms and programmable hardware. However, we primarily target model deployment on FPGAs because they can be programmed to fully exploit custom accumulator bit widths. Our experimentation shows accumulator bit width significantly impacts the resource efficiency of FPGA-based accelerators. On average across our benchmarks, A2Q offers up to a 2.3x reduction in resource utilization over 32-bit accumulator counterparts with 99.2% of the floating-point model accuracy.
This paper studies distributed Nash equilibrium (NE) seeking under Denial-of-Service (DoS) attacks and quantization. The players can only exchange information with their own direct neighbors. The transmitted information is subject to quantization and packet losses induced by malicious DoS attacks. We propose a quantized distributed NE seeking strategy based on the approach of dynamic quantized consensus. To solve the quantizer saturation problem caused by DoS attacks, the quantization mechanism is equipped to have zooming-in and holding capabilities, in which the holding capability is consistent with the results in quantized consensus under DoS. A sufficient condition on the number of quantizer levels is provided, under which the quantizers are free from saturation under DoS attacks. The proposed distributed quantized NE seeking strategy is shown to have the so-called maximum resilience to DoS attacks. Namely, if the bound characterizing the maximum resilience is violated, an attacker can deny all the transmissions and hence distributed NE seeking is impossible.
In recent years, audio coding technology has been standardized based on several frameworks that incorporate linear predictive coding (LPC). However, coding the transient signal using frequency-domain LP residual signals remains a challenge. To address this, temporal noise shaping (TNS) can be adapted, although it cannot be effectively operated since the estimated temporal envelope in the modified discrete cosine transform (MDCT) domain is accompanied by the time-domain aliasing (TDA) terms. In this study, we propose the modulated complex lapped transform-based coding framework integrated with transform coded excitation (TCX) and complex LPC-based TNS (CTNS). Our approach uses a 50\% overlap window and switching scheme for the CTNS to improve the coding efficiency. Additionally, an adaptive calculation of the target bits for the sub-bands using the frequency envelope information based on the quantized LPC coefficients is proposed. To minimize the quantization mismatch between both modes, an integrated quantization for real and complex values and a TDA augmentation method that compensates for the artificially generated TDA components during switching operations are proposed. The proposed coding framework shows a superior performance in both objective metrics and subjective listening tests, thereby demonstrating its low bit-rate audio coding.
Digitalizing real-world analog signals typically involves sampling in time and discretizing in amplitude. Subsequent signal reconstructions inevitably incur an error that depends on the amplitude resolution and the temporal density of the acquired samples. From an implementation viewpoint, consistent signal reconstruction methods have proven a profitable error-rate decay as the sampling rate increases. Despite that, these results are obtained under offline settings. Therefore, a research gap exists regarding methods for consistent signal reconstruction from data streams. Solving this problem is of great importance because such methods could run at a lower computational cost than the existing offline ones or be used under real-time requirements without losing the benefits of ensuring consistency. In this paper, we formalize for the first time the concept of consistent signal reconstruction from streaming time-series data. Then, we present a signal reconstruction method able to enforce consistency and also exploit the spatiotemporal dependencies of streaming multivariate time-series data to further reduce the signal reconstruction error. Our experiments show that our proposed method achieves a favorable error-rate decay with the sampling rate compared to a similar but non-consistent reconstruction.
With the ever-growing complexity of deep learning models for face recognition, it becomes hard to deploy these systems in real life. Researchers have two options: 1) use smaller models; 2) compress their current models. Since the usage of smaller models might lead to concerning biases, compression gains relevance. However, compressing might be also responsible for an increase in the bias of the final model. We investigate the overall performance, the performance on each ethnicity subgroup and the racial bias of a State-of-the-Art quantization approach when used with synthetic and real data. This analysis provides a few more details on potential benefits of performing quantization with synthetic data, for instance, the reduction of biases on the majority of test scenarios. We tested five distinct architectures and three different training datasets. The models were evaluated on a fourth dataset which was collected to infer and compare the performance of face recognition models on different ethnicity.
Statistical and stochastic analysis based on thermodynamics has been the main analysis framework for stochastic global optimization. Recently, appearing quantum annealing or quantum tunneling algorithm for global optimization, we require a new researching framework for global optimization algorithms. In this paper, we provide the analysis for quantization-based optimization based on the Schr\"odinger equation to reveal what property in quantum mechanics enables global optimization. We present that the tunneling effect derived by the Schr\"odinger equation in quantization-based optimization enables to escape of a local minimum. Additionally, we confirm that this tunneling effect is the same property included in quantum mechanics-based global optimization. Experiments with standard multi-modal benchmark functions represent that the proposed analysis is valid.
In this work, we propose distributed and networked energy management scenarios to optimize the production and reservation of energy among a set of distributed energy nodes. In other words, the idea is to optimally allocate the generated and reserved powers based on nodes' local cost gradient information while meeting the demand energy. One main concern is the all-time (or anytime) resource-demand feasibility, implying that at all iterations of the scheduling algorithm, the balance between the produced power and demand plus reserved power must hold. The other concern is to design algorithms to tolerate communication time-delays and changes in the network. Further, one can incorporate possible model nonlinearity in the algorithm to address both inherent (e.g., saturation and quantization) and purposefully-added (e.g., signum-based) nonlinearities in the model. The proposed optimal allocation algorithm addresses all the above concerns, while it benefits from possible features of the distributed (or networked) solutions such as no-single-node-of-failure and distributed information processing. We show both the all-time feasibility of the proposed scheme and its convergence under certain bound on the step-rate using Lyapunov-type proofs.
Super-resolution (SR) networks have been investigated for a while, with their mobile and lightweight versions gaining noticeable popularity recently. Quantization, the procedure of decreasing the precision of network parameters (mostly FP32 to INT8), is also utilized in SR networks for establishing mobile compatibility. This study focuses on a very important but mostly overlooked post-training quantization (PTQ) step: representative dataset (RD), which adjusts the quantization range for PTQ. We propose a novel pipeline (clip-free quantization pipeline, CFQP) backed up with extensive experimental justifications to cleverly augment RD images by only using outputs of the FP32 model. Using the proposed pipeline for RD, we can successfully eliminate unwanted clipped activation layers, which nearly all mobile SR methods utilize to make the model more robust to PTQ in return for a large overhead in runtime. Removing clipped activations with our method significantly benefits overall increased stability, decreased inference runtime up to 54% on some SR models, better visual quality results compared to INT8 clipped models - and outperforms even some FP32 non-quantized models, both in runtime and visual quality, without the need for retraining with clipped activation.
There has been many papers in academic literature on quantizing weight tensors in deep learning models to reduce inference latency and memory footprint. TVM also has the ability to quantize weights and support low-bit computations. Although quantization is typically expected to improve inference time, in TVM, the performance of 8-bit quantization does not meet the expectations. Typically, when applying 8-bit quantization to a deep learning model, it is usually expected to achieve around 50% of the full-precision inference time. However, in this particular case, not only does the quantized version fail to achieve the desired performance boost, but it actually performs worse, resulting in an inference time that is about 2 times as slow as the non-quantized version. In this project, we thoroughly investigate the reasons behind the underperformance and assess the compatibility and optimization opportunities of 8-bit quantization in TVM. We discuss the optimization of two different types of tasks: computation-bound and memory-bound, and provide a detailed comparison of various optimization techniques in TVM. Through the identification of performance issues, we have successfully improved quantization by addressing a bug in graph building. Furthermore, we analyze multiple optimization strategies to achieve the optimal quantization result. The best experiment achieves 163.88% improvement compared with the TVM compiled baseline in inference time for the compute-bound task and 194.98% for the memory-bound task.
Large Language Models (LLMs) have achieved state-of-the-art performance across various language tasks but pose challenges for practical deployment due to their substantial memory requirements. Furthermore, the latest generative models suffer from high inference costs caused by the memory bandwidth bottleneck in the auto-regressive decoding process. To address these issues, we propose an efficient weight-only quantization method that reduces memory consumption and accelerates inference for LLMs. To ensure minimal quality degradation, we introduce a simple and effective heuristic approach that utilizes only the model weights of a pre-trained model. This approach is applicable to both Mixture-of-Experts (MoE) and dense models without requiring additional fine-tuning. To demonstrate the effectiveness of our proposed method, we first analyze the challenges and issues associated with LLM quantization. Subsequently, we present our heuristic approach, which adaptively finds the granularity of quantization, effectively addressing these problems. Furthermore, we implement highly efficient GPU GEMMs that perform on-the-fly matrix multiplication and dequantization, supporting the multiplication of fp16 or bf16 activations with int8 or int4 weights. We evaluate our approach on large-scale open source models such as OPT-175B and internal MoE models, showcasing minimal accuracy loss while achieving up to 3.65 times higher throughput on the same number of GPUs.