The deployment of large language models (LLMs) is often constrained by memory bandwidth, where the primary bottleneck is the cost of transferring model parameters from the GPU's global memory to its registers. When coupled with custom kernels that fuse the dequantization and matmul operations, weight-only quantization can thus enable faster inference by reducing the amount of memory movement. However, developing high-performance kernels for weight-quantized LLMs presents substantial challenges, especially when the weights are compressed to non-evenly-divisible bit widths (e.g., 3 bits) with non-uniform, lookup table (LUT) quantization. This paper describes FLUTE, a flexible lookup table engine for LUT-quantized LLMs, which uses offline restructuring of the quantized weight matrix to minimize bit manipulations associated with unpacking, and vectorization and duplication of the lookup table to mitigate shared memory bandwidth constraints. At batch sizes < 32 and quantization group size of 128 (typical in LLM inference), the FLUTE kernel can be 2-4x faster than existing GEMM kernels. As an application of FLUTE, we explore a simple extension to lookup table-based NormalFloat quantization and apply it to quantize LLaMA3 to various configurations, obtaining competitive quantization performance against strong baselines while obtaining an end-to-end throughput increase of 1.5 to 2 times.
Large language models (LLMs) have numerous applications across various domains, but their high computational and memory demands pose significant deployment challenges. Weight quantization is an effective technique for reducing the decoding latency and memory requirements of LLMs. Existing approaches primarily aim to maintain the quality of quantized models by preserving outliers in input features, but they still suffer significant quality loss at lower bit widths. Our approach builds on Optimal Brain Quantization (OBQ), an iterative weight-update-based quantization framework. We identify a key limitation of OBQ, specifically that its uniform quantization grid is suboptimal for maintaining model quality, as it introduces large errors to the task loss. To address this, we propose LeanQuant, which learns a loss-error-aware quantization grid by leveraging the inverse diagonal Hessian. Extensive empirical evaluations demonstrate that LeanQuant is both efficient and accurate; it can quantize a 70-billion-parameter model in 6 hours using a single 32GB GPU and performs favorably compared to competitive baselines in the 4-bit, 3-bit, and 2-bit regions.
Semi-supervised learning aims to leverage numerous unlabeled data to improve the model performance. Current semi-supervised 3D object detection methods typically use a teacher to generate pseudo labels for a student, and the quality of the pseudo labels is essential for the final performance. In this paper, we propose PatchTeacher, which focuses on partial scene 3D object detection to provide high-quality pseudo labels for the student. Specifically, we divide a complete scene into a series of patches and feed them to our PatchTeacher sequentially. PatchTeacher leverages the low memory consumption advantage of partial scene detection to process point clouds with a high-resolution voxelization, which can minimize the information loss of quantization and extract more fine-grained features. However, it is non-trivial to train a detector on fractions of the scene. Therefore, we introduce three key techniques, i.e., Patch Normalizer, Quadrant Align, and Fovea Selection, to improve the performance of PatchTeacher. Moreover, we devise PillarMix, a strong data augmentation strategy that mixes truncated pillars from different LiDAR scans to generate diverse training samples and thus help the model learn more general representation. Extensive experiments conducted on Waymo and ONCE datasets verify the effectiveness and superiority of our method and we achieve new state-of-the-art results, surpassing existing methods by a large margin. Codes are available at https://github.com/LittlePey/PTPM.
Diffusion models have transformed the landscape of image generation and now show remarkable potential for image compression. Most of the recent diffusion-based compression methods require training and are tailored for a specific bit-rate. In this work, we propose Posterior Sampling-based Compression (PSC) - a zero-shot compression method that leverages a pre-trained diffusion model as its sole neural network component, thus enabling the use of diverse, publicly available models without additional training. Our approach is inspired by transform coding methods, which encode the image in some pre-chosen transform domain. However, PSC constructs a transform that is adaptive to the image. This is done by employing a zero-shot diffusion-based posterior sampler so as to progressively construct the rows of the transform matrix. Each new chunk of rows is chosen to reduce the uncertainty about the image given the quantized measurements collected thus far. Importantly, the same adaptive scheme can be replicated at the decoder, thus avoiding the need to encode the transform itself. We demonstrate that even with basic quantization and entropy coding, PSC's performance is comparable to established training-based methods in terms of rate, distortion, and perceptual quality. This is while providing greater flexibility, allowing to choose at inference time any desired rate or distortion.
In the past few years, large-scale pre-trained vision-language models like CLIP have achieved tremendous success in various fields. Naturally, how to transfer the rich knowledge in such huge pre-trained models to downstream tasks and datasets becomes a hot topic. During downstream adaptation, the most challenging problems are overfitting and catastrophic forgetting, which can cause the model to overly focus on the current data and lose more crucial domain-general knowledge. Existing works use classic regularization techniques to solve the problems. As solutions become increasingly complex, the ever-growing storage and inference costs are also a significant problem that urgently needs to be addressed. While in this paper, we start from an observation that proper random noise can suppress overfitting and catastrophic forgetting. Then we regard quantization error as a kind of noise, and explore quantization for regularizing vision-language model, which is quite efficiency and effective. Furthermore, to improve the model's generalization capability while maintaining its specialization capacity at minimal cost, we deeply analyze the characteristics of the weight distribution in prompts, conclude several principles for quantization module design and follow such principles to create several competitive baselines. The proposed method is significantly efficient due to its inherent lightweight nature, making it possible to adapt on extremely resource-limited devices. Our method can be fruitfully integrated into many existing approaches like MaPLe, enhancing accuracy while reducing storage overhead, making it more powerful yet versatile. Extensive experiments on 11 datasets shows great superiority of our method sufficiently. Code is available at https://github.com/beyondhtx/QPrompt.
Field-aware Factorization Machines (FFMs) have emerged as a powerful model for click-through rate prediction, particularly excelling in capturing complex feature interactions. In this work, we present an in-depth analysis of our in-house, Rust-based Deep FFM implementation, and detail its deployment on a CPU-only, multi-data-center scale. We overview key optimizations devised for both training and inference, demonstrated by previously unpublished benchmark results in efficient model search and online training. Further, we detail an in-house weight quantization that resulted in more than an order of magnitude reduction in bandwidth footprint related to weight transfers across data-centres. We disclose the engine and associated techniques under an open-source license to contribute to the broader machine learning community. This paper showcases one of the first successful CPU-only deployments of Deep FFMs at such scale, marking a significant stride in practical, low-footprint click-through rate prediction methodologies.
The modeling of users' behaviors is crucial in modern recommendation systems. A lot of research focuses on modeling users' lifelong sequences, which can be extremely long and sometimes exceed thousands of items. These models use the target item to search for the most relevant items from the historical sequence. However, training lifelong sequences in click through rate (CTR) prediction or personalized search ranking (PSR) is extremely difficult due to the insufficient learning problem of ID embedding, especially when the IDs in the lifelong sequence features do not exist in the samples of training dataset. Additionally, existing target attention mechanisms struggle to learn the multi-modal representations of items in the sequence well. The distribution of multi-modal embedding (text, image and attributes) output of user's interacted items are not properly aligned and there exist divergence across modalities. We also observe that users' search query sequences and item browsing sequences can fully depict users' intents and benefit from each other. To address these challenges, we propose a unified lifelong multi-modal sequence model called SEMINAR-Search Enhanced Multi-Modal Interest Network and Approximate Retrieval. Specifically, a network called Pretraining Search Unit (PSU) learns the lifelong sequences of multi-modal query-item pairs in a pretraining-finetuning manner with multiple objectives: multi-modal alignment, next query-item pair prediction, query-item relevance prediction, etc. After pretraining, the downstream model restores the pretrained embedding as initialization and finetunes the network. To accelerate the online retrieval speed of multi-modal embedding, we propose a multi-modal codebook-based product quantization strategy to approximate the exact attention calculati
This report introduces the Qwen2 series, the latest addition to our large language models and large multimodal models. We release a comprehensive suite of foundational and instruction-tuned language models, encompassing a parameter range from 0.5 to 72 billion, featuring dense models and a Mixture-of-Experts model. Qwen2 surpasses most prior open-weight models, including its predecessor Qwen1.5, and exhibits competitive performance relative to proprietary models across diverse benchmarks on language understanding, generation, multilingual proficiency, coding, mathematics, and reasoning. The flagship model, Qwen2-72B, showcases remarkable performance: 84.2 on MMLU, 37.9 on GPQA, 64.6 on HumanEval, 89.5 on GSM8K, and 82.4 on BBH as a base language model. The instruction-tuned variant, Qwen2-72B-Instruct, attains 9.1 on MT-Bench, 48.1 on Arena-Hard, and 35.7 on LiveCodeBench. Moreover, Qwen2 demonstrates robust multilingual capabilities, proficient in approximately 30 languages, spanning English, Chinese, Spanish, French, German, Arabic, Russian, Korean, Japanese, Thai, Vietnamese, and more, underscoring its versatility and global reach. To foster community innovation and accessibility, we have made the Qwen2 model weights openly available on Hugging Face and ModelScope, and the supplementary materials including example code on GitHub. These platforms also include resources for quantization, fine-tuning, and deployment, facilitating a wide range of applications and research endeavors.
With the rapid development of deep learning, Deep Spiking Neural Networks (DSNNs) have emerged as promising due to their unique spike event processing and asynchronous computation. When deployed on neuromorphic chips, DSNNs offer significant power advantages over Deep Artificial Neural Networks (DANNs) and eliminate time and energy consuming multiplications due to the binary nature of spikes (0 or 1). Additionally, DSNNs excel in processing temporal information, making them potentially superior for handling temporal data compared to DANNs. However, their deep network structure and numerous parameters result in high computational costs and energy consumption, limiting real-life deployment. To enhance DSNNs efficiency, researchers have adapted methods from DANNs, such as pruning, quantization, and knowledge distillation, and developed specific techniques like reducing spike firing and pruning time steps. While previous surveys have covered DSNNs algorithms, hardware deployment, and general overviews, focused research on DSNNs compression and efficiency has been lacking. This survey addresses this gap by concentrating on efficient DSNNs and their compression methods. It begins with an exploration of DSNNs' biological background and computational units, highlighting differences from DANNs. It then delves into various compression methods, including pruning, quantization, knowledge distillation, and reducing spike firing, and concludes with suggestions for future research directions.
When Large Language Models (LLMs) are compressed using techniques such as quantization, the predominant way to demonstrate the validity of such techniques is by measuring the model's accuracy on various benchmarks.If the accuracies of the baseline model and the compressed model are close, it is assumed that there was negligible degradation in quality.However, even when the accuracy of baseline and compressed model are similar, we observe the phenomenon of flips, wherein answers change from correct to incorrect and vice versa in proportion.We conduct a detailed study of metrics across multiple compression techniques, models and datasets, demonstrating that the behavior of compressed models as visible to end-users is often significantly different from the baseline model, even when accuracy is similar.We further evaluate compressed models qualitatively and quantitatively using MT-Bench and show that compressed models are significantly worse than baseline models in this free-form generative task.Thus, we argue that compression techniques should also be evaluated using distance metrics.We propose two such metrics, KL-Divergence and flips, and show that they are well correlated.
Various reversible deep neural networks (DNN) models have been proposed to reduce memory consumption in the training process. However, almost all existing reversible DNNs either require special non-standard architectures or are constructed by modifying existing DNN architectures considerably to enable reversibility. In this work we present the BDIA-transformer, which is an exact bit-level reversible transformer that uses an unchanged standard architecture for inference. The basic idea is to first treat each transformer block as the Euler integration approximation for solving an ordinary differential equation (ODE) and then incorporate the technique of bidirectional integration approximation (BDIA) into the neural architecture, together with activation quantization to make it exactly bit-level reversible. In the training process, we let a hyper-parameter $\gamma$ in BDIA-transformer randomly take one of the two values $\{0.5, -0.5\}$ per training sample per transformer block for averaging every two consecutive integration approximations. As a result, BDIA-transformer can be viewed as training an ensemble of ODE solvers parameterized by a set of binary random variables, which regularizes the model and results in improved validation accuracy. Lightweight side information per transformer block is required to be stored in the forward process to account for binary quantization loss to enable exact bit-level reversibility. In the inference procedure, the expectation $\mathbb{E}(\gamma)=0$ is taken to make the resulting architectures of BDIA-transformer identical to transformers up to activation quantization. Our experiments in both image classification and language translation show that BDIA-transformers outperform their conventional counterparts significantly in terms of validation performance while also requiring considerably less training memory.
As Deep Neural Networks (DNNs) rapidly advance in various fields, including speech verification, they typically involve high computational costs and substantial memory consumption, which can be challenging to manage on mobile systems. Quantization of deep models offers a means to reduce both computational and memory expenses. Our research proposes an optimization framework for the quantization of the speaker verification model. By analyzing performance changes and model size reductions in each layer of a pre-trained speaker verification model, we have effectively minimized performance degradation while significantly reducing the model size. Our quantization algorithm is the first attempt to maintain the performance of the state-of-the-art pre-trained speaker verification model, ECAPATDNN, while significantly compressing its model size. Overall, our quantization approach resulted in reducing the model size by half, with an increase in EER limited to 0.07%.
We present MELLE, a novel continuous-valued token based language modeling approach for text-to-speech synthesis (TTS). MELLE autoregressively generates continuous mel-spectrogram frames directly from text condition, bypassing the need for vector quantization, which is typically designed for audio compression and sacrifices fidelity compared to continuous representations. Specifically, (i) instead of cross-entropy loss, we apply regression loss with a proposed spectrogram flux loss function to model the probability distribution of the continuous-valued tokens; (ii) we have incorporated variational inference into MELLE to facilitate sampling mechanisms, thereby enhancing the output diversity and model robustness. Experiments demonstrate that, compared to the two-stage codec language model VALL-E and its variants, the single-stage MELLE mitigates robustness issues by avoiding the inherent flaws of sampling vector-quantized codes, achieves superior performance across multiple metrics, and, most importantly, offers a more streamlined paradigm. The demos of our work are provided at https://aka.ms/melle.
Federated Learning (FL) can protect the privacy of the vehicles in vehicle edge computing (VEC) to a certain extent through sharing the gradients of vehicles' local models instead of local data. The gradients of vehicles' local models are usually large for the vehicular artificial intelligence (AI) applications, thus transmitting such large gradients would cause large per-round latency. Gradient quantization has been proposed as one effective approach to reduce the per-round latency in FL enabled VEC through compressing gradients and reducing the number of bits, i.e., the quantization level, to transmit gradients. The selection of quantization level and thresholds determines the quantization error, which further affects the model accuracy and training time. To do so, the total training time and quantization error (QE) become two key metrics for the FL enabled VEC. It is critical to jointly optimize the total training time and QE for the FL enabled VEC. However, the time-varying channel condition causes more challenges to solve this problem. In this paper, we propose a distributed deep reinforcement learning (DRL)-based quantization level allocation scheme to optimize the long-term reward in terms of the total training time and QE. Extensive simulations identify the optimal weighted factors between the total training time and QE, and demonstrate the feasibility and effectiveness of the proposed scheme.
NLP(natural language processsing) has achieved great success through the transformer model.However, the model has hundreds of millions or billions parameters,which is huge burden for its deployment on personal computer or small scale of server.To deal with it, we either make the model's weight matrix relatively sparser, or compress attention layer. Pattern pruning ,one of the most important pruning methods, permits selecting fixed number of parameters in each divided pattern block and prunes it. However, the effect of pattern pruning is strictly limited by the sparsity within a region of weights in each layer. In this paper,we first introduced Alternating Direction Method of Multipliers(ADMM) based pattern pruning framework to reshape the distribution of activation map. Specifically, we propose to formulate the pattern pruning on transformer as a constrained optimization and use ADMM to optimize the problem. In this way, the initial dense feature maps is transformed to rather regionally sparsified ones.Therefore, we can then achieve higher compression ratio with better performance based on pattern pruning method. Additionally, this paper provides a theoretical derivations of the ADMM with local sparsity. Finally, we also extend the proposed ADMM based framework with SR-STE to demonstrate its generalization and to avoid gradient vanishing problem. We conduct extensive experiments on classification tasks over GLUE datasets. Significantly, we achieve 50% percent compression ratio while maintaining overall score 80.1 on GLUE dataset.
Training Large Language Models (LLMs) is memory-intensive due to the large number of parameters and associated optimization states. GaLore, a recent method, reduces memory usage by projecting weight gradients into a low-rank subspace without compromising performance. However, GaLore relies on time-consuming Singular Value Decomposition (SVD) operations to identify the subspace, and the frequent subspace updates lead to significant training time overhead. Moreover, GaLore offers minimal improvements in accuracy and efficiency compared to LoRA in more accessible fine-tuning scenarios. To address these limitations, we introduce Q-Galore, a novel approach that substantially reduces memory usage by combining quantization and low-rank projection, surpassing the benefits of GaLore. Our method is based on two key observations: (i) the gradient subspace exhibits diverse properties, with some layers converging early in training while others are subject to frequent changes; (ii) the projection matrices are highly resilient to low-bit quantization. Leveraging these insights, Q-GaLore adaptively updates the gradient subspace based on its convergence statistics, achieving comparable performance while significantly reducing the number of SVD operations. We maintain the projection matrices in INT4 format and weights in INT8 format, incorporating stochastic rounding to capture accumulated gradient information. This approach enables a high-precision training trajectory using only low-precision weights. We demonstrate that Q-GaLore achieves highly competitive performance with exceptional memory efficiency. At pre-training, Q-GaLore facilitates training a LLaMA-7B model from scratch on a single NVIDIA RTX 4060 Ti with only 16 GB memory. At fine-tuning, it reduces memory consumption by up to 50% compared to LoRA and GaLore, while consistently outperforming QLoRA at the same memory cost.
Attention, as a core layer of the ubiquitous Transformer architecture, is the bottleneck for large language models and long-context applications. FlashAttention elaborated an approach to speed up attention on GPUs through minimizing memory reads/writes. However, it has yet to take advantage of new capabilities present in recent hardware, with FlashAttention-2 achieving only 35% utilization on the H100 GPU. We develop three main techniques to speed up attention on Hopper GPUs: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-specialization and (2) interleave block-wise matmul and softmax operations, and (3) block quantization and incoherent processing that leverages hardware support for FP8 low-precision. We demonstrate that our method, FlashAttention-3, achieves speedup on H100 GPUs by 1.5-2.0$\times$ with FP16 reaching up to 740 TFLOPs/s (75% utilization), and with FP8 reaching close to 1.2 PFLOPs/s. We validate that FP8 FlashAttention-3 achieves 2.6$\times$ lower numerical error than a baseline FP8 attention.
Low-Rank Adaptation (LoRA), as a representative Parameter-Efficient Fine-Tuning (PEFT)method, significantly enhances the training efficiency by updating only a small portion of the weights in Large Language Models (LLMs). Recently, weight-only quantization techniques have also been applied to LoRA methods to reduce the memory footprint of fine-tuning. However, applying weight-activation quantization to the LoRA pipeline is under-explored, and we observe substantial performance degradation primarily due to the presence of activation outliers. In this work, we propose RoLoRA, the first LoRA-based scheme for effective weight-activation quantization. RoLoRA utilizes rotation for outlier elimination and proposes rotation-aware fine-tuning to preserve the outlier-free characteristics in rotated LLMs. Experimental results show RoLoRA consistently improves low-bit LoRA convergence and post-training quantization robustness in weight-activation settings. We evaluate RoLoRA across LLaMA2-7B/13B, LLaMA3-8B models, achieving up to 29.5% absolute accuracy gain of 4-bit weight-activation quantized LLaMA2- 13B on commonsense reasoning tasks compared to LoRA baseline. We further demonstrate its effectiveness on Large Multimodal Models (LLaVA-1.5-7B). Codes are available at https://github.com/HuangOwen/RoLoRA
Deep learning has made remarkable progress recently, largely due to the availability of large, well-labeled datasets. However, the training on such datasets elevates costs and computational demands. To address this, various techniques like coreset selection, dataset distillation, and dataset quantization have been explored in the literature. Unlike traditional techniques that depend on uniform sample distributions across different classes, our research demonstrates that maintaining performance is feasible even with uneven distributions. We find that for certain classes, the variation in sample quantity has a minimal impact on performance. Inspired by this observation, an intuitive idea is to reduce the number of samples for stable classes and increase the number of samples for sensitive classes to achieve a better performance with the same sampling ratio. Then the question arises: how can we adaptively select samples from a dataset to achieve optimal performance? In this paper, we propose a novel active learning based adaptive sampling strategy, Dataset Quantization with Active Learning based Adaptive Sampling (DQAS), to optimize the sample selection. In addition, we introduce a novel pipeline for dataset quantization, utilizing feature space from the final stage of dataset quantization to generate more precise dataset bins. Our comprehensive evaluations on the multiple datasets show that our approach outperforms the state-of-the-art dataset compression methods.
Low-resolution analog-to-digital converters (ADCs) and hybrid beamforming have emerged as efficient solutions to reduce power consumption with satisfactory spectral efficiency (SE) in massive multiple-input multiple-output (MIMO) systems. In this paper, we investigate the performance of a hybrid receiver in massive MIMO orthogonal frequency-division multiplexing (OFDM) uplink systems with low-resolution ADCs and oversampling. Considering both the temporal and spatial correlation of the quantization distortion (QD), we derive a closed-form approximation of the frequency-domain QD covariance matrix, which facilitates the evaluation of the system's SE. Then we jointly design the analog and digital combiners of the hybrid receiver to maximize the SE. The formulated problem is challenging due to the constant-modulus constraint of the analog combiner and its coupling with the digital one. To overcome these challenges, we transform the objective function into an equivalent but more tractable form and then iteratively update the analog and digital combiners. Numerical simulations verify the superiority of the proposed algorithm over the considered benchmarks and show the resilience of the hybrid receiver to beam squint with low-resolution ADCs. Furthermore, the proposed hybrid receiver design with oversampling can achieve significantly higher energy efficiency compared with the fully digital one.