As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.
Zero-shot singing voice synthesis (SVS) with style transfer and style control aims to generate high-quality singing voices with unseen timbres and styles (including singing method, emotion, rhythm, technique, and pronunciation) from audio and text prompts. However, the multifaceted nature of singing styles poses a significant challenge for effective modeling, transfer, and control. Furthermore, current SVS models often fail to generate singing voices rich in stylistic nuances for unseen singers. To address these challenges, we introduce TCSinger, the first zero-shot SVS model for style transfer across cross-lingual speech and singing styles, along with multi-level style control. Specifically, TCSinger proposes three primary modules: 1) the clustering style encoder employs a clustering vector quantization model to stably condense style information into a compact latent space; 2) the Style and Duration Language Model (S\&D-LM) concurrently predicts style information and phoneme duration, which benefits both; 3) the style adaptive decoder uses a novel mel-style adaptive normalization method to generate singing voices with enhanced details. Experimental results show that TCSinger outperforms all baseline models in synthesis quality, singer similarity, and style controllability across various tasks, including zero-shot style transfer, multi-level style control, cross-lingual style transfer, and speech-to-singing style transfer. Singing voice samples can be accessed at https://aaronz345.github.io/TCSingerDemo/.
Dynamic quantization has attracted rising attention in image super-resolution (SR) as it expands the potential of heavy SR models onto mobile devices while preserving competitive performance. Existing methods explore layer-to-bit configuration upon varying local regions, adaptively allocating the bit to each layer and patch. Despite the benefits, they still fall short in the trade-off of SR accuracy and quantization efficiency. Apart from this, adapting the quantization level for each layer individually can disturb the original inter-layer relationships, thus diminishing the representation capability of quantized models. In this work, we propose Granular-DQ, which capitalizes on the intrinsic characteristics of images while dispensing with the previous consideration for layer sensitivity in quantization. Granular-DQ conducts a multi-granularity analysis of local patches with further exploration of their information densities, achieving a distinctive patch-wise and layer-invariant dynamic quantization paradigm. Specifically, Granular-DQ initiates by developing a granularity-bit controller (GBC) to apprehend the coarse-to-fine granular representations of different patches, matching their proportional contribution to the entire image to determine the proper bit-width allocation. On this premise, we investigate the relation between bit-width and information density, devising an entropy-to-bit (E2B) mechanism that enables further fine-grained dynamic bit adaption of high-bit patches. Extensive experiments validate the superiority and generalization ability of Granular-DQ over recent state-of-the-art methods on various SR models. Code and supplementary statement can be found at \url{https://github.com/MmmingS/Granular-DQ.git}.
Disentangled representation learning aims to represent the underlying generative factors of a dataset in a latent representation independently of one another. In our work, we propose a discrete variational autoencoder (VAE) based model where the ground truth information about the generative factors are not provided to the model. We demonstrate the advantages of learning discrete representations over learning continuous representations in facilitating disentanglement. Furthermore, we propose incorporating an inductive bias into the model to further enhance disentanglement. Precisely, we propose scalar quantization of the latent variables in a latent representation with scalar values from a global codebook, and we add a total correlation term to the optimization as an inductive bias. Our method called FactorQVAE combines optimization based disentanglement approaches with discrete representation learning, and it outperforms the former disentanglement methods in terms of two disentanglement metrics (DCI and InfoMEC) while improving the reconstruction performance. Our code can be found at https://github.com/ituvisionlab/FactorQVAE.
This work introduces MICSim, an open-source, pre-circuit simulator designed for early-stage evaluation of chip-level software performance and hardware overhead of mixed-signal compute-in-memory (CIM) accelerators. MICSim features a modular design, allowing easy multi-level co-design and design space exploration. Modularized from the state-of-the-art CIM simulator NeuroSim, MICSim provides a highly configurable simulation framework supporting multiple quantization algorithms, diverse circuit/architecture designs, and different memory devices. This modular approach also allows MICSim to be effectively extended to accommodate new designs. MICSim natively supports evaluating accelerators' software and hardware performance for CNNs and Transformers in Python, leveraging the popular PyTorch and HuggingFace Transformers frameworks. These capabilities make MICSim highly adaptive when simulating different networks and user-friendly. This work demonstrates that MICSim can easily be combined with optimization strategies to perform design space exploration and used for chip-level Transformers CIM accelerators evaluation. Also, MICSim can achieve a 9x - 32x speedup of NeuroSim through a statistic-based average mode proposed by this work.
Quantization has gained attention as a promising solution for the cost-effective deployment of large and small language models. However, most prior work has been limited to perplexity or basic knowledge tasks and lacks a comprehensive evaluation of recent models like Llama-3.3. In this paper, we conduct a comprehensive evaluation of instruction-tuned models spanning 1B to 405B parameters, applying four quantization methods across 13 datasets. Our findings reveal that (1) quantized models generally surpass smaller FP16 baselines, yet they often struggle with instruction-following and hallucination detection; (2) FP8 consistently emerges as the most robust option across tasks, and AWQ tends to outperform GPTQ in weight-only quantization; (3) smaller models can suffer severe accuracy drops at 4-bit quantization, while 70B-scale models maintain stable performance; (4) notably, \textit{hard} tasks do not always experience the largest accuracy losses, indicating that quantization magnifies a model's inherent weaknesses rather than simply correlating with task difficulty; and (5) an LLM-based judge (MT-Bench) highlights significant performance declines in Coding and STEM tasks, though it occasionally reports improvements in reasoning.
Approximate nearest neighbor (ANN) query in high-dimensional Euclidean space is a key operator in database systems. For this query, quantization is a popular family of methods developed for compressing vectors and reducing memory consumption. Recently, a method called RaBitQ achieves the state-of-the-art performance among these methods. It produces better empirical performance in both accuracy and efficiency when using the same compression rate and provides rigorous theoretical guarantees. However, the method is only designed for compressing vectors at high compression rates (32x) and lacks support for achieving higher accuracy by using more space. In this paper, we introduce a new quantization method to address this limitation by extending RaBitQ. The new method inherits the theoretical guarantees of RaBitQ and achieves the asymptotic optimality in terms of the trade-off between space and error bounds as to be proven in this study. Additionally, we present efficient implementations of the method, enabling its application to ANN queries to reduce both space and time consumption. Extensive experiments on real-world datasets confirm that our method consistently outperforms the state-of-the-art baselines in both accuracy and efficiency when using the same amount of memory.
Training deep neural networks (DNNs) is costly. Fortunately, Nvidia Ampere and Hopper GPUs can accelerate matrix multiplications twice as fast as a dense equivalent by implementing 2:4 sparsity. However, previous STE-based 2:4 pre-training methods (e.g. STE with hard-thresholding, SR-STE) suffer from optimization difficulties because of discontinuous pruning function. In this study, we comprehensively analyse the bottleneck of traditional N:M sparse training and recognize three drawbacks with discontinuity: incorrect descending direction, inability to predict the amount of descent and sparse mask oscillation. In light of this, we propose S-STE, a simple yet powerful 2:4 training method that contains two parts: to continuously project weights to be 2:4 sparse, and to rescale sparse weights with a per-tensor fixed scaling factor. Besides, we adopt minimum-variance unbiased estimation for activation gradient and FP8 quantization for whole process. Results show that our method surpasses previous 2:4 pre-training recipes and is comparable even with full parameter models. Our toolkit is available at https://github.com/huyz2023/2by4-pretrain.
The advent of large language models (LLMs) revolutionized natural language processing applications, and running LLMs on edge devices has become increasingly attractive for reasons including reduced latency, data localization, and personalized user experiences. This comprehensive review examines the challenges of deploying computationally expensive LLMs on resource-constrained devices and explores innovative solutions across multiple domains. The paper investigates the development of on-device language models, their efficient architectures, including parameter sharing and modular designs, as well as state-of-the-art compression techniques like quantization, pruning, and knowledge distillation. Hardware acceleration strategies and collaborative edge-cloud deployment approaches are analyzed, highlighting the intricate balance between performance and resource utilization. Case studies of on-device language models from major mobile manufacturers demonstrate real-world applications and potential benefits. The review also addresses critical aspects such as adaptive learning, multi-modal capabilities, and personalization. By identifying key research directions and open challenges, this paper provides a roadmap for future advancements in on-device language models, emphasizing the need for interdisciplinary efforts to realize the full potential of ubiquitous, intelligent computing while ensuring responsible and ethical deployment. For a comprehensive review of research work and educational resources on on-device large language models (LLMs), please visit https://github.com/NexaAI/Awesome-LLMs-on-device. To download and run on-device LLMs, visit https://www.nexaai.com/models.
We present BigCodec, a low-bitrate neural speech codec. While recent neural speech codecs have shown impressive progress, their performance significantly deteriorates at low bitrates (around 1 kbps). Although a low bitrate inherently restricts performance, other factors, such as model capacity, also hinder further improvements. To address this problem, we scale up the model size to 159M parameters that is more than 10 times larger than popular codecs with about 10M parameters. Besides, we integrate sequential models into traditional convolutional architectures to better capture temporal dependency and adopt low-dimensional vector quantization to ensure a high code utilization. Comprehensive objective and subjective evaluations show that BigCodec, with a bitrate of 1.04 kbps, significantly outperforms several existing low-bitrate codecs. Furthermore, BigCodec achieves objective performance comparable to popular codecs operating at 4-6 times higher bitrates, and even delivers better subjective perceptual quality than the ground truth.
Diffusion Transformers (DiTs) have recently attracted significant interest from both industry and academia due to their enhanced capabilities in visual generation, surpassing the performance of traditional diffusion models that employ U-Net. However, the improved performance of DiTs comes at the expense of higher parameter counts and implementation costs, which significantly limits their deployment on resource-constrained devices like mobile phones. We propose DiTAS, a data-free post-training quantization (PTQ) method for efficient DiT inference. DiTAS relies on the proposed temporal-aggregated smoothing techniques to mitigate the impact of the channel-wise outliers within the input activations, leading to much lower quantization error under extremely low bitwidth. To further enhance the performance of the quantized DiT, we adopt the layer-wise grid search strategy to optimize the smoothing factor. Moreover, we integrate a training-free LoRA module for weight quantization, leveraging alternating optimization to minimize quantization errors without additional fine-tuning. Experimental results demonstrate that our approach enables 4-bit weight, 8-bit activation (W4A8) quantization for DiTs while maintaining comparable performance as the full-precision model.
Bit-level sparsity methods skip ineffectual zero-bit operations and are typically applicable within bit-serial deep learning accelerators. This type of sparsity at the bit-level is especially interesting because it is both orthogonal and compatible with other deep neural network (DNN) efficiency methods such as quantization and pruning. In this work, we improve the practicality and efficiency of bitlevel sparsity through a novel algorithmic bit-pruning, averaging, and compression method, and a co-designed efficient bit-serial hardware accelerator. On the algorithmic side, we introduce bidirectional bit sparsity (BBS). The key insight of BBS is that we can leverage bit sparsity in a symmetrical way to prune either zero-bits or one-bits. This significantly improves the load balance of bit-serial computing and guarantees the level of sparsity to be more than 50%. On top of BBS, we further propose two bit-level binary pruning methods that require no retraining, and can be seamlessly applied to quantized DNNs. Combining binary pruning with a new tensor encoding scheme, BBS can both skip computation and reduce the memory footprint associated with bi-directional sparse bit columns. On the hardware side, we demonstrate the potential of BBS through BitVert, a bitserial architecture with an efficient PE design to accelerate DNNs with low overhead, exploiting our proposed binary pruning. Evaluation on seven representative DNN models shows that our approach achieves: (1) on average 1.66$\times$ reduction in model sizewith negligible accuracy loss of < 0.5%; (2) up to 3.03$\times$ speedupand 2.44$\times$ energy saving compared to prior DNN accelerators.
Recent large language models (LLMs) have enabled the development of advanced agentic systems that can integrate various tools and APIs to fulfill user queries through function calling. However, the deployment of these LLMs on the edge has not been explored since they typically require cloud-based infrastructure due to their substantial model size and computational demands. To this end, we present TinyAgent, an end-to-end framework for training and deploying task-specific small language model agents capable of function calling for driving agentic systems at the edge. We first show how to enable accurate function calling for open-source models via the LLMCompiler framework. We then systematically curate a high-quality dataset for function calling, which we use to fine-tune two small language models, TinyAgent-1.1B and 7B. For efficient inference, we introduce a novel tool retrieval method to reduce the input prompt length and utilize quantization to further accelerate the inference speed. As a driving application, we demonstrate a local Siri-like system for Apple's MacBook that can execute user commands through text or voice input. Our results show that our models can achieve, and even surpass, the function-calling capabilities of larger models like GPT-4-Turbo, while being fully deployed at the edge. We open-source our dataset, models, and installable package and provide a demo video for our MacBook assistant agent.
This paper addresses the limitations of conventional vector quantization algorithms, particularly K-Means and its variant K-Means++, and investigates the Stochastic Quantization (SQ) algorithm as a scalable alternative for high-dimensional unsupervised and semi-supervised learning tasks. Traditional clustering algorithms often suffer from inefficient memory utilization during computation, necessitating the loading of all data samples into memory, which becomes impractical for large-scale datasets. While variants such as Mini-Batch K-Means partially mitigate this issue by reducing memory usage, they lack robust theoretical convergence guarantees due to the non-convex nature of clustering problems. In contrast, the Stochastic Quantization algorithm provides strong theoretical convergence guarantees, making it a robust alternative for clustering tasks. We demonstrate the computational efficiency and rapid convergence of the algorithm on an image classification problem with partially labeled data, comparing model accuracy across various ratios of labeled to unlabeled data. To address the challenge of high dimensionality, we employ a Triplet Network to encode images into low-dimensional representations in a latent space, which serve as a basis for comparing the efficiency of both the Stochastic Quantization algorithm and traditional quantization algorithms. Furthermore, we enhance the algorithm's convergence speed by introducing modifications with an adaptive learning rate.
This paper focuses on modern efficient training and inference technologies on foundation models and illustrates them from two perspectives: model and system design. Model and System Design optimize LLM training and inference from different aspects to save computational resources, making LLMs more efficient, affordable, and more accessible. The paper list repository is available at https://github.com/NoakLiu/Efficient-Foundation-Models-Survey.
Normalizing flows, a category of probabilistic models famed for their capabilities in modeling complex data distributions, have exhibited remarkable efficacy in unsupervised anomaly detection. This paper explores the potential of normalizing flows in multi-class anomaly detection, wherein the normal data is compounded with multiple classes without providing class labels. Through the integration of vector quantization (VQ), we empower the flow models to distinguish different concepts of multi-class normal data in an unsupervised manner, resulting in a novel flow-based unified method, named VQ-Flow. Specifically, our VQ-Flow leverages hierarchical vector quantization to estimate two relative codebooks: a Conceptual Prototype Codebook (CPC) for concept distinction and its concomitant Concept-Specific Pattern Codebook (CSPC) to capture concept-specific normal patterns. The flow models in VQ-Flow are conditioned on the concept-specific patterns captured in CSPC, capable of modeling specific normal patterns associated with different concepts. Moreover, CPC further enables our VQ-Flow for concept-aware distribution modeling, faithfully mimicking the intricate multi-class normal distribution through a mixed Gaussian distribution reparametrized on the conceptual prototypes. Through the introduction of vector quantization, the proposed VQ-Flow advances the state-of-the-art in multi-class anomaly detection within a unified training scheme, yielding the Det./Loc. AUROC of 99.5%/98.3% on MVTec AD. The codebase is publicly available at https://github.com/cool-xuan/vqflow.
The rapid growth of large models' size has far outpaced that of computing resources. To bridge this gap, encouraged by the parsimonious relationship between genotype and phenotype in the brain's growth and development, we propose the so-called hyper-compression that turns the model compression into the issue of parameter representation via a hyperfunction. Specifically, it is known that the trajectory of some low-dimensional dynamic systems can fill the high-dimensional space eventually. Thus, hyper-compression, using these dynamic systems as the hyperfunctions, represents the parameters of the target network by their corresponding composition number or trajectory length. This suggests a novel mechanism for model compression, substantially different from the existing pruning, quantization, distillation, and decomposition. Along this direction, we methodologically identify a suitable dynamic system with the irrational winding as the hyperfunction and theoretically derive its associated error bound. Next, guided by our theoretical insights, we propose several engineering twists to make the hyper-compression pragmatic and effective. Lastly, systematic and comprehensive experiments confirm that hyper-compression enjoys the following \textbf{PNAS} merits: 1) \textbf{P}referable compression ratio; 2) \textbf{N}o post-hoc retraining; 3) \textbf{A}ffordable inference time; and 4) \textbf{S}hort compression time. It compresses LLaMA2-7B in an hour and achieves close-to-int4-quantization performance, without retraining and with a performance drop of less than 1\%. We have open-sourced our code in https://github.com/Juntongkuki/Hyper-Compression.git for free download and evaluation.
Background and Aims: This study evaluates the medical reasoning performance of large language models (LLMs) and vision language models (VLMs) in gastroenterology. Methods: We used 300 gastroenterology board exam-style multiple-choice questions, 138 of which contain images to systematically assess the impact of model configurations and parameters and prompt engineering strategies utilizing GPT-3.5. Next, we assessed the performance of proprietary and open-source LLMs (versions), including GPT (3.5, 4, 4o, 4omini), Claude (3, 3.5), Gemini (1.0), Mistral, Llama (2, 3, 3.1), Mixtral, and Phi (3), across different interfaces (web and API), computing environments (cloud and local), and model precisions (with and without quantization). Finally, we assessed accuracy using a semiautomated pipeline. Results: Among the proprietary models, GPT-4o (73.7%) and Claude3.5-Sonnet (74.0%) achieved the highest accuracy, outperforming the top open-source models: Llama3.1-405b (64%), Llama3.1-70b (58.3%), and Mixtral-8x7b (54.3%). Among the quantized open-source models, the 6-bit quantized Phi3-14b (48.7%) performed best. The scores of the quantized models were comparable to those of the full-precision models Llama2-7b, Llama2--13b, and Gemma2-9b. Notably, VLM performance on image-containing questions did not improve when the images were provided and worsened when LLM-generated captions were provided. In contrast, a 10% increase in accuracy was observed when images were accompanied by human-crafted image descriptions. Conclusion: In conclusion, while LLMs exhibit robust zero-shot performance in medical reasoning, the integration of visual data remains a challenge for VLMs. Effective deployment involves carefully determining optimal model configurations, encouraging users to consider either the high performance of proprietary models or the flexible adaptability of open-source models.
How can balance be quantified in game settings? This question is crucial for game designers, especially in player-versus-player (PvP) games, where analyzing the strength relations among predefined team compositions-such as hero combinations in multiplayer online battle arena (MOBA) games or decks in card games-is essential for enhancing gameplay and achieving balance. We have developed two advanced measures that extend beyond the simplistic win rate to quantify balance in zero-sum competitive scenarios. These measures are derived from win value estimations, which employ strength rating approximations via the Bradley-Terry model and counter relationship approximations via vector quantization, significantly reducing the computational complexity associated with traditional win value estimations. Throughout the learning process of these models, we identify useful categories of compositions and pinpoint their counter relationships, aligning with the experiences of human players without requiring specific game knowledge. Our methodology hinges on a simple technique to enhance codebook utilization in discrete representation with a deterministic vector quantization process for an extremely small state space. Our framework has been validated in popular online games, including Age of Empires II, Hearthstone, Brawl Stars, and League of Legends. The accuracy of the observed strength relations in these games is comparable to traditional pairwise win value predictions, while also offering a more manageable complexity for analysis. Ultimately, our findings contribute to a deeper understanding of PvP game dynamics and present a methodology that significantly improves game balance evaluation and design.
Recent advancements in audio generation have been significantly propelled by the capabilities of Large Language Models (LLMs). The existing research on audio LLM has primarily focused on enhancing the architecture and scale of audio language models, as well as leveraging larger datasets, and generally, acoustic codecs, such as EnCodec, are used for audio tokenization. However, these codecs were originally designed for audio compression, which may lead to suboptimal performance in the context of audio LLM. Our research aims to address the shortcomings of current audio LLM codecs, particularly their challenges in maintaining semantic integrity in generated audio. For instance, existing methods like VALL-E, which condition acoustic token generation on text transcriptions, often suffer from content inaccuracies and elevated word error rates (WER) due to semantic misinterpretations of acoustic tokens, resulting in word skipping and errors. To overcome these issues, we propose a straightforward yet effective approach called X-Codec. X-Codec incorporates semantic features from a pre-trained semantic encoder before the Residual Vector Quantization (RVQ) stage and introduces a semantic reconstruction loss after RVQ. By enhancing the semantic ability of the codec, X-Codec significantly reduces WER in speech synthesis tasks and extends these benefits to non-speech applications, including music and sound generation. Our experiments in text-to-speech, music continuation, and text-to-sound tasks demonstrate that integrating semantic information substantially improves the overall performance of language models in audio generation. Our code and demo are available (Demo: https://x-codec-audio.github.io Code: https://github.com/zhenye234/xcodec)