Quantization reduces the model's hardware costs, such as data movement, storage, and operations like multiply and addition. It also affects the model's behavior by degrading the output quality. Therefore, there is a need for methods that preserve the model's behavior when quantizing model parameters. More exotic numerical encodings, such as block-scaled number formats, have shown advantages for utilizing a fixed bit budget to encode model parameters. This paper presents error diffusion (ED), a hyperparameter-free method for post-training quantization with support for block-scaled data formats. Our approach does not rely on backpropagation or Hessian information. We describe how to improve the quantization process by viewing the neural model as a composite function and diffusing the quantization error in every layer. In addition, we introduce TensorCast, an open-source library based on PyTorch to emulate a variety of number formats, including the block-scaled ones, to aid the research in neural model quantization. We demonstrate the efficacy of our algorithm through rigorous testing on various architectures, including vision and large language models (LLMs), where it consistently delivers competitive results. Our experiments confirm that block-scaled data formats provide a robust choice for post-training quantization and could be used effectively to enhance the practical deployment of advanced neural networks.
We introduce Latent Action Pretraining for general Action models (LAPA), an unsupervised method for pretraining Vision-Language-Action (VLA) models without ground-truth robot action labels. Existing Vision-Language-Action models require action labels typically collected by human teleoperators during pretraining, which significantly limits possible data sources and scale. In this work, we propose a method to learn from internet-scale videos that do not have robot action labels. We first train an action quantization model leveraging VQ-VAE-based objective to learn discrete latent actions between image frames, then pretrain a latent VLA model to predict these latent actions from observations and task descriptions, and finally finetune the VLA on small-scale robot manipulation data to map from latent to robot actions. Experimental results demonstrate that our method significantly outperforms existing techniques that train robot manipulation policies from large-scale videos. Furthermore, it outperforms the state-of-the-art VLA model trained with robotic action labels on real-world manipulation tasks that require language conditioning, generalization to unseen objects, and semantic generalization to unseen instructions. Training only on human manipulation videos also shows positive transfer, opening up the potential for leveraging web-scale data for robotics foundation model.
The edge-device environment imposes severe resource limitations, encompassing computation costs, hardware resource usage, and energy consumption for deploying deep neural network models. Ultra-low-bit quantization and hardware accelerators have been explored as promising approaches to address these challenges. Ultra-low-bit quantization significantly reduces the model size and the computational cost. Despite progress so far, many competitive ultra-low-bit models still partially rely on float or non-ultra-low-bit quantized computation such as the input and output layer. We introduce Efficiera Residual Networks (ERNs), a model optimized for low-resource edge devices. ERNs achieve full ultra-low-bit quantization, with all weights, including the initial and output layers, being binary, and activations set at 2 bits. We introduce the shared constant scaling factor technique to enable integer-valued computation in residual connections, allowing our model to operate without float values until the final convolution layer. Demonstrating competitiveness, ERNs achieve an ImageNet top-1 accuracy of 72.5pt with a ResNet50-compatible architecture and 63.6pt with a model size less than 1MB. Moreover, ERNs exhibit impressive inference times, reaching 300FPS with the smallest model and 60FPS with the largest model on a cost-efficient FPGA device.
Quantization has been substantially adopted to accelerate inference and reduce memory consumption of large language models (LLMs). While activation-weight joint quantization speeds up the inference process through low-precision kernels, we demonstrate that it suffers severe performance degradation on multi-step reasoning tasks, rendering it ineffective. We propose a novel quantization paradigm called QSPEC, which seamlessly integrates two complementary quantization schemes for speculative decoding. Leveraging nearly cost-free execution switching, QSPEC drafts tokens with low-precision, fast activation-weight quantization, and verifies them with high-precision weight-only quantization, effectively combining the strengths of both quantization schemes. Compared to high-precision quantization methods, QSPEC empirically boosts token generation throughput by up to 1.64x without any quality compromise, distinguishing it from other low-precision quantization approaches. This enhancement is also consistent across various serving tasks, model sizes, quantization methods, and batch sizes. Compared to state-of-art speculative decoding methods, our approach reuses weights and the KV cache, avoiding extra memory overhead while achieving up to 1.55x speedup in batched serving with a high acceptance rate. Furthermore, QSPEC offers a plug-and-play advantage without requiring any training. We believe that QSPEC demonstrates unique strengths for future deployment of high-fidelity quantization schemes, particularly in memory-constrained scenarios (e.g., edge devices).
Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks--referred to as Streaming Heads--do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU. Code is provided in https://github.com/mit-han-lab/duo-attention.
Recently, quantization has been widely used for the compression and acceleration of large language models~(LLMs). Due to the outliers in LLMs, it is crucial to flatten weights and activations to minimize quantization error with the equally spaced quantization points. Prior research explores various pre-quantization transformations to suppress outliers, such as per-channel scaling and Hadamard transformation. However, we observe that these transformed weights and activations can still remain steep and outspread. In this paper, we propose FlatQuant (Fast and Learnable Affine Transformation), a new post-training quantization approach to enhance flatness of weights and activations. Our approach identifies optimal affine transformations tailored to each linear layer, calibrated in hours via a lightweight objective. To reduce runtime overhead, we apply Kronecker decomposition to the transformation matrices, and fuse all operations in FlatQuant into a single kernel. Extensive experiments show that FlatQuant sets up a new state-of-the-art quantization benchmark. For instance, it achieves less than $\textbf{1}\%$ accuracy drop for W4A4 quantization on the LLaMA-3-70B model, surpassing SpinQuant by $\textbf{7.5}\%$. For inference latency, FlatQuant reduces the slowdown induced by pre-quantization transformation from 0.26x of QuaRot to merely $\textbf{0.07x}$, bringing up to $\textbf{2.3x}$ speedup for prefill and $\textbf{1.7x}$ speedup for decoding, respectively. Code is available at: \url{https://github.com/ruikangliu/FlatQuant}.
The vector quantization is a widely used method to map continuous representation to discrete space and has important application in tokenization for generative mode, bottlenecking information and many other tasks in machine learning. Vector Quantized Variational Autoencoder (VQ-VAE) is a type of variational autoencoder using discrete embedding as latent. We generalize the technique further, enriching the probabilistic framework with a Gaussian mixture as the underlying generative model. This framework leverages a codebook of latent means and adaptive variances to capture complex data distributions. This principled framework avoids various heuristics and strong assumptions that are needed with the VQ-VAE to address training instability and to improve codebook utilization. This approach integrates the benefits of both discrete and continuous representations within a variational Bayesian framework. Furthermore, by introducing the \textit{Aggregated Categorical Posterior Evidence Lower Bound} (ALBO), we offer a principled alternative optimization objective that aligns variational distributions with the generative model. Our experiments demonstrate that GM-VQ improves codebook utilization and reduces information loss without relying on handcrafted heuristics.
We propose GALA, a novel representation of 3D shapes that (i) excels at capturing and reproducing complex geometry and surface details, (ii) is computationally efficient, and (iii) lends itself to 3D generative modelling with modern, diffusion-based schemes. The key idea of GALA is to exploit both the global sparsity of surfaces within a 3D volume and their local surface properties. Sparsity is promoted by covering only the 3D object boundaries, not empty space, with an ensemble of tree root voxels. Each voxel contains an octree to further limit storage and compute to regions that contain surfaces. Adaptivity is achieved by fitting one local and geometry-aware coordinate frame in each non-empty leaf node. Adjusting the orientation of the local grid, as well as the anisotropic scales of its axes, to the local surface shape greatly increases the amount of detail that can be stored in a given amount of memory, which in turn allows for quantization without loss of quality. With our optimized C++/CUDA implementation, GALA can be fitted to an object in less than 10 seconds. Moreover, the representation can efficiently be flattened and manipulated with transformer networks. We provide a cascaded generation pipeline capable of generating 3D shapes with great geometric detail.
Private deep neural network (DNN) inference based on secure two-party computation (2PC) enables secure privacy protection for both the server and the client. However, existing secure 2PC frameworks suffer from a high inference latency due to enormous communication. As the communication of both linear and non-linear DNN layers reduces with the bit widths of weight and activation, in this paper, we propose PrivQuant, a framework that jointly optimizes the 2PC-based quantized inference protocols and the network quantization algorithm, enabling communication-efficient private inference. PrivQuant proposes DNN architecture-aware optimizations for the 2PC protocols for communication-intensive quantized operators and conducts graph-level operator fusion for communication reduction. Moreover, PrivQuant also develops a communication-aware mixed precision quantization algorithm to improve inference efficiency while maintaining high accuracy. The network/protocol co-optimization enables PrivQuant to outperform prior-art 2PC frameworks. With extensive experiments, we demonstrate PrivQuant reduces communication by $11\times, 2.5\times \mathrm{and}~ 2.8\times$, which results in $8.7\times, 1.8\times ~ \mathrm{and}~ 2.4\times$ latency reduction compared with SiRNN, COINN, and CoPriv, respectively.
Training neural networks is computationally heavy and energy-intensive. Many methodologies were developed to save computational requirements and energy by reducing the precision of network weights at inference time and introducing techniques such as rounding, stochastic rounding, and quantization. However, most of these techniques still require full gradient precision at training time, which makes training such models prohibitive on edge devices. This work presents a novel technique for training neural networks without needing gradients. This enables a training process where all the weights are one or two bits, without any hidden full precision computations. We show that it is possible to train models without gradient-based optimization techniques by identifying erroneous contributions of each neuron towards the expected classification and flipping the relevant bits using logical operations. We tested our method on several standard datasets and achieved performance comparable to corresponding gradient-based baselines with a fraction of the compute power.
Conventional model compression techniques for LLMs address high memory consumption and slow inference challenges but typically require computationally expensive retraining to preserve accuracy. In contrast, one-shot compression methods eliminate retraining cost, but struggle to achieve accuracy comparable to dense models. This paper presents SLIM, a new one-shot compression framework that holistically integrates hardware-friendly quantization, sparsity, and low-rank approximation into a unified process. First, we formulate the quantization process using a probabilistic approach (SLIM-Quant) that enables us to apply uniform quantization. Then, we use an existing one-shot pruning method to apply semi-structured sparsity on top of the quantized weights. Finally, to compensate for the introduced aggregated quantization and sparsity error, we use a novel saliency function with unique invertible and additive features that enables us to mathematically compute the value of low-rank adapters. SLIM improves model accuracy by up to 5.66% (LLaMA-2-7B) for 2:4 sparsity with 4-bit weight quantization, outperforming prior methods. Models compressed with SLIM achieve up to 4.3x and 3.8x on Nvidia RTX3060 and A100 GPUs, respectively. Additionally, they achieve up to 0.23x end-to-end memory reduction in comparison to their dense counterparts. We also propose an optional PEFT recipe that further improves accuracy by up to 1.66% (LLaMA-2-13B) compared to SLIM without fine-tuning.
Language Models (LMs) assign significant attention to the first token, even if it is not semantically important, which is known as attention sink. This phenomenon has been widely adopted in applications such as streaming/long context generation, KV cache optimization, inference acceleration, model quantization, and others. Despite its widespread use, a deep understanding of attention sink in LMs is still lacking. In this work, we first demonstrate that attention sinks exist universally in LMs with various inputs, even in small models. Furthermore, attention sink is observed to emerge during the LM pre-training, motivating us to investigate how optimization, data distribution, loss function, and model architecture in LM pre-training influence its emergence. We highlight that attention sink emerges after effective optimization on sufficient training data. The sink position is highly correlated with the loss function and data distribution. Most importantly, we find that attention sink acts more like key biases, storing extra attention scores, which could be non-informative and not contribute to the value computation. We also observe that this phenomenon (at least partially) stems from tokens' inner dependence on attention scores as a result of softmax normalization. After relaxing such dependence by replacing softmax attention with other attention operations, such as sigmoid attention without normalization, attention sinks do not emerge in LMs up to 1B parameters. The code is available at https://github.com/sail-sg/Attention-Sink.
The ever increasing sizes of Large Language Models (LLMs) beyond hundreds of billions of parameters have generated enormous pressure on the manufacturers of dedicated hardware accelerators and made the innovative design of the latter one of the most rapidly expanding fields of the AI industry. Various approaches have been explored to enable efficient and accurate processing of LLMs on the available accelerators given their computational and storage limitations. Among these, various quantization techniques have become the main focus of the community as a means of reducing the compute, communication and storage requirements. Quantization to lower precision formats naturally poses a number of challenges caused by the limited range of the available value representations. When it comes to processing the popular Transformer models on hardware, one of the main issues becomes calculation of the LayerNorm simply because accumulation of the variance requires a much wider dynamic range than the hardware enables. In this article, we address this matter and propose a computationally-efficient scaling technique that can be easily applied to Transformer models during inference. Our method suggests a straightforward way of scaling the LayerNorm inputs based on the static weights of the immediately preceding linear layers. The scaling factors are computed offline, based solely on the linear layer weights, hence no latency or computational overhead is added during inference. Most importantly, our technique ensures that no numerical issues such as overflow or underflow could happen during the compute. This approach offers smooth, accurate and resource-effective inference across a wide range of hardware architectures. The article provides theoretical justification as well as supporting numerical simulations.
With the rapid growth in the use of fine-tuning for large language models (LLMs), optimizing fine-tuning while keeping inference efficient has become highly important. However, this is a challenging task as it requires improvements in all aspects, including inference speed, fine-tuning speed, memory consumption, and, most importantly, model quality. Previous studies have attempted to achieve this by combining quantization with fine-tuning, but they have failed to enhance all four aspects simultaneously. In this study, we propose a new lightweight technique called Quantization for Efficient Fine-Tuning (QEFT). QEFT accelerates both inference and fine-tuning, is supported by robust theoretical foundations, offers high flexibility, and maintains good hardware compatibility. Our extensive experiments demonstrate that QEFT matches the quality and versatility of full-precision parameter-efficient fine-tuning, while using fewer resources. Our code is available at https://github.com/xvyaward/qeft.
The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottleneck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform sparse attention mechanism solely on those important tokens, reducing the latency in the prefill phase. Tokens deemed less important will be discarded to reduce KV cache size, alleviating the memory bottleneck in the decoding phase. Our experiments demonstrate that ZipVL can accelerate the prefill phase by 2.3$\times$ and improve decoding throughput by 2.8$\times$, with a minimal accuracy reduction of only 0.5\% on VQAv2 benchmark over LLaVA-Next-13B model, effectively enhancing the generation efficiency of LVLMs.
Large language models achieve exceptional performance on various downstream tasks through supervised fine-tuning. However, the diversity of downstream tasks and practical requirements makes deploying multiple full-parameter fine-tuned models challenging. Current methods that compress the delta weight struggle to achieve ultra-high compression, failing to minimize the deployment overhead. To address the above issue, we propose a novel distribution-driven delta compression framework DeltaDQ, which utilizes Group-wise Dropout and Separate Quantization to achieve ultra-high compression for the delta weight. We have observed that the matrix-computed intermediate results for the delta weight exhibit extremely small variance and min-max range characteristics, referred to as Balanced Intermediate Results. Exploiting this phenomenon, we introduce Group-wise Dropout to perform dropout on the delta weight using an optimal group size. Furthermore, using Separate Quantization, sparse weights are quantized and decomposed to achieve a lower bit. Experimental results show that DeltaDQ achieves 16x compression with improved accuracy compared to baselines for WizardMath and WizardCoder models across different parameter scales. Moreover, DeltaDQ demonstrates the ability for ultra-high compression ratio, achieving 128x compression for the WizardMath-7B model and 512x compression for the WizardMath-70B model.
We present the WaveFormer, a neural network architecture based on a linear attention transformer to enable long sequence inference for TinyML devices. Waveformer achieves a new state-of-the-art accuracy of 98.8 % and 99.1 % on the Google Speech V2 keyword spotting (KWS) dataset for the 12 and 35 class problems with only 130 kB of weight storage, compatible with MCU class devices. Top-1 accuracy is improved by 0.1 and 0.9 percentage points while reducing the model size and number of operations by 2.5× and 4.7× compared to the state of the art. We also propose a hardware-friendly 8-bit integer quantization algorithm for the linear attention operator, enabling efficient deployment on low-cost, ultra-low-power microcontrollers without loss of accuracy.
Large language models (LLMs) are primarily designed to understand unstructured text. When directly applied to structured formats such as tabular data, they may struggle to discern inherent relationships and overlook critical patterns. While tabular representation learning methods can address some of these limitations, existing efforts still face challenges with sparse high-cardinality fields, precise numerical reasoning, and column-heavy tables. Furthermore, leveraging these learned representations for downstream tasks through a language based interface is not apparent. In this paper, we present an innovative and scalable solution to these challenges. Concretely, our approach introduces a multi-tier partitioning mechanism that utilizes power-law dynamics to handle large vocabularies, an adaptive quantization mechanism to impose priors on numerical continuity, and a distinct treatment of core-columns and meta-information columns. To facilitate instruction tuning on LLMs, we propose a parameter efficient decoder that interleaves transaction and text modalities using a series of adapter layers, thereby exploiting rich cross-task knowledge. We validate the efficacy of our solution on a large-scale dataset of synthetic payments transactions.
The spatio-temporal complexity of video data presents significant challenges in tasks such as compression, generation, and inpainting. We present four key contributions to address the challenges of spatiotemporal video processing. First, we introduce the 3D Mobile Inverted Vector-Quantization Variational Autoencoder (3D-MBQ-VAE), which combines Variational Autoencoders (VAEs) with masked token modeling to enhance spatiotemporal video compression. The model achieves superior temporal consistency and state-of-the-art (SOTA) reconstruction quality by employing a novel training strategy with full frame masking. Second, we present MotionAura, a text-to-video generation framework that utilizes vector-quantized diffusion models to discretize the latent space and capture complex motion dynamics, producing temporally coherent videos aligned with text prompts. Third, we propose a spectral transformer-based denoising network that processes video data in the frequency domain using the Fourier Transform. This method effectively captures global context and long-range dependencies for high-quality video generation and denoising. Lastly, we introduce a downstream task of Sketch Guided Video Inpainting. This task leverages Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. Our models achieve SOTA performance on a range of benchmarks. Our work offers robust frameworks for spatiotemporal modeling and user-driven video content manipulation. We will release the code, datasets, and models in open-source.
Diffusion models have become the dominant approach for visual generation. They are trained by denoising a Markovian process which gradually adds noise to the input. We argue that the Markovian property limits the model's ability to fully utilize the generation trajectory, leading to inefficiencies during training and inference. In this paper, we propose DART, a transformer-based model that unifies autoregressive (AR) and diffusion within a non-Markovian framework. DART iteratively denoises image patches spatially and spectrally using an AR model that has the same architecture as standard language models. DART does not rely on image quantization, which enables more effective image modeling while maintaining flexibility. Furthermore, DART seamlessly trains with both text and image data in a unified model. Our approach demonstrates competitive performance on class-conditioned and text-to-image generation tasks, offering a scalable, efficient alternative to traditional diffusion models. Through this unified framework, DART sets a new benchmark for scalable, high-quality image synthesis.