As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging, as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays. CacheGen is a fast context-loading module for LLM systems. First, CacheGen uses a custom tensor encoder, leveraging KV cache's distributional properties to encode a KV cache into more compact bitstream representations with negligible decoding overhead, to save bandwidth usage. Second, CacheGen adapts the compression level of different parts of a KV cache to cope with changes in available bandwidth, in order to maintain low context-loading delay and high generation quality. % When available bandwidth drops, CacheGen may raise the compression level for a part of the context or recompute its KV cache on the fly. We test CacheGen on popular LLMs and datasets. Compared to the recent systems that reuse the KV cache, CacheGen reduces the KV cache size by 3.5-4.3x and the total delay in fetching and processing contexts by 3.2-3.7x with negligible impact on the LLM response quality. Our code is at: https://github.com/UChi-JCL/CacheGen.
Large Language Models (LLMs) have showcased remarkable impacts across a wide spectrum of natural language processing tasks. Fine-tuning these pre-trained models on downstream datasets provides further significant performance gains, but this process has been challenging due to its extraordinary resource requirements. To this end, existing efforts focus on parameter-efficient fine-tuning, which, unfortunately, fail to capitalize on the powerful potential of full-parameter fine-tuning. In this work, we propose QFT, a novel Quantized Full-parameter Tuning framework for LLMs that enables memory-efficient fine-tuning without harming performance. Our framework incorporates two novel ideas: (i) we adopt the efficient Lion optimizer, which only keeps track of the momentum and has consistent update magnitudes for each parameter, an inherent advantage for robust quantization; and (ii) we quantize all model states and store them as integer values, and present a gradient flow and parameter update scheme for the quantized weights. As a result, QFT reduces the model state memory to 21% of the standard solution while achieving comparable performance, e.g., tuning a LLaMA-7B model requires only <30GB of memory, satisfied by a single A6000 GPU.
Real-time visual localization often utilizes online computing, for which query images or videos are transmitted to remote servers for visual place recognition (VPR). However, limited network bandwidth necessitates image-quality reduction and thus the degradation of global image descriptors, reducing VPR accuracy. We address this issue at the descriptor extraction level with a knowledge-distillation methodology that learns feature representations from high-quality images to extract more discriminative descriptors from low-quality images. Our approach includes the Inter-channel Correlation Knowledge Distillation (ICKD) loss, Mean Squared Error (MSE) loss, and Triplet loss. We validate the proposed losses on multiple VPR methods and datasets subjected to JPEG compression, resolution reduction, and video quantization. We obtain significant improvements in VPR recall rates under all three tested modalities of lowered image quality. Furthermore, we fill a gap in VPR literature on video-based data and its influence on VPR performance. This work contributes to more reliable place recognition in resource-constrained environments.
Vector-quantized image modeling has shown great potential in synthesizing high-quality images. However, generating high-resolution images remains a challenging task due to the quadratic computational overhead of the self-attention process. In this study, we seek to explore a more efficient two-stage framework for high-resolution image generation with improvements in the following three aspects. (1) Based on the observation that the first quantization stage has solid local property, we employ a local attention-based quantization model instead of the global attention mechanism used in previous methods, leading to better efficiency and reconstruction quality. (2) We emphasize the importance of multi-grained feature interaction during image generation and introduce an efficient attention mechanism that combines global attention (long-range semantic consistency within the whole image) and local attention (fined-grained details). This approach results in faster generation speed, higher generation fidelity, and improved resolution. (3) We propose a new generation pipeline incorporating autoencoding training and autoregressive generation strategy, demonstrating a better paradigm for image synthesis. Extensive experiments demonstrate the superiority of our approach in high-quality and high-resolution image reconstruction and generation.
Large Language Models (LLMs) pose significant hardware challenges related to memory requirements and computational ability. There are two mainstream quantization schemes for LLMs: coarse-grained ($\textit{e.g.,}$ channel-wise) quantization and fine-grained ($\textit{e.g.,}$ group-wise) quantization. Fine-grained quantization has smaller quantization loss, consequently achieving superior performance. However, when applied to weight-activation quantization, it disrupts continuous integer matrix multiplication, leading to inefficient inference. In this paper, we introduce Dual Grained Quantization (DGQ), a novel A8W4 quantization for LLM that maintains superior performance while ensuring fast inference speed. DSQ dequantizes the fine-grained INT4 weight into coarse-grained INT8 representation and preform matrix multiplication using INT8 kernels. Besides, we develop a two-phase grid search algorithm to simplify the determination of fine-grained and coarse-grained quantization scales. We also devise a percentile clipping schema for smoothing the activation outliers without the need for complex optimization techniques. Experimental results demonstrate that DGQ consistently outperforms prior methods across various LLM architectures and a wide range of tasks. Remarkably, by our implemented efficient CUTLASS kernel, we achieve $\textbf{1.12}$ $\times$ memory reduction and $\textbf{3.24}$ $\times$ speed gains comparing A16W4 implementation. These advancements enable efficient deployment of A8W4 LLMs for real-world applications.
Vision Transformer (ViT) architectures represent images as collections of high-dimensional vectorized tokens, each corresponding to a rectangular non-overlapping patch. This representation trades spatial granularity for embedding dimensionality, and results in semantically rich but spatially coarsely quantized feature maps. In order to retrieve spatial details beneficial to fine-grained inference tasks we propose a training-free method inspired by "stochastic resonance". Specifically, we perform sub-token spatial transformations to the input data, and aggregate the resulting ViT features after applying the inverse transformation. The resulting "Stochastic Resonance Transformer" (SRT) retains the rich semantic information of the original representation, but grounds it on a finer-scale spatial domain, partly mitigating the coarse effect of spatial tokenization. SRT is applicable across any layer of any ViT architecture, consistently boosting performance on several tasks including segmentation, classification, depth estimation, and others by up to 14.9% without the need for any fine-tuning.
In this paper, we address the problem of learning a binary (positive vs. negative) classifier given Positive and Unlabeled data commonly referred to as PU learning. Although rudimentary techniques like clustering, out-of-distribution detection, or positive density estimation can be used to solve the problem in low-dimensional settings, their efficacy progressively deteriorates with higher dimensions due to the increasing complexities in the data distribution. In this paper we propose to learn a neural network-based data representation using a loss function that can be used to project the unlabeled data into two (positive and negative) clusters that can be easily identified using simple clustering techniques, effectively emulating the phenomenon observed in low-dimensional settings. We adopt a vector quantization technique for the learned representations to amplify the separation between the learned unlabeled data clusters. We conduct experiments on simulated PU data that demonstrate the improved performance of our proposed method compared to the current state-of-the-art approaches. We also provide some theoretical justification for our two cluster-based approach and our algorithmic choices.
Recently, intelligent reflecting surface (IRS)-assisted communication has gained considerable attention due to its advantage in extending the coverage and compensating the path loss with low-cost passive metasurface. This paper considers the uplink channel estimation for IRS-aided multiuser massive MISO communications with one-bit ADCs at the base station (BS). The use of one-bit ADC is impelled by the low-cost and power efficient implementation of massive antennas techniques. However, the passiveness of IRS and the lack of signal level information after one-bit quantization make the IRS channel estimation challenging. To tackle this problem, we exploit the structured sparsity of the user-IRS-BS cascaded channels and develop three channel estimators, each of which utilizes the structured sparsity at different levels. Specifically, the first estimator exploits the elementwise sparsity of the cascaded channel and employs the sparse Bayesian learning (SBL) to infer the channel responses via the type-II maximum likelihood (ML) estimation. However, due to the one-bit quantization, the type-II ML in general is intractable. As such, a variational expectation-maximization (EM) algorithm is custom-derived to iteratively compute an ML solution. The second estimator utilizes the common row-structured sparsity induced by the IRS-to-BS channel shared among the users, and develops another type-II ML solution via the block SBL (BSBL) and the variational EM. To further improve the performance of BSBL, a third two-stage estimator is proposed, which can utilize both the common row-structured sparsity and the column-structured sparsity arising from the limited scattering around the users. Simulation results show that the more diverse structured sparsity is exploited, the better estimation performance is achieved, and that the proposed estimators are superior to state-of-the-art one-bit estimators.
The information bottleneck auto-encoder is a tool for disentanglement commonly used for voice transformation. The successful disentanglement relies on the right choice of bottleneck size. Previous bottleneck auto-encoders created the bottleneck by the dimension of the latent space or through vector quantization and had no means to change the bottleneck size of a specific model. As the bottleneck removes information from the disentangled representation, the choice of bottleneck size is a trade-off between disentanglement and synthesis quality. We propose to build the information bottleneck using dropout which allows us to change the bottleneck through the dropout rate and investigate adapting the bottleneck size depending on the context. We experimentally explore into using the adaptive bottleneck for pitch transformation and demonstrate that the adaptive bottleneck leads to improved disentanglement of the F0 parameter for both, speech and singing voice leading to improved synthesis quality. Using the variable bottleneck size, we were able to achieve disentanglement for singing voice including extremely high pitches and create a universal voice model, that works on both speech and singing voice with improved synthesis quality.
Optical processors, built with "optical neurons", can efficiently perform high-dimensional linear operations at the speed of light. Thus they are a promising avenue to accelerate large-scale linear computations. With the current advances in micro-fabrication, such optical processors can now be 3D fabricated, but with a limited precision. This limitation translates to quantization of learnable parameters in optical neurons, and should be handled during the design of the optical processor in order to avoid a model mismatch. Specifically, optical neurons should be trained or designed within the physical-constraints at a predefined quantized precision level. To address this critical issues we propose a physics-informed quantization-aware training framework. Our approach accounts for physical constraints during the training process, leading to robust designs. We demonstrate that our approach can design state of the art optical processors using diffractive networks for multiple physics based tasks despite quantized learnable parameters. We thus lay the foundation upon which improved optical processors may be 3D fabricated in the future.
Continual learning is a desirable feature in many modern machine learning applications, which allows in-field adaptation and updating, ranging from accommodating distribution shift, to fine-tuning, and to learning new tasks. For applications with privacy and low latency requirements, the compute and memory demands imposed by continual learning can be cost-prohibitive for resource-constraint edge platforms. Reducing computational precision through fully quantized training (FQT) simultaneously reduces memory footprint and increases compute efficiency for both training and inference. However, aggressive quantization especially integer FQT typically degrades model accuracy to unacceptable levels. In this paper, we propose a technique that leverages inexpensive Hadamard transforms to enable low-precision training with only integer matrix multiplications. We further determine which tensors need stochastic rounding and propose tiled matrix multiplication to enable low-bit width accumulators. We demonstrate the effectiveness of our technique on several human activity recognition datasets and CIFAR100 in a class incremental learning setting. We achieve less than 0.5% and 3% accuracy degradation while we quantize all matrix multiplications inputs down to 4-bits with 8-bit accumulators.
Diffusion models have demonstrated remarkable capabilities in image synthesis and related generative tasks. Nevertheless, their practicality for real-world applications is constrained by substantial computational costs and latency issues. Quantization is a dominant way to compress and accelerate diffusion models, where post-training quantization (PTQ) and quantization-aware training (QAT) are two main approaches, each bearing its own properties. While PTQ exhibits efficiency in terms of both time and data usage, it may lead to diminished performance in low bit-width. On the other hand, QAT can alleviate performance degradation but comes with substantial demands on computational and data resources. In this paper, we introduce a data-free and parameter-efficient fine-tuning framework for low-bit diffusion models, dubbed EfficientDM, to achieve QAT-level performance with PTQ-like efficiency. Specifically, we propose a quantization-aware variant of the low-rank adapter (QALoRA) that can be merged with model weights and jointly quantized to low bit-width. The fine-tuning process distills the denoising capabilities of the full-precision model into its quantized counterpart, eliminating the requirement for training data. We also introduce scale-aware optimization and temporal learned step-size quantization to further enhance performance. Extensive experimental results demonstrate that our method significantly outperforms previous PTQ-based diffusion models while maintaining similar time and data efficiency. Specifically, there is only a 0.05 sFID increase when quantizing both weights and activations of LDM-4 to 4-bit on ImageNet 256x256. Compared to QAT-based methods, our EfficientDM also boasts a 16.2x faster quantization speed with comparable generation quality. Code is available at \href{https://github.com/ThisisBillhe/EfficientDM}{this hrl}.
Quantization has emerged as a promising direction for model compression. Recently, data-free quantization has been widely studied as a promising method to avoid privacy concerns, which synthesizes images as an alternative to real training data. Existing methods use classification loss to ensure the reliability of the synthesized images. Unfortunately, even if these images are well-classified by the pre-trained model, they still suffer from low semantics and homogenization issues. Intuitively, these low-semantic images are sensitive to perturbations, and the pre-trained model tends to have inconsistent output when the generator synthesizes an image with poor semantics. To this end, we propose Robustness-Guided Image Synthesis (RIS), a simple but effective method to enrich the semantics of synthetic images and improve image diversity, further boosting the performance of downstream data-free compression tasks. Concretely, we first introduce perturbations on input and model weight, then define the inconsistency metrics at feature and prediction levels before and after perturbations. On the basis of inconsistency on two levels, we design a robustness optimization objective to enhance the semantics of synthetic images. Moreover, we also make our approach diversity-aware by forcing the generator to synthesize images with small correlations in the label space. With RIS, we achieve state-of-the-art performance for various settings on data-free quantization and can be extended to other data-free compression tasks.
Large Mixture of Experts (MoE) models could achieve state-of-the-art quality on various language tasks, including machine translation task, thanks to the efficient model scaling capability with expert parallelism. However, it has brought a fundamental issue of larger memory consumption and increased memory bandwidth bottleneck at deployment time. In this paper, we propose Mixture of Quantized Experts (MoQE) which is a simple weight-only quantization method applying ultra low-bit down to 2-bit quantizations only to expert weights for mitigating the increased memory and latency issues of MoE models. We show that low-bit quantization together with the MoE architecture delivers a reliable model performance while reducing the memory size significantly even without any additional training in most cases. In particular, expert layers in MoE models are much more robust to the quantization than conventional feedforward networks (FFN) layers. In our comprehensive analysis, we show that MoE models with 2-bit expert weights can deliver better model performance than the dense model trained on the same dataset. As a result of low-bit quantization, we show the model size can be reduced by 79.6% of the original half precision floating point (fp16) MoE model. Combined with an optimized GPU runtime implementation, it also achieves 1.24X speed-up on A100 GPUs.
We present Junk DNA Hypothesis by adopting a novel task-centric angle for the pre-trained weights of large language models (LLMs). It has been believed that weights in LLMs contain significant redundancy, leading to the conception that a considerable chunk of the parameters can be removed by pruning without compromising performance. Contrary to this belief, this paper presents a counter-argument: small-magnitude weights of pre-trained model weights encode vital knowledge essential for tackling difficult downstream tasks - manifested as the monotonic relationship between the performance drop of downstream tasks across the difficulty spectrum, as we prune more pre-trained weights by magnitude. Moreover, we reveal that these seemingly inconsequential weights can result in irreparable loss of knowledge and performance degradation in difficult tasks, even when downstream continual training is allowed. Interestingly, our evaluations show that the other popular compression, namely quantization, fails to exhibit similar monotonic effect and does not as convincingly disentangle this task-difficulty information. To study formally, we introduce several quantifiable metrics to gauge the downstream task difficulty: (1) within the same task category, and (2) across different task categories. Our extensive experiments substantiate the Junk DNA Hypothesis across a diverse range of model sizes, tasks, datasets, and even pruning methods. Codes are available at: https://github.com/VITA-Group/Junk_DNA_Hypothesis.git.
Vector Quantization (VQ) is a well-known technique in deep learning for extracting informative discrete latent representations. VQ-embedded models have shown impressive results in a range of applications including image and speech generation. VQ operates as a parametric K-means algorithm that quantizes inputs using a single codebook vector in the forward pass. While powerful, this technique faces practical challenges including codebook collapse, non-differentiability and lossy compression. To mitigate the aforementioned issues, we propose Soft Convex Quantization (SCQ) as a direct substitute for VQ. SCQ works like a differentiable convex optimization (DCO) layer: in the forward pass, we solve for the optimal convex combination of codebook vectors that quantize the inputs. In the backward pass, we leverage differentiability through the optimality conditions of the forward solution. We then introduce a scalable relaxation of the SCQ optimization and demonstrate its efficacy on the CIFAR-10, GTSRB and LSUN datasets. We train powerful SCQ autoencoder models that significantly outperform matched VQ-based architectures, observing an order of magnitude better image reconstruction and codebook usage with comparable quantization runtime.
Reinforcement learning (RL) is a promising approach for solving robotic manipulation tasks. However, it is challenging to apply the RL algorithms directly in the real world. For one thing, RL is data-intensive and typically requires millions of interactions with environments, which are impractical in real scenarios. For another, it is necessary to make heavy engineering efforts to design reward functions manually. To address these issues, we leverage foundation models in this paper. We propose Reinforcement Learning with Foundation Priors (RLFP) to utilize guidance and feedback from policy, value, and success-reward foundation models. Within this framework, we introduce the Foundation-guided Actor-Critic (FAC) algorithm, which enables embodied agents to explore more efficiently with automatic reward functions. The benefits of our framework are threefold: (1) \textit{sample efficient}; (2) \textit{minimal and effective reward engineering}; (3) \textit{agnostic to foundation model forms and robust to noisy priors}. Our method achieves remarkable performances in various manipulation tasks on both real robots and in simulation. Across 5 dexterous tasks with real robots, FAC achieves an average success rate of 86\% after one hour of real-time learning. Across 8 tasks in the simulated Meta-world, FAC achieves 100\% success rates in 7/8 tasks under less than 100k frames (about 1-hour training), outperforming baseline methods with manual-designed rewards in 1M frames. We believe the RLFP framework can enable future robots to explore and learn autonomously in the physical world for more tasks. Visualizations and code are available at \url{https://yewr.github.io/rlfp}.
The field of Artificial Intelligence has witnessed remarkable progress in recent years, especially with the emergence of powerful large language models (LLMs) based on the transformer architecture. Cloud-based LLMs, such as OpenAI's ChatGPT, offer impressive capabilities but come with concerns regarding latency and privacy due to network dependencies. This article presents an innovative approach to LLM inference, envisioning a future where LLMs with billions of parameters can be executed directly on mobile devices without network connectivity. The article showcases a fine-tuned GPT LLM with 3 billion parameters that can operate smoothly on devices with as low as 4GB of memory. Through the integration of native code and model quantization techniques, the application not only serves as a general-purpose assistant but also facilitates seamless mobile interactions with text-to-actions features. The article provides insights into the training pipeline, implementation details, test results, and future directions of on-device LLM inference. This breakthrough technology opens up possibilities for empowering users with sophisticated AI capabilities while preserving their privacy and eliminating latency concerns.
Symbolic systems are powerful frameworks for modeling cognitive processes as they encapsulate the rules and relationships fundamental to many aspects of human reasoning and behavior. Central to these models are systematicity, compositionality, and productivity, making them invaluable in both cognitive science and artificial intelligence. However, certain limitations remain. For instance, the integration of structured symbolic processes and latent sub-symbolic processes has been implemented at the computational level through fiat methods such as quantization or softmax sampling, which assume, rather than derive, the operations underpinning discretization and symbolicization. In this work, we introduce a novel neural stochastic dynamical systems model that integrates attractor dynamics with symbolic representations to model cognitive processes akin to the probabilistic language of thought (PLoT). Our model segments the continuous representational space into discrete basins, with attractor states corresponding to symbolic sequences, that reflect the semanticity and compositionality characteristic of symbolic systems through unsupervised learning, rather than relying on pre-defined primitives. Moreover, like PLoT, our model learns to sample a diverse distribution of attractor states that reflect the mutual information between the input data and the symbolic encodings. This approach establishes a unified framework that integrates both symbolic and sub-symbolic processing through neural dynamics, a neuro-plausible substrate with proven expressivity in AI, offering a more comprehensive model that mirrors the complex duality of cognitive operations.
Neural video codecs have recently become competitive with standard codecs such as HEVC in the low-delay setting. However, most neural codecs are large floating-point networks that use pixel-dense warping operations for temporal modeling, making them too computationally expensive for deployment on mobile devices. Recent work has demonstrated that running a neural decoder in real time on mobile is feasible, but shows this only for 720p RGB video. This work presents the first neural video codec that decodes 1080p YUV420 video in real time on a mobile device. Our codec relies on two major contributions. First, we design an efficient codec that uses a block-based motion compensation algorithm available on the warping core of the mobile accelerator, and we show how to quantize this model to integer precision. Second, we implement a fast decoder pipeline that concurrently runs neural network components on the neural signal processor, parallel entropy coding on the mobile GPU, and warping on the warping core. Our codec outperforms the previous on-device codec by a large margin with up to 48% BD-rate savings, while reducing the MAC count on the receiver side by $10 \times$. We perform a careful ablation to demonstrate the effect of the introduced motion compensation scheme, and ablate the effect of model quantization.