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}.
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.
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.
Audiovisual segmentation (AVS) is a challenging task that aims to segment visual objects in videos according to their associated acoustic cues. With multiple sound sources and background disturbances involved, establishing robust correspondences between audio and visual contents poses unique challenges due to (1) complex entanglement across sound sources and (2) frequent changes in the occurrence of distinct sound events. Assuming sound events occur independently, the multi-source semantic space can be represented as the Cartesian product of single-source sub-spaces. We are motivated to decompose the multi-source audio semantics into single-source semantics for more effective interactions with visual content. We propose a semantic decomposition method based on product quantization, where the multi-source semantics can be decomposed and represented by several disentangled and noise-suppressed single-source semantics. Furthermore, we introduce a global-to-local quantization mechanism, which distills knowledge from stable global (clip-level) features into local (frame-level) ones, to handle frequent changes in audio semantics. Extensive experiments demonstrate that our semantically decomposed audio representation significantly improves AVS performance, e.g., +21.2% mIoU on the challenging AVS-Semantic benchmark with ResNet50 backbone. https://github.com/lxa9867/QSD.
Despite their remarkable achievements, modern Large Language Models (LLMs) face exorbitant computational and memory footprints. Recently, several works have shown significant success in training-free and data-free compression (pruning and quantization) of LLMs that achieve 50 - 60% sparsity and reduce the bit width to 3 or 4 bits per weight, with negligible degradation of perplexity over the uncompressed baseline. As recent research efforts are focused on developing increasingly sophisticated compression methods, our work takes a step back and re-evaluates the effectiveness of existing SoTA compression methods, which rely on a fairly simple and widely questioned metric, perplexity (even for dense LLMs). We introduce Knowledge-Intensive Compressed LLM BenchmarK (LLM-KICK), a collection of carefully curated tasks to redefine the evaluation protocol for compressed LLMs, which have significant alignment with their dense counterparts and perplexity fail to capture subtle change in their true capabilities. LLM-KICK unveils many favorable merits and unfortunate plights of current SoTA compression methods: all pruning methods suffer significant performance degradation, sometimes at trivial sparsity ratios (e.g., 25-30%), and fail for N:M sparsity in knowledge-intensive tasks; current quantization methods are more successful than pruning; yet, pruned LLMs even at $\geq 50$% sparsity are robust in-context retrieval and summarization systems; among others. LLM-KICK is designed to holistically access compressed LLMs' ability for language understanding, reasoning, generation, in-context retrieval, in-context summarization, etc. We hope our study can foster the development of better LLM compression methods. The reproduced codes are available at https://github.com/VITA-Group/llm-kick.
This paper explores network binarization, a radical form of quantization, compressing model weights to a single bit, specifically for Large Language Models (LLMs) compression. Due to previous binarization methods collapsing LLMs, we propose a novel approach, Partially-Binarized LLM (PB-LLM), which can achieve extreme low-bit quantization while maintaining the linguistic reasoning capacity of quantized LLMs. Specifically, our exploration first uncovers the ineffectiveness of naive applications of existing binarization algorithms and highlights the imperative role of salient weights in achieving low-bit quantization. Thus, PB-LLM filters a small ratio of salient weights during binarization, allocating them to higher-bit storage, i.e., partially-binarization. PB-LLM is extended to recover the capacities of quantized LMMs, by analyzing from the perspective of post-training quantization (PTQ) and quantization-aware training (QAT). Under PTQ, combining the concepts from GPTQ, we reconstruct the binarized weight matrix guided by the Hessian matrix and successfully recover the reasoning capacity of PB-LLM in low-bit. Under QAT, we freeze the salient weights during training, explore the derivation of optimal scaling factors crucial for minimizing the quantization error, and propose a scaling mechanism based on this derived scaling strategy for residual binarized weights. Those explorations and the developed methodologies significantly contribute to rejuvenating the performance of low-bit quantized LLMs and present substantial advancements in the field of network binarization for LLMs.The code is available at https://github.com/hahnyuan/BinaryLLM.
COMpression with Bayesian Implicit NEural Representations (COMBINER) is a recent data compression method that addresses a key inefficiency of previous Implicit Neural Representation (INR)-based approaches: it avoids quantization and enables direct optimization of the rate-distortion performance. However, COMBINER still has significant limitations: 1) it uses factorized priors and posterior approximations that lack flexibility; 2) it cannot effectively adapt to local deviations from global patterns in the data; and 3) its performance can be susceptible to modeling choices and the variational parameters' initializations. Our proposed method, Robust and Enhanced COMBINER (RECOMBINER), addresses these issues by 1) enriching the variational approximation while retaining a low computational cost via a linear reparameterization of the INR weights, 2) augmenting our INRs with learnable positional encodings that enable them to adapt to local details and 3) splitting high-resolution data into patches to increase robustness and utilizing expressive hierarchical priors to capture dependency across patches. We conduct extensive experiments across several data modalities, showcasing that RECOMBINER achieves competitive results with the best INR-based methods and even outperforms autoencoder-based codecs on low-resolution images at low bitrates. Our PyTorch implementation is available at https://github.com/cambridge-mlg/RECOMBINER/.
In this work, we establish a direct connection between generative diffusion models (DMs) and stochastic quantization (SQ). The DM is realized by approximating the reversal of a stochastic process dictated by the Langevin equation, generating samples from a prior distribution to effectively mimic the target distribution. Using numerical simulations, we demonstrate that the DM can serve as a global sampler for generating quantum lattice field configurations in two-dimensional $\phi^4$ theory. We demonstrate that DMs can notably reduce autocorrelation times in the Markov chain, especially in the critical region where standard Markov Chain Monte-Carlo (MCMC) algorithms experience critical slowing down. The findings can potentially inspire further advancements in lattice field theory simulations, in particular in cases where it is expensive to generate large ensembles.
Accurate localization of cephalometric landmarks holds great importance in the fields of orthodontics and orthognathics due to its potential for automating key point labeling. In the context of landmark detection, particularly in cephalometrics, it has been observed that existing methods often lack standardized pipelines and well-designed bias reduction processes, which significantly impact their performance. In this paper, we revisit a related task, human pose estimation (HPE), which shares numerous similarities with cephalometric landmark detection (CLD), and emphasize the potential for transferring techniques from the former field to benefit the latter. Motivated by this insight, we have developed a robust and adaptable benchmark based on the well-established HPE codebase known as MMPose. This benchmark can serve as a dependable baseline for achieving exceptional CLD performance. Furthermore, we introduce an upscaling design within the framework to further enhance performance. This enhancement involves the incorporation of a lightweight and efficient super-resolution module, which generates heatmap predictions on high-resolution features and leads to further performance refinement, benefiting from its ability to reduce quantization bias. In the MICCAI CLDetection2023 challenge, our method achieves 1st place ranking on three metrics and 3rd place on the remaining one. The code for our method is available at https://github.com/5k5000/CLdetection2023.
We introduce Transformer-VQ, a decoder-only transformer computing softmax-based dense self-attention in linear time. Transformer-VQ's efficient attention is enabled by vector-quantized keys and a novel caching mechanism. In our large-scale experiments, Transformer-VQ is shown highly competitive in quality, obtaining 0.99 bpb on Enwik8, 26.6 ppl on PG-19, and 3.16 bpb on ImageNet64. In addition, the optimized implementation of Transformer-VQ is over 3x faster than a comparable quadratic-time transformer at sequence length 8k, is over 12x faster at 32k, and can scale to 131k with similar throughput. Code available: \url{https://github.com/transformer-vq/transformer_vq}
We propose a memory-efficient finetuning algorithm for large language models (LLMs) that supports finetuning LLMs with 65B parameters in 2/3/4-bit precision on as little as one 24GB GPU. Our method, modular low-rank adaptation (ModuLoRA), integrates any user-specified weight quantizer with finetuning via low-rank adapters (LoRAs). Our approach relies on a simple quantization-agnostic backward pass that adaptively materializes low-precision LLM weights from a custom black-box quantization module. This approach enables finetuning 2-bit and 3-bit LLMs for the first time -- leveraging state-of-the-art 2-bit QuIP\# quantization and 3-bit OPTQ quantization -- outperforming finetuning that relies on less sophisticated 4-bit and 8-bit methods. In our experiments, \lplora~attains competitive performance on text classification, natural language inference, and instruction following tasks using significantly less memory than existing approaches, and we also surpass the state-of-the-art ROUGE score on a popular summarization task. We release \lplora~together with a series of low-precision models as part of \llmtune, a user-friendly library for quantizing, running, and finetuning LLMs on consumer GPUs.
Large Language Models (LLMs) have recently demonstrated remarkable success across various tasks. However, efficiently serving LLMs has been a challenge due to the large memory bottleneck, specifically in small batch inference settings (e.g. mobile devices). Weight-only quantization can be a promising approach, but sub-4 bit quantization remains a challenge due to large-magnitude activation outliers. To mitigate the undesirable outlier effect, we first propose per-IC quantization, a simple yet effective method that creates quantization groups within each input channel (IC) rather than the conventional per-output-channel (per-OC). Our method is motivated by the observation that activation outliers affect the input dimension of the weight matrix, so similarly grouping the weights in the IC direction can isolate outliers within a group. We also find that activation outliers do not dictate quantization difficulty, and inherent weight sensitivities also exist. With per-IC quantization as a new outlier-friendly scheme, we propose Adaptive Dimensions (AdaDim), a versatile quantization framework that can adapt to various weight sensitivity patterns. We demonstrate the effectiveness of AdaDim by augmenting prior methods such as Round-To-Nearest and GPTQ, showing significant improvements across various language modeling benchmarks for both base (up to +4.7% on MMLU) and instruction-tuned (up to +10% on HumanEval) LLMs. Code is available at https://github.com/johnheo/adadim-llm
We propose to replace vector quantization (VQ) in the latent representation of VQ-VAEs with a simple scheme termed finite scalar quantization (FSQ), where we project the VAE representation down to a few dimensions (typically less than 10). Each dimension is quantized to a small set of fixed values, leading to an (implicit) codebook given by the product of these sets. By appropriately choosing the number of dimensions and values each dimension can take, we obtain the same codebook size as in VQ. On top of such discrete representations, we can train the same models that have been trained on VQ-VAE representations. For example, autoregressive and masked transformer models for image generation, multimodal generation, and dense prediction computer vision tasks. Concretely, we employ FSQ with MaskGIT for image generation, and with UViM for depth estimation, colorization, and panoptic segmentation. Despite the much simpler design of FSQ, we obtain competitive performance in all these tasks. We emphasize that FSQ does not suffer from codebook collapse and does not need the complex machinery employed in VQ (commitment losses, codebook reseeding, code splitting, entropy penalties, etc.) to learn expressive discrete representations.
Recent advances in deep learning methods such as LLMs and Diffusion models have created a need for improved quantization methods that can meet the computational demands of these modern architectures while maintaining accuracy. Towards this goal, we study the advantages of FP8 data formats for post-training quantization across 75 unique network architectures covering a wide range of tasks, including machine translation, language modeling, text generation, image classification, generation, and segmentation. We examine three different FP8 representations (E5M2, E4M3, and E3M4) to study the effects of varying degrees of trade-off between dynamic range and precision on model accuracy. Based on our extensive study, we developed a quantization workflow that generalizes across different network architectures. Our empirical results show that FP8 formats outperform INT8 in multiple aspects, including workload coverage (92.64% vs. 65.87%), model accuracy and suitability for a broader range of operations. Furthermore, our findings suggest that E4M3 is better suited for NLP models, whereas E3M4 performs marginally better than E4M3 on computer vision tasks. The code is publicly available on Intel Neural Compressor: https://github.com/intel/neural-compressor.
Learned B-frame video compression aims to adopt bi-directional motion estimation and motion compensation (MEMC) coding for middle frame reconstruction. However, previous learned approaches often directly extend neural P-frame codecs to B-frame relying on bi-directional optical-flow estimation or video frame interpolation. They suffer from inaccurate quantized motions and inefficient motion compensation. To address these issues, we propose a simple yet effective structure called Interpolation-driven B-frame Video Compression (IBVC). Our approach only involves two major operations: video frame interpolation and artifact reduction compression. IBVC introduces a bit-rate free MEMC based on interpolation, which avoids optical-flow quantization and additional compression distortions. Later, to reduce duplicate bit-rate consumption and focus on unaligned artifacts, a residual guided masking encoder is deployed to adaptively select the meaningful contexts with interpolated multi-scale dependencies. In addition, a conditional spatio-temporal decoder is proposed to eliminate location errors and artifacts instead of using MEMC coding in other methods. The experimental results on B-frame coding demonstrate that IBVC has significant improvements compared to the relevant state-of-the-art methods. Meanwhile, our approach can save bit rates compared with the random access (RA) configuration of H.266 (VTM). The code will be available at https://github.com/ruhig6/IBVC.
Model quantization, which aims to compress deep neural networks and accelerate inference speed, has greatly facilitated the development of cumbersome models on mobile and edge devices. There is a common assumption in quantization methods from prior works that training data is available. In practice, however, this assumption cannot always be fulfilled due to reasons of privacy and security, rendering these methods inapplicable in real-life situations. Thus, data-free network quantization has recently received significant attention in neural network compression. Causal reasoning provides an intuitive way to model causal relationships to eliminate data-driven correlations, making causality an essential component of analyzing data-free problems. However, causal formulations of data-free quantization are inadequate in the literature. To bridge this gap, we construct a causal graph to model the data generation and discrepancy reduction between the pre-trained and quantized models. Inspired by the causal understanding, we propose the Causality-guided Data-free Network Quantization method, Causal-DFQ, to eliminate the reliance on data via approaching an equilibrium of causality-driven intervened distributions. Specifically, we design a content-style-decoupled generator, synthesizing images conditioned on the relevant and irrelevant factors; then we propose a discrepancy reduction loss to align the intervened distributions of the pre-trained and quantized models. It is worth noting that our work is the first attempt towards introducing causality to data-free quantization problem. Extensive experiments demonstrate the efficacy of Causal-DFQ. The code is available at https://github.com/42Shawn/Causal-DFQ.
Weight-sharing quantization has emerged as a technique to reduce energy expenditure during inference in large neural networks by constraining their weights to a limited set of values. However, existing methods for weight-sharing quantization often make assumptions about the treatment of weights based on value alone that neglect the unique role weight position plays. This paper proposes a probabilistic framework based on Bayesian neural networks (BNNs) and a variational relaxation to identify which weights can be moved to which cluster centre and to what degree based on their individual position-specific learned uncertainty distributions. We introduce a new initialisation setting and a regularisation term which allow for the training of BNNs under complex dataset-model combinations. By leveraging the flexibility of weight values captured through a probability distribution, we enhance noise resilience and downstream compressibility. Our iterative clustering procedure demonstrates superior compressibility and higher accuracy compared to state-of-the-art methods on both ResNet models and the more complex transformer-based architectures. In particular, our method outperforms the state-of-the-art quantization method top-1 accuracy by 1.6% on ImageNet using DeiT-Tiny, with its 5 million+ weights now represented by only 296 unique values.
Knowledge distillation (KD) compresses deep neural networks by transferring task-related knowledge from cumbersome pre-trained teacher models to compact student models. However, current KD methods for super-resolution (SR) networks overlook the nature of SR task that the outputs of the teacher model are noisy approximations to the ground-truth distribution of high-quality images (GT), which shades the teacher model's knowledge to result in limited KD effects. To utilize the teacher model beyond the GT upper-bound, we present the Data Upcycling Knowledge Distillation (DUKD), to transfer the teacher model's knowledge to the student model through the upcycled in-domain data derived from training data. Besides, we impose label consistency regularization to KD for SR by the paired invertible augmentations to improve the student model's performance and robustness. Comprehensive experiments demonstrate that the DUKD method significantly outperforms previous arts on several SR tasks.
Understanding indoor scenes is crucial for urban studies. Considering the dynamic nature of indoor environments, effective semantic segmentation requires both real-time operation and high accuracy.To address this, we propose AsymFormer, a novel network that improves real-time semantic segmentation accuracy using RGB-D multi-modal information without substantially increasing network complexity. AsymFormer uses an asymmetrical backbone for multimodal feature extraction, reducing redundant parameters by optimizing computational resource distribution. To fuse asymmetric multimodal features, a Local Attention-Guided Feature Selection (LAFS) module is used to selectively fuse features from different modalities by leveraging their dependencies. Subsequently, a Cross-Modal Attention-Guided Feature Correlation Embedding (CMA) module is introduced to further extract cross-modal representations. The AsymFormer demonstrates competitive results with 54.1% mIoU on NYUv2 and 49.1% mIoU on SUNRGBD. Notably, AsymFormer achieves an inference speed of 65 FPS (79 FPS after implementing mixed precision quantization) on RTX3090, demonstrating that AsymFormer can strike a balance between high accuracy and efficiency.
Efficient training of large-scale graph neural networks (GNNs) has been studied with a specific focus on reducing their memory consumption. Work by Liu et al. (2022) proposed extreme activation compression (EXACT) which demonstrated drastic reduction in memory consumption by performing quantization of the intermediate activation maps down to using INT2 precision. They showed little to no reduction in performance while achieving large reductions in GPU memory consumption. In this work, we present an improvement to the EXACT strategy by using block-wise quantization of the intermediate activation maps. We experimentally analyze different block sizes and show further reduction in memory consumption (>15%), and runtime speedup per epoch (about 5%) even when performing extreme extents of quantization with similar performance trade-offs as with the original EXACT. Further, we present a correction to the assumptions on the distribution of intermediate activation maps in EXACT (assumed to be uniform) and show improved variance estimations of the quantization and dequantization steps.