This paper introduces a highly efficient greedy decoding algorithm for Transducer-based speech recognition models. We redesign the standard nested-loop design for RNN-T decoding, swapping loops over frames and labels: the outer loop iterates over labels, while the inner loop iterates over frames searching for the next non-blank symbol. Additionally, we represent partial hypotheses in a special structure using CUDA tensors, supporting parallelized hypotheses manipulations. Experiments show that the label-looping algorithm is up to 2.0X faster than conventional batched decoding when using batch size 32. It can be further combined with other compiler or GPU call-related techniques to achieve even more speedup. Our algorithm is general-purpose and can work with both conventional Transducers and Token-and-Duration Transducers. We open-source our implementation to benefit the research community.
Parameter-Efficient Fine-tuning (PEFT) facilitates the fine-tuning of Large Language Models (LLMs) under limited resources. However, the fine-tuning performance with PEFT on complex, knowledge-intensive tasks is limited due to the constrained model capacity, which originates from the limited number of additional trainable parameters. To overcome this limitation, we introduce a novel mechanism that fine-tunes LLMs with adapters of larger size yet memory-efficient. This is achieved by leveraging the inherent activation sparsity in the Feed-Forward Networks (FFNs) of LLMs and utilizing the larger capacity of Central Processing Unit (CPU) memory compared to Graphics Processing Unit (GPU). We store and update the parameters of larger adapters on the CPU. Moreover, we employ a Mixture of Experts (MoE)-like architecture to mitigate unnecessary CPU computations and reduce the communication volume between the GPU and CPU. This is particularly beneficial over the limited bandwidth of PCI Express (PCIe). Our method can achieve fine-tuning results comparable to those obtained with larger memory capacities, even when operating under more limited resources such as a 24GB memory single GPU setup, with acceptable loss in training efficiency. Our codes are available at https://github.com/CURRENTF/MEFT.
Neural operators effectively solve PDE problems from data without knowing the explicit equations, which learn the map from the input sequences of observed samples to the predicted values. Most existing works build the model in the original geometric space, leading to high computational costs when the number of sample points is large. We present the Latent Neural Operator (LNO) solving PDEs in the latent space. In particular, we first propose Physics-Cross-Attention (PhCA) transforming representation from the geometric space to the latent space, then learn the operator in the latent space, and finally recover the real-world geometric space via the inverse PhCA map. Our model retains flexibility that can decode values in any position not limited to locations defined in the training set, and therefore can naturally perform interpolation and extrapolation tasks particularly useful for inverse problems. Moreover, the proposed LNO improves both prediction accuracy and computational efficiency. Experiments show that LNO reduces the GPU memory by 50%, speeds up training 1.8 times, and reaches state-of-the-art accuracy on four out of six benchmarks for forward problems and a benchmark for inverse problem. Code is available at https://github.com/L-I-M-I-T/LatentNeuralOperator.
Quality-Diversity (QD) algorithms have emerged as a powerful optimization paradigm with the aim of generating a set of high-quality and diverse solutions. To achieve such a challenging goal, QD algorithms require maintaining a large archive and a large population in each iteration, which brings two main issues, sample and resource efficiency. Most advanced QD algorithms focus on improving the sample efficiency, while the resource efficiency is overlooked to some extent. Particularly, the resource overhead during the training process has not been touched yet, hindering the wider application of QD algorithms. In this paper, we highlight this important research question, i.e., how to efficiently train QD algorithms with limited resources, and propose a novel and effective method called RefQD to address it. RefQD decomposes a neural network into representation and decision parts, and shares the representation part with all decision parts in the archive to reduce the resource overhead. It also employs a series of strategies to address the mismatch issue between the old decision parts and the newly updated representation part. Experiments on different types of tasks from small to large resource consumption demonstrate the excellent performance of RefQD: it not only uses significantly fewer resources (e.g., 16\% GPU memories on QDax and 3.7\% on Atari) but also achieves comparable or better performance compared to sample-efficient QD algorithms. Our code is available at \url{https://github.com/lamda-bbo/RefQD}.
In this paper, we propose Flash3D, a method for scene reconstruction and novel view synthesis from a single image which is both very generalisable and efficient. For generalisability, we start from a "foundation" model for monocular depth estimation and extend it to a full 3D shape and appearance reconstructor. For efficiency, we base this extension on feed-forward Gaussian Splatting. Specifically, we predict a first layer of 3D Gaussians at the predicted depth, and then add additional layers of Gaussians that are offset in space, allowing the model to complete the reconstruction behind occlusions and truncations. Flash3D is very efficient, trainable on a single GPU in a day, and thus accessible to most researchers. It achieves state-of-the-art results when trained and tested on RealEstate10k. When transferred to unseen datasets like NYU it outperforms competitors by a large margin. More impressively, when transferred to KITTI, Flash3D achieves better PSNR than methods trained specifically on that dataset. In some instances, it even outperforms recent methods that use multiple views as input. Code, models, demo, and more results are available at https://www.robots.ox.ac.uk/~vgg/research/flash3d/.
Training LLMs relies on distributed implementations using multiple GPUs to compute gradients in parallel with sharded optimizers. However, synchronizing gradients in data parallel setups introduces communication overhead that grows with the number of workers, limiting parallelization efficiency. Local optimization algorithms reduce communications but incur high memory costs as they prevent optimizer state sharding, hindering scalability. To address this, we propose \textbf{AC}cumulate while \textbf{CO}mmunicate (\acco), a memory-efficient optimization algorithm for distributed LLM training. By synchronizing delayed gradients while computing new ones, \acco~reduces GPU idle time and supports heterogeneous hardware. To mitigate the convergence issues caused by delayed updates, we introduce a novel technique ensuring training dynamics align with standard distributed optimization. Compared to ZeRO-1, our approach is significantly faster and scales effectively across heterogeneous hardware.
The majority of existing speech bandwidth extension (BWE) methods operate under the constraint of fixed source and target sampling rates, which limits their flexibility in practical applications. In this paper, we propose a multi-stage speech BWE model named MS-BWE, which can handle a set of source and target sampling rate pairs and achieve flexible extensions of frequency bandwidth. The proposed MS-BWE model comprises a cascade of BWE blocks, with each block featuring a dual-stream architecture to realize amplitude and phase extension, progressively painting the speech frequency bands stage by stage. The teacher-forcing strategy is employed to mitigate the discrepancy between training and inference. Experimental results demonstrate that our proposed MS-BWE is comparable to state-of-the-art speech BWE methods in speech quality. Regarding generation efficiency, the one-stage generation of MS-BWE can achieve over one thousand times real-time on GPU and about sixty times on CPU.
Recent advancements in Latent Diffusion Models (LDMs) have propelled them to the forefront of various generative tasks. However, their iterative sampling process poses a significant computational burden, resulting in slow generation speeds and limiting their application in text-to-audio generation deployment. In this work, we introduce AudioLCM, a novel consistency-based model tailored for efficient and high-quality text-to-audio generation. AudioLCM integrates Consistency Models into the generation process, facilitating rapid inference through a mapping from any point at any time step to the trajectory's initial point. To overcome the convergence issue inherent in LDMs with reduced sample iterations, we propose the Guided Latent Consistency Distillation with a multi-step Ordinary Differential Equation (ODE) solver. This innovation shortens the time schedule from thousands to dozens of steps while maintaining sample quality, thereby achieving fast convergence and high-quality generation. Furthermore, to optimize the performance of transformer-based neural network architectures, we integrate the advanced techniques pioneered by LLaMA into the foundational framework of transformers. This architecture supports stable and efficient training, ensuring robust performance in text-to-audio synthesis. Experimental results on text-to-sound generation and text-to-music synthesis tasks demonstrate that AudioLCM needs only 2 iterations to synthesize high-fidelity audios, while it maintains sample quality competitive with state-of-the-art models using hundreds of steps. AudioLCM enables a sampling speed of 333x faster than real-time on a single NVIDIA 4090Ti GPU, making generative models practically applicable to text-to-audio generation deployment. Our extensive preliminary analysis shows that each design in AudioLCM is effective.
Due to the high similarity of disparity between consecutive frames in video sequences, the area where disparity changes is defined as the residual map, which can be calculated. Based on this, we propose RecSM, a network based on residual estimation with a flexible recursive structure for video stereo matching. The RecSM network accelerates stereo matching using a Multi-scale Residual Estimation Module (MREM), which employs the temporal context as a reference and rapidly calculates the disparity for the current frame by computing only the residual values between the current and previous frames. To further reduce the error of estimated disparities, we use the Disparity Optimization Module (DOM) and Temporal Attention Module (TAM) to enforce constraints between each module, and together with MREM, form a flexible Stackable Computation Structure (SCS), which allows for the design of different numbers of SCS based on practical scenarios. Experimental results demonstrate that with a stack count of 3, RecSM achieves a 4x speed improvement compared to ACVNet, running at 0.054 seconds based on one NVIDIA RTX 2080TI GPU, with an accuracy decrease of only 0.7%. Code is available at https://github.com/Y0uchenZ/RecSM.
Visual anomaly detection aims to identify anomalous regions in images through unsupervised learning paradigms, with increasing application demand and value in fields such as industrial inspection and medical lesion detection. Despite significant progress in recent years, there is a lack of comprehensive benchmarks to adequately evaluate the performance of various mainstream methods across different datasets under the practical multi-class setting. The absence of standardized experimental setups can lead to potential biases in training epochs, resolution, and metric results, resulting in erroneous conclusions. This paper addresses this issue by proposing a comprehensive visual anomaly detection benchmark, ADer, which is a modular framework that is highly extensible for new methods. The benchmark includes multiple datasets from industrial and medical domains, implementing fifteen state-of-the-art methods and nine comprehensive metrics. Additionally, we have proposed the GPU-assisted ADEval package to address the slow evaluation problem of metrics like time-consuming mAU-PRO on large-scale data, significantly reducing evaluation time by more than \textit{1000-fold}. Through extensive experimental results, we objectively reveal the strengths and weaknesses of different methods and provide insights into the challenges and future directions of multi-class visual anomaly detection. We hope that ADer will become a valuable resource for researchers and practitioners in the field, promoting the development of more robust and generalizable anomaly detection systems. Full codes are open-sourced at https://github.com/zhangzjn/ader.
Matrix multiplication (MatMul) typically dominates the overall computational cost of large language models (LLMs). This cost only grows as LLMs scale to larger embedding dimensions and context lengths. In this work, we show that MatMul operations can be completely eliminated from LLMs while maintaining strong performance at billion-parameter scales. Our experiments show that our proposed MatMul-free models achieve performance on-par with state-of-the-art Transformers that require far more memory during inference at a scale up to at least 2.7B parameters. We investigate the scaling laws and find that the performance gap between our MatMul-free models and full precision Transformers narrows as the model size increases. We also provide a GPU-efficient implementation of this model which reduces memory usage by up to 61% over an unoptimized baseline during training. By utilizing an optimized kernel during inference, our model's memory consumption can be reduced by more than 10x compared to unoptimized models. To properly quantify the efficiency of our architecture, we build a custom hardware solution on an FPGA which exploits lightweight operations beyond what GPUs are capable of. We processed billion-parameter scale models at 13W beyond human readable throughput, moving LLMs closer to brain-like efficiency. This work not only shows how far LLMs can be stripped back while still performing effectively, but also points at the types of operations future accelerators should be optimized for in processing the next generation of lightweight LLMs. Our code implementation is available at https://github.com/ridgerchu/matmulfreellm.
This paper introduces Helix, a distributed system for high-throughput, low-latency large language model (LLM) serving in heterogeneous GPU clusters. The key idea behind Helix is to formulate inference computation of LLMs over heterogeneous GPUs and network connections as a max-flow problem on directed, weighted graphs, whose nodes represent GPU instances and edges capture both GPU and network heterogeneity through their capacities. Helix then uses a mixed integer linear programming (MILP) algorithm to discover highly optimized strategies to serve LLMs on heterogeneous GPUs. This approach allows Helix to jointly optimize model placement and request scheduling, two highly entangled tasks in heterogeneous LLM serving. Our evaluation on several heterogeneous clusters ranging from 24 to 42 GPU nodes shows that Helix improves serving throughput by up to 3.3x and reduces prompting and decoding latency by up to 66% and 24%, respectively, compared to existing approaches. Helix is available at https://github.com/Thesys-lab/Helix-ASPLOS25.
Training models with longer in-context lengths is a significant challenge for multimodal model due to substantial GPU memory and computational costs. This exploratory study does not present state-of-the-art models; rather, it introduces an innovative method designed to increase in-context text length in multi-modality large language models (MLLMs) efficiently. We present Visualized In-Context Text Processing (VisInContext), which processes long in-context text using visual tokens. This technique significantly reduces GPU memory usage and floating point operations (FLOPs) for both training and inferenceing stage. For instance, our method expands the pre-training in-context text length from 256 to 2048 tokens with nearly same FLOPs for a 56 billion parameter MOE model. Experimental results demonstrate that model trained with VisInContext delivers superior performance on common downstream benchmarks for in-context few-shot evaluation. Additionally, VisInContext is complementary to existing methods for increasing in-context text length and enhances document understanding capabilities, showing great potential in document QA tasks and sequential document retrieval.
Large language models (LLMs) have shown amazing capabilities in knowledge memorization and the present. However, when it comes to domain-specific knowledge and downstream tasks like medical, general LLMs are often unable to give precise answers. In addition, when people want LLMs to answer classification questions, they usually go through instruction tuning first. However, LLMs do not always give a direct index of the categorization after instruction tuning. In this paper, we proposed LlamaCare, a fine-tuned medical language model, and Extended Classification Integration(ECI), a module to handle classification problems of LLMs. Our contributions are : (i) We fine-tuned a large language model of medical knowledge with very low carbon emissions and achieved similar performance with ChatGPT by a 24G GPU. (ii) We solved the problem of redundant categorical answers and improved the performance of LLMs by proposing a new module called Extended Classification Integration. (iii) We released our processed data for one-shot and few-shot training for some benchmarks such as PubMedQA and USMLE 1-3 step. Our method achieves a close performance comparable to some state-of-the-art models with the same quantity of parameters on benchmarks, while being more environmentally friendly by using less GPU computation time. Our models, codes, and datasets can be found at \url{https://github.com/Stephen-SMJ/LLamaCare}.
In this paper, we propose an efficient, fast, and versatile distillation method to accelerate the generation of pre-trained diffusion models: Flash Diffusion. The method reaches state-of-the-art performances in terms of FID and CLIP-Score for few steps image generation on the COCO2014 and COCO2017 datasets, while requiring only several GPU hours of training and fewer trainable parameters than existing methods. In addition to its efficiency, the versatility of the method is also exposed across several tasks such as text-to-image, inpainting, face-swapping, super-resolution and using different backbones such as UNet-based denoisers (SD1.5, SDXL) or DiT (Pixart-$\alpha$), as well as adapters. In all cases, the method allowed to reduce drastically the number of sampling steps while maintaining very high-quality image generation. The official implementation is available at https://github.com/gojasper/flash-diffusion.
Video generation has made remarkable progress in recent years, especially since the advent of the video diffusion models. Many video generation models can produce plausible synthetic videos, e.g., Stable Video Diffusion (SVD). However, most video models can only generate low frame rate videos due to the limited GPU memory as well as the difficulty of modeling a large set of frames. The training videos are always uniformly sampled at a specified interval for temporal compression. Previous methods promote the frame rate by either training a video interpolation model in pixel space as a postprocessing stage or training an interpolation model in latent space for a specific base video model. In this paper, we propose a training-free video interpolation method for generative video diffusion models, which is generalizable to different models in a plug-and-play manner. We investigate the non-linearity in the feature space of video diffusion models and transform a video model into a self-cascaded video diffusion model with incorporating the designed hidden state correction modules. The self-cascaded architecture and the correction module are proposed to retain the temporal consistency between key frames and the interpolated frames. Extensive evaluations are preformed on multiple popular video models to demonstrate the effectiveness of the propose method, especially that our training-free method is even comparable to trained interpolation models supported by huge compute resources and large-scale datasets.
Despite impressive advancements in diffusion-based video editing models in altering video attributes, there has been limited exploration into modifying motion information while preserving the original protagonist's appearance and background. In this paper, we propose MotionFollower, a lightweight score-guided diffusion model for video motion editing. To introduce conditional controls to the denoising process, MotionFollower leverages two of our proposed lightweight signal controllers, one for poses and the other for appearances, both of which consist of convolution blocks without involving heavy attention calculations. Further, we design a score guidance principle based on a two-branch architecture, including the reconstruction and editing branches, which significantly enhance the modeling capability of texture details and complicated backgrounds. Concretely, we enforce several consistency regularizers and losses during the score estimation. The resulting gradients thus inject appropriate guidance to the intermediate latents, forcing the model to preserve the original background details and protagonists' appearances without interfering with the motion modification. Experiments demonstrate the competitive motion editing ability of MotionFollower qualitatively and quantitatively. Compared with MotionEditor, the most advanced motion editing model, MotionFollower achieves an approximately 80% reduction in GPU memory while delivering superior motion editing performance and exclusively supporting large camera movements and actions.
The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. While recent research has investigated various speculative decoding techniques for multi-token generation, these efforts have primarily focused on improving processing speed such as throughput. Crucially, they often neglect other metrics essential for real-life deployments, such as memory consumption and training cost. To overcome these limitations, we propose a novel parallel prompt decoding that requires only $0.0002$% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. Inspired by the human natural language generation process, $PPD$ approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. This approach partially recovers the missing conditional dependency information necessary for multi-token generation, resulting in up to a 28% higher acceptance rate for long-range predictions. Furthermore, we present a hardware-aware dynamic sparse tree technique that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs. Through extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49$\times$ speedup and maintains a minimal runtime memory overhead of just $0.0004$%. More importantly, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding, showing up to $1.22\times$ further speed improvement. Our code is available at https://github.com/hmarkc/parallel-prompt-decoding.
Accurate prediction of Cardiovascular disease (CVD) risk in medical imaging is central to effective patient health management. Previous studies have demonstrated that imaging features in computed tomography (CT) can help predict CVD risk. However, CT entails notable radiation exposure, which may result in adverse health effects for patients. In contrast, chest X-ray emits significantly lower levels of radiation, offering a safer option. This rationale motivates our investigation into the feasibility of using chest X-ray for predicting CVD risk. Convolutional Neural Networks (CNNs) and Transformers are two established network architectures for computer-aided diagnosis. However, they struggle to model very high resolution chest X-ray due to the lack of large context modeling power or quadratic time complexity. Inspired by state space sequence models (SSMs), a new class of network architectures with competitive sequence modeling power as Transfomers and linear time complexity, we propose Bidirectional Image Mamba (BI-Mamba) to complement the unidirectional SSMs with opposite directional information. BI-Mamba utilizes parallel forward and backwark blocks to encode longe-range dependencies of multi-view chest X-rays. We conduct extensive experiments on images from 10,395 subjects in National Lung Screening Trail (NLST). Results show that BI-Mamba outperforms ResNet-50 and ViT-S with comparable parameter size, and saves significant amount of GPU memory during training. Besides, BI-Mamba achieves promising performance compared with previous state of the art in CT, unraveling the potential of chest X-ray for CVD risk prediction.
Spatial Message Passing Graph Neural Networks (MPGNNs) are widely used for learning on graph-structured data. However, key limitations of l-step MPGNNs are that their "receptive field" is typically limited to the l-hop neighborhood of a node and that information exchange between distant nodes is limited by over-squashing. Motivated by these limitations, we propose Spatio-Spectral Graph Neural Networks (S$^2$GNNs) -- a new modeling paradigm for Graph Neural Networks (GNNs) that synergistically combines spatially and spectrally parametrized graph filters. Parameterizing filters partially in the frequency domain enables global yet efficient information propagation. We show that S$^2$GNNs vanquish over-squashing and yield strictly tighter approximation-theoretic error bounds than MPGNNs. Further, rethinking graph convolutions at a fundamental level unlocks new design spaces. For example, S$^2$GNNs allow for free positional encodings that make them strictly more expressive than the 1-Weisfeiler-Lehman (WL) test. Moreover, to obtain general-purpose S$^2$GNNs, we propose spectrally parametrized filters for directed graphs. S$^2$GNNs outperform spatial MPGNNs, graph transformers, and graph rewirings, e.g., on the peptide long-range benchmark tasks, and are competitive with state-of-the-art sequence modeling. On a 40 GB GPU, S$^2$GNNs scale to millions of nodes.