Quantizing large language models has become a standard way to reduce their memory and computational costs. Typically, existing methods focus on breaking down the problem into individual layer-wise sub-problems, and minimizing per-layer error, measured via various metrics. Yet, this approach currently lacks theoretical justification and the metrics employed may be sub-optimal. In this paper, we present a "linearity theorem" establishing a direct relationship between the layer-wise $\ell_2$ reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint in the medium-bitwidth regime, obtained by reduction to dynamic programming. On the practical side, we demonstrate improved accuracy-compression trade-offs on Llama-3.1 and 3.2-family models, as well as on Qwen-family models. Further, we show that our method can be efficiently supported in terms of GPU kernels at various batch sizes, advancing both data-free and non-uniform quantization for LLMs.
Inference for Large Language Models (LLMs) is computationally demanding. To reduce the cost of auto-regressive decoding, Key-Value (KV) cache is used to store intermediate activations, which significantly lowers the computational overhead for token generation. However, the memory required for the KV cache grows rapidly, often exceeding the capacity of GPU memory. A cost-effective alternative is to offload KV cache to CPU memory, which alleviates GPU memory pressure, but shifts the bottleneck to the limited bandwidth of the PCIe connection between the CPU and GPU. Existing methods attempt to address these issues by overlapping GPU computation with I/O or employing CPU-GPU heterogeneous execution, but they are hindered by excessive data movement and dependence on CPU capabilities. Fully overlapping PCIe communication latency gets challenging as the size of the KV cache grows and/or the GPU compute capabilities increase. In this paper, we introduce KVPR, an efficient I/O-aware LLM inference method where the CPU first transfers a partial set of activations, from which the GPU can start recomputing the KV cache values. While the GPU recomputes the partial KV cache, the remaining portion of the KV cache is transferred concurrently from the CPU. This approach overlaps GPU recomputation with KV cache transfer to minimize idle GPU time and maximize inference performance. KVPR is fully automated by integrating a profiler module that utilizes input characteristics and system hardware information, a scheduler module to optimize the distribution of computation and communication workloads, and a runtime module to efficiently execute the derived execution plan. Experimental results show that KVPR achieves up to 35.8% lower latency and 46.2% higher throughput during decoding compared to state-of-the-art approaches. The code is available at https://github.com/chaoyij/KVPR.
Autoregressive models have demonstrated remarkable success across various fields, from large language models (LLMs) to large multimodal models (LMMs) and 2D content generation, moving closer to artificial general intelligence (AGI). Despite these advances, applying autoregressive approaches to 3D object generation and understanding remains largely unexplored. This paper introduces Scale AutoRegressive 3D (SAR3D), a novel framework that leverages a multi-scale 3D vector-quantized variational autoencoder (VQVAE) to tokenize 3D objects for efficient autoregressive generation and detailed understanding. By predicting the next scale in a multi-scale latent representation instead of the next single token, SAR3D reduces generation time significantly, achieving fast 3D object generation in just 0.82 seconds on an A6000 GPU. Additionally, given the tokens enriched with hierarchical 3D-aware information, we finetune a pretrained LLM on them, enabling multimodal comprehension of 3D content. Our experiments show that SAR3D surpasses current 3D generation methods in both speed and quality and allows LLMs to interpret and caption 3D models comprehensively.
Every year, thousands of people suffer from skull damage and require personalized implants to fill the cranial cavity. Unfortunately, the waiting time for reconstruction surgery can extend to several weeks or even months, especially in less developed countries. One factor contributing to the extended waiting period is the intricate process of personalized implant modeling. Currently, the preparation of these implants by experienced biomechanical experts is both costly and time-consuming. Recent advances in artificial intelligence, especially in deep learning, offer promising potential for automating the process. However, deep learning-based cranial reconstruction faces several challenges: (i) the limited size of training datasets, (ii) the high resolution of the volumetric data, and (iii) significant data heterogeneity. In this work, we propose a novel approach to address these challenges by enhancing the reconstruction through learnable symmetry enforcement. We demonstrate that it is possible to train a neural network dedicated to calculating skull symmetry, which can be utilized either as an additional objective function during training or as a post-reconstruction objective during the refinement step. We quantitatively evaluate the proposed method using open SkullBreak and SkullFix datasets, and qualitatively using real clinical cases. The results indicate that the symmetry-preserving reconstruction network achieves considerably better outcomes compared to the baseline (0.94/0.94/1.31 vs 0.84/0.76/2.43 in terms of DSC, bDSC, and HD95). Moreover, the results are comparable to the best-performing methods while requiring significantly fewer computational resources (< 500 vs > 100,000 GPU hours). The proposed method is a considerable contribution to the field of applied artificial intelligence in medicine and is a step toward automatic cranial defect reconstruction in clinical practice.
The forecasting skill of numerical weather prediction (NWP) models critically depends on the accurate initial conditions, also known as analysis, provided by data assimilation (DA). Traditional DA methods often face a trade-off between computational cost and accuracy due to complex linear algebra computations and the high dimensionality of the model, especially in nonlinear systems. Moreover, processing massive data in real-time requires substantial computational resources. To address this, we introduce an artificial intelligence-based data assimilation framework (ADAF) to generate high-quality kilometer-scale analysis. This study is the pioneering work using real-world observations from varied locations and multiple sources to verify the AI method's efficacy in DA, including sparse surface weather observations and satellite imagery. We implemented ADAF for four near-surface variables in the Contiguous United States (CONUS). The results indicate that ADAF surpasses the High Resolution Rapid Refresh Data Assimilation System (HRRRDAS) in accuracy by 16% to 33% for near-surface atmospheric conditions, aligning more closely with actual observations, and can effectively reconstruct extreme events, such as tropical cyclone wind fields. Sensitivity experiments reveal that ADAF can generate high-quality analysis even with low-accuracy backgrounds and extremely sparse surface observations. ADAF can assimilate massive observations within a three-hour window at low computational cost, taking about two seconds on an AMD MI200 graphics processing unit (GPU). ADAF has been shown to be efficient and effective in real-world DA, underscoring its potential role in operational weather forecasting.
Graph neural architecture search (GNAS) can customize high-performance graph neural network architectures for specific graph tasks or datasets. However, existing GNAS methods begin searching for architectures from a zero-knowledge state, ignoring the prior knowledge that may improve the search efficiency. The available knowledge base (e.g. NAS-Bench-Graph) contains many rich architectures and their multiple performance metrics, such as the accuracy (#Acc) and number of parameters (#Params). This study proposes exploiting such prior knowledge to accelerate the multi-objective evolutionary search on a new graph dataset, named knowledge-aware evolutionary GNAS (KEGNAS). KEGNAS employs the knowledge base to train a knowledge model and a deep multi-output Gaussian process (DMOGP) in one go, which generates and evaluates transfer architectures in only a few GPU seconds. The knowledge model first establishes a dataset-to-architecture mapping, which can quickly generate candidate transfer architectures for a new dataset. Subsequently, the DMOGP with architecture and dataset encodings is designed to predict multiple performance metrics for candidate transfer architectures on the new dataset. According to the predicted metrics, non-dominated candidate transfer architectures are selected to warm-start the multi-objective evolutionary algorithm for optimizing the #Acc and #Params on a new dataset. Empirical studies on NAS-Bench-Graph and five real-world datasets show that KEGNAS swiftly generates top-performance architectures, achieving 4.27% higher accuracy than advanced evolutionary baselines and 11.54% higher accuracy than advanced differentiable baselines. In addition, ablation studies demonstrate that the use of prior knowledge significantly improves the search performance.
Graph Neural Networks (GNNs) have shown significant promise in various domains, such as recommendation systems, bioinformatics, and network analysis. However, the irregularity of graph data poses unique challenges for efficient computation, leading to the development of specialized GNN accelerator architectures that surpass traditional CPU and GPU performance. Despite this, the structural diversity of input graphs results in varying performance across different GNN accelerators, depending on their dataflows. This variability in performance due to differing dataflows and graph properties remains largely unexplored, limiting the adaptability of GNN accelerators. To address this, we propose a data-driven framework for dataflow-aware latency prediction in GNN inference. Our approach involves training regressors to predict the latency of executing specific graphs on particular dataflows, using simulations on synthetic graphs. Experimental results indicate that our regressors can predict the optimal dataflow for a given graph with up to 91.28% accuracy and a Mean Absolute Percentage Error (MAPE) of 3.78%. Additionally, we introduce an online scheduling algorithm that uses these regressors to enhance scheduling decisions. Our experiments demonstrate that this algorithm achieves up to $3.17\times$ speedup in mean completion time and $6.26\times$ speedup in mean execution time compared to the best feasible baseline across all datasets.
Previous research on lightweight models has primarily focused on CNNs and Transformer-based designs. CNNs, with their local receptive fields, struggle to capture long-range dependencies, while Transformers, despite their global modeling capabilities, are limited by quadratic computational complexity in high-resolution scenarios. Recently, state-space models have gained popularity in the visual domain due to their linear computational complexity. Despite their low FLOPs, current lightweight Mamba-based models exhibit suboptimal throughput. In this work, we propose the MobileMamba framework, which balances efficiency and performance. We design a three-stage network to enhance inference speed significantly. At a fine-grained level, we introduce the Multi-Receptive Field Feature Interaction(MRFFI) module, comprising the Long-Range Wavelet Transform-Enhanced Mamba(WTE-Mamba), Efficient Multi-Kernel Depthwise Convolution(MK-DeConv), and Eliminate Redundant Identity components. This module integrates multi-receptive field information and enhances high-frequency detail extraction. Additionally, we employ training and testing strategies to further improve performance and efficiency. MobileMamba achieves up to 83.6% on Top-1, surpassing existing state-of-the-art methods which is maximum x21 faster than LocalVim on GPU. Extensive experiments on high-resolution downstream tasks demonstrate that MobileMamba surpasses current efficient models, achieving an optimal balance between speed and accuracy.
Arbor is a software library designed for efficient simulation of large-scale networks of biological neurons with detailed morphological structures. It combines customizable neuronal and synaptic mechanisms with high-performance computing, supporting multi-core CPU and GPU systems. In humans and other animals, synaptic plasticity processes play a vital role in cognitive functions, including learning and memory. Recent studies have shown that intracellular molecular processes in dendrites significantly influence single-neuron dynamics. However, for understanding how the complex interplay between dendrites and synaptic processes influences network dynamics, computational modeling is required. To enable the modeling of large-scale networks of morphologically detailed neurons with diverse plasticity processes, we have extended the Arbor library to the Plastic Arbor framework, supporting simulations of a large variety of spike-driven plasticity paradigms. To showcase the features of the new framework, we present examples of computational models, beginning with single-synapse dynamics, progressing to multi-synapse rules, and finally scaling up to large recurrent networks. While cross-validating our implementations by comparison with other simulators, we show that Arbor allows simulating plastic networks of multi-compartment neurons at nearly no additional cost in runtime compared to point-neuron simulations. Using the new framework, we have already been able to investigate the impact of dendritic structures on network dynamics across a timescale of several hours, showing a relation between the length of dendritic trees and the ability of the network to efficiently store information. By our extension of Arbor, we aim to provide a valuable tool that will support future studies on the impact of synaptic plasticity, especially, in conjunction with neuronal morphology, in large networks.
This review report discusses the cold start latency in serverless inference and existing solutions. It particularly reviews the ServerlessLLM method, a system designed to address the cold start problem in serverless inference for large language models. Traditional serverless approaches struggle with high latency due to the size of LLM checkpoints and the overhead of initializing GPU resources. ServerlessLLM introduces a multitier checkpoint loading system, leveraging underutilized GPU memory and storage to reduce startup times by 6--8x compared to existing methods. It also proposes live inference migration and a startup-time-optimized model scheduler, ensuring efficient resource allocation and minimizing delays. This system significantly improves performance and scalability in serverless environments for LLM workloads. Besides ServerlessLLM, several other methods from recent research literature, including Rainbowcake, are reviewed in this paper. Further discussions explore how FaaS providers tackle cold starts and the possible future scopes.
Night unmanned aerial vehicle (UAV) tracking is impeded by the challenges of poor illumination, with previous daylight-optimized methods demonstrating suboptimal performance in low-light conditions, limiting the utility of UAV applications. To this end, we propose an efficient mamba-based tracker, leveraging dual enhancement techniques to boost night UAV tracking. The mamba-based low-light enhancer, equipped with an illumination estimator and a damage restorer, achieves global image enhancement while preserving the details and structure of low-light images. Additionally, we advance a cross-modal mamba network to achieve efficient interactive learning between vision and language modalities. Extensive experiments showcase that our method achieves advanced performance and exhibits significantly improved computation and memory efficiency. For instance, our method is 2.8$\times$ faster than CiteTracker and reduces 50.2$\%$ GPU memory. Our codes are available at \url{https://github.com/983632847/Awesome-Multimodal-Object-Tracking}.
Although large language models (LLMs) have achieved remarkable success across various domains, their considerable scale necessitates substantial computational resources, posing significant challenges for deployment in resource-constrained environments. Layer pruning, as a simple yet effective compression method, removes layers of a model directly, reducing computational overhead. However, what are the best practices for layer pruning in LLMs? Are sophisticated layer selection metrics truly effective? Does the LoRA (Low-Rank Approximation) family, widely regarded as a leading method for pruned model fine-tuning, truly meet expectations when applied to post-pruning fine-tuning? To answer these questions, we dedicate thousands of GPU hours to benchmarking layer pruning in LLMs and gaining insights across multiple dimensions. Our results demonstrate that a simple approach, i.e., pruning the final 25\% of layers followed by fine-tuning the \texttt{lm\_head} and the remaining last three layer, yields remarkably strong performance. Following this guide, we prune Llama-3.1-8B-It and obtain a model that outperforms many popular LLMs of similar size, such as ChatGLM2-6B, Vicuna-7B-v1.5, Qwen1.5-7B and Baichuan2-7B. We release the optimal model weights on Huggingface, and the code is available on GitHub.
Organs-at-risk segmentation is critical for ensuring the safety and precision of radiotherapy and surgical procedures. However, existing methods for organs-at-risk image segmentation often suffer from uncertainties and biases in target selection, as well as insufficient model validation experiments, limiting their generality and reliability in practical applications. To address these issues, we propose an innovative cascaded network architecture called the Multi-scale Cascaded Fusing Network (MCFNet), which effectively captures complex multi-scale and multi-resolution features. MCFNet includes a Sharp Extraction Backbone and a Flexible Connection Backbone, which respectively enhance feature extraction in the downsampling and skip-connection stages. This design not only improves segmentation accuracy but also ensures computational efficiency, enabling precise detail capture even in low-resolution images. We conduct experiments using the A6000 GPU on diverse datasets from 671 patients, including 36,131 image-mask pairs across 10 different datasets. MCFNet demonstrates strong robustness, performing consistently well across 10 datasets. Additionally, MCFNet exhibits excellent generalizability, maintaining high accuracy in different clinical scenarios. We also introduce an adaptive loss aggregation strategy to further optimize the model training process, improving both segmentation accuracy and efficiency. Through extensive validation, MCFNet demonstrates superior performance compared to existing methods, providing more reliable image-guided support. Our solution aims to significantly improve the precision and safety of radiotherapy and surgical procedures, advancing personalized treatment. The code has been made available on GitHub:https://github.com/Henry991115/MCFNet.
Deep learning models often require specially designed architectures to process data of different dimensions, such as 1D time series, 2D images, and 3D volumetric data. Existing bidirectional models mainly focus on sequential data, making it difficult to scale effectively to higher dimensions. To address this issue, we propose a novel multi-dimensional bidirectional neural network architecture, named Nd-BiMamba2, which efficiently handles 1D, 2D, and 3D data. Nd-BiMamba2 is based on the Mamba2 module and introduces innovative bidirectional processing mechanisms and adaptive padding strategies to capture bidirectional information in multi-dimensional data while maintaining computational efficiency. Unlike existing methods that require designing specific architectures for different dimensional data, Nd-BiMamba2 adopts a unified architecture with a modular design, simplifying development and maintenance costs. To verify the portability and flexibility of Nd-BiMamba2, we successfully exported it to ONNX and TorchScript and tested it on different hardware platforms (e.g., CPU, GPU, and mobile devices). Experimental results show that Nd-BiMamba2 runs efficiently on multiple platforms, demonstrating its potential in practical applications. The code is open-source: https://github.com/Human9000/nd-Mamba2-torch
The widely-used, weight-only quantized large language models (LLMs), which leverage low-bit integer (INT) weights and retain floating-point (FP) activations, reduce storage requirements while maintaining accuracy. However, this shifts the energy and latency bottlenecks towards the FP activations that are associated with costly memory accesses and computations. Existing LLM accelerators focus primarily on computation optimizations, overlooking the potential of jointly optimizing FP computations and data movement, particularly for the dominant FP-INT GeMM operations in LLM inference. To address these challenges, we investigate the sensitivity of activation precision across various LLM modules and its impact on overall model accuracy. Based on our findings, we first propose the Anda data type: an adaptive data format with group-shared exponent bits and dynamic mantissa bit allocation. Secondly, we develop an iterative post-training adaptive precision search algorithm that optimizes the bit-width for different LLM modules to balance model accuracy, energy efficiency, and inference speed. Lastly, a suite of hardware optimization techniques is proposed to maximally exploit the benefits of the Anda format. These include a bit-plane-based data organization scheme, Anda-enhanced processing units with bit-serial computation, and a runtime bit-plane Anda compressor to simultaneously optimize storage, computation, and memory footprints. Our evaluations on FPINT GeMM operations show that Anda achieves a 2.4x speedup, 4.0x area efficiency, and 3.1x energy efficiency improvement on average for popular LLMs including OPT, LLaMA, and LLaMA-2 series over the GPU-like FP-FP baseline. Anda demonstrates strong adaptability across various application scenarios, accuracy requirements, and system performance, enabling efficient LLM inference across a wide range of deployment scenarios.
Burn injuries present a significant global health challenge. Among the most severe long-term consequences are contractures, which can lead to functional impairments and disfigurement. Understanding and predicting the evolution of post-burn wounds is essential for developing effective treatment strategies. Traditional mathematical models, while accurate, are often computationally expensive and time-consuming, limiting their practical application. Recent advancements in machine learning, particularly in deep learning, offer promising alternatives for accelerating these predictions. This study explores the use of a deep operator network (DeepONet), a type of neural operator, as a surrogate model for finite element simulations, aimed at predicting post-burn contraction across multiple wound shapes. A DeepONet was trained on three distinct initial wound shapes, with enhancement made to the architecture by incorporating initial wound shape information and applying sine augmentation to enforce boundary conditions. The performance of the trained DeepONet was evaluated on a test set including finite element simulations based on convex combinations of the three basic wound shapes. The model achieved an $R^2$ score of $0.99$, indicating strong predictive accuracy and generalization. Moreover, the model provided reliable predictions over an extended period of up to one year, with speedups of up to 128-fold on CPU and 235-fold on GPU, compared to the numerical model. These findings suggest that DeepONets can effectively serve as a surrogate for traditional finite element methods in simulating post-burn wound evolution, with potential applications in medical treatment planning.
Contrastive Language-Image Pre-training (CLIP) has attracted a surge of attention for its superior zero-shot performance and excellent transferability to downstream tasks. However, training such large-scale models usually requires substantial computation and storage, which poses barriers for general users with consumer-level computers. Motivated by this observation, in this paper we investigate how to achieve competitive performance on only one Nvidia RTX3090 GPU and with one terabyte for storing dataset. On one hand, we simplify the transformer block structure and combine Weight Inheritance with multi-stage Knowledge Distillation (WIKD), thereby reducing the parameters and improving the inference speed during training along with deployment. On the other hand, confronted with the convergence challenge posed by small dataset, we generate synthetic captions for each sample as data augmentation, and devise a novel Pair Matching (PM) loss to fully exploit the distinguishment among positive and negative image-text pairs. Extensive experiments demonstrate that our model can achieve a new state-of-the-art datascale-parameter-accuracy tradeoff, which could further popularize the CLIP model in the related research community.
The applications of LLM Agents are becoming increasingly complex and diverse, leading to a high demand for structured outputs that can be parsed into code, structured function calls, and embodied agent commands. These developments bring significant demands for structured generation in LLM inference. Context-free grammar is a flexible approach to enable structured generation via constrained decoding. However, executing context-free grammar requires going through several stack states over all tokens in vocabulary during runtime, bringing non-negligible overhead for structured generation. In this paper, we propose XGrammar, a flexible and efficient structure generation engine for large language models. XGrammar accelerates context-free grammar execution by dividing the vocabulary into context-independent tokens that can be prechecked and context-dependent tokens that need to be interpreted during runtime. We further build transformations to expand the grammar context and reduce the number of context-independent tokens. Additionally, we build an efficient persistent stack to accelerate the context-dependent token checks. Finally, we co-design the grammar engine with LLM inference engine to overlap grammar computation with GPU executions. Evaluation results show that XGrammar can achieve up to 100x speedup over existing solutions. Combined with an LLM inference engine, it can generate near-zero overhead structure generation in end-to-end low-LLM serving.
The widespread adoption of language models (LMs) across multiple industries has caused huge surge in demand for GPUs. Training LMs requires tens of thousands of GPUs and housing them in the same datacenter (DCs) is becoming challenging. We focus on training such models across multiple DCs connected via Wide-Area-Network (WAN). We build ATLAS that speeds up such training time using novel temporal bandwidth sharing and many other design choices. While ATLAS improves the training time, it does not eliminate the bubbles (idle GPU cycles). We built BUBBLETEA that runs prefill-as-a-service (part of LM inference) during the bubbles that improves the GPU utilization substantially without any impact of training. Together, ATLAS and BUBBLETEA improve training time by up to 17X and achieve GPU utilization of up to 94%.
We present FAST-Splat for fast, ambiguity-free semantic Gaussian Splatting, which seeks to address the main limitations of existing semantic Gaussian Splatting methods, namely: slow training and rendering speeds; high memory usage; and ambiguous semantic object localization. In deriving FAST-Splat , we formulate open-vocabulary semantic Gaussian Splatting as the problem of extending closed-set semantic distillation to the open-set (open-vocabulary) setting, enabling FAST-Splat to provide precise semantic object localization results, even when prompted with ambiguous user-provided natural-language queries. Further, by exploiting the explicit form of the Gaussian Splatting scene representation to the fullest extent, FAST-Splat retains the remarkable training and rendering speeds of Gaussian Splatting. Specifically, while existing semantic Gaussian Splatting methods distill semantics into a separate neural field or utilize neural models for dimensionality reduction, FAST-Splat directly augments each Gaussian with specific semantic codes, preserving the training, rendering, and memory-usage advantages of Gaussian Splatting over neural field methods. These Gaussian-specific semantic codes, together with a hash-table, enable semantic similarity to be measured with open-vocabulary user prompts and further enable FAST-Splat to respond with unambiguous semantic object labels and 3D masks, unlike prior methods. In experiments, we demonstrate that FAST-Splat is 4x to 6x faster to train with a 13x faster data pre-processing step, achieves between 18x to 75x faster rendering speeds, and requires about 3x smaller GPU memory, compared to the best-competing semantic Gaussian Splatting methods. Further, FAST-Splat achieves relatively similar or better semantic segmentation performance compared to existing methods. After the review period, we will provide links to the project website and the codebase.