Root Cause Analysis (RCA) plays a pivotal role in the incident diagnosis process for cloud services, requiring on-call engineers to identify the primary issues and implement corrective actions to prevent future recurrences. Improving the incident RCA process is vital for minimizing service downtime, customer impact and manual toil. Recent advances in artificial intelligence have introduced state-of-the-art Large Language Models (LLMs) like GPT-4, which have proven effective in tackling various AIOps problems, ranging from code authoring to incident management. Nonetheless, the GPT-4 model's immense size presents challenges when trying to fine-tune it on user data because of the significant GPU resource demand and the necessity for continuous model fine-tuning with the emergence of new data. To address the high cost of fine-tuning LLM, we propose an in-context learning approach for automated root causing, which eliminates the need for fine-tuning. We conduct extensive study over 100,000 production incidents, comparing several large language models using multiple metrics. The results reveal that our in-context learning approach outperforms the previous fine-tuned large language models such as GPT-3 by an average of 24.8\% across all metrics, with an impressive 49.7\% improvement over the zero-shot model. Moreover, human evaluation involving actual incident owners demonstrates its superiority over the fine-tuned model, achieving a 43.5\% improvement in correctness and an 8.7\% enhancement in readability. The impressive results demonstrate the viability of utilizing a vanilla GPT model for the RCA task, thereby avoiding the high computational and maintenance costs associated with a fine-tuned model.
This paper presents MoE-Infinity, an efficient MoE inference system designed for personal machines with limited GPU memory capacity. The key idea for MoE-Infinity is that on personal machines, which are often single-user environments, MoE-based LLMs typically operate with a batch size of one. In this setting, MoE models exhibit a high degree of activation sparsity, meaning a small number of experts are frequently reused in generating tokens during the decode phase. Leveraging this idea, we design a sparsity-aware expert cache, which can trace the sparse activation of experts during inference and carefully select the trace that represents the sparsity pattern. By analyzing these selected traces, MoE-Infinity guides the replacement and prefetching of the expert cache, providing 3.1-16.7x per-token latency improvements over numerous state-of-the-art systems, including vLLM, Ollama, DeepSpeed and BrainStorm across various MoE models (DeepSeek and Mixtral) when handling different LLM tasks. MoE-Infinity's source code is publicly available at https://github.com/EfficientMoE/MoE-Infinity
This paper presents ServerlessLLM, a distributed system designed to support low-latency serverless inference for Large Language Models (LLMs). By harnessing the substantial near-GPU storage and memory capacities of inference servers, ServerlessLLM achieves effective local checkpoint storage, minimizing the need for remote checkpoint downloads and ensuring efficient checkpoint loading. The design of ServerlessLLM features three core contributions: (i) \emph{fast multi-tier checkpoint loading}, featuring a new loading-optimized checkpoint format and a multi-tier loading system, fully utilizing the bandwidth of complex storage hierarchies on GPU servers; (ii) \emph{efficient live migration of LLM inference}, which enables newly initiated inferences to capitalize on local checkpoint storage while ensuring minimal user interruption; and (iii) \emph{startup-time-optimized model scheduling}, which assesses the locality statuses of checkpoints on each server and schedules the model onto servers that minimize the time to start the inference. Comprehensive evaluations, including microbenchmarks and real-world scenarios, demonstrate that ServerlessLLM dramatically outperforms state-of-the-art serverless systems, reducing latency by 10 - 200X across various LLM inference workloads.
Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code is publicly available at https://github.com/usyd-fsalab/fp6_llm.
Neural networks have become popular due to their versatility and state-of-the-art results in many applications, such as image classification, natural language processing, speech recognition, forecasting, etc. These applications are also used in resource-constrained environments such as embedded devices. In this work, the susceptibility of neural network implementations to reverse engineering is explored on the NVIDIA Jetson Nano microcomputer via side-channel analysis. To this end, an architecture extraction attack is presented. In the attack, 15 popular convolutional neural network architectures (EfficientNets, MobileNets, NasNet, etc.) are implemented on the GPU of Jetson Nano and the electromagnetic radiation of the GPU is analyzed during the inference operation of the neural networks. The results of the analysis show that neural network architectures are easily distinguishable using deep learning-based side-channel analysis.
Recent years have seen an exponential rise in complex software-driven functionality in vehicles, leading to a rising number of electronic control units (ECUs), network capabilities, and interfaces. These expanded capabilities also bring-in new planes of vulnerabilities making intrusion detection and management a critical capability; however, this can often result in more ECUs and network elements due to the high computational overheads. In this paper, we present a consolidated ECU architecture incorporating an Intrusion Detection System (IDS) for Automotive Controller Area Network (CAN) along with traditional ECU functionality on an off-the-shelf hybrid FPGA device, with near-zero overhead for the ECU functionality. We propose two quantised multi-layer perceptrons (QMLP's) as isolated IDSs for detecting a range of attack vectors including Denial-of-Service, Fuzzing and Spoofing, which are accelerated using off-the-shelf deep-learning processing unit (DPU) IP block from Xilinx, operating fully transparently to the software on the ECU. The proposed models achieve the state-of-the-art classification accuracy for all the attacks, while we observed a 15x reduction in power consumption when compared against the GPU-based implementation of the same models quantised using Nvidia libraries. We also achieved a 2.3x speed up in per-message processing latency (at 0.24 ms from the arrival of a CAN message) to meet the strict end-to-end latency on critical CAN nodes and a 2.6x reduction in power consumption for inference when compared to the state-of-the-art IDS models on embedded IDS and loosely coupled IDS accelerators (GPUs) discussed in the literature.
This paper introduces InverseMatrixVT3D, an efficient method for transforming multi-view image features into 3D feature volumes for 3D semantic occupancy prediction. Existing methods for constructing 3D volumes often rely on depth estimation, device-specific operators, or transformer queries, which hinders the widespread adoption of 3D occupancy models. In contrast, our approach leverages two projection matrices to store the static mapping relationships and matrix multiplications to efficiently generate global Bird's Eye View (BEV) features and local 3D feature volumes. Specifically, we achieve this by performing matrix multiplications between multi-view image feature maps and two sparse projection matrices. We introduce a sparse matrix handling technique for the projection matrices to optimize GPU memory usage. Moreover, a global-local attention fusion module is proposed to integrate the global BEV features with the local 3D feature volumes to obtain the final 3D volume. We also employ a multi-scale supervision mechanism to enhance performance further. Extensive experiments performed on the nuScenes and SemanticKITTI datasets reveal that our approach not only stands out for its simplicity and effectiveness but also achieves the top performance in detecting vulnerable road users (VRU), crucial for autonomous driving and road safety. The code has been made available at: https://github.com/DanielMing123/InverseMatrixVT3D
Accurate and rapid railway track segmentation can assist automatic train driving and is a key step in early warning to fixed or moving obstacles on the railway track. However, certain existing algorithms tailored for track segmentation often struggle to meet the requirements of real-time and efficiency on resource-constrained edge devices. Considering this challenge, we propose an edge-enabled real-time railway track segmentation algorithm, which is optimized to be suitable for edge applications by optimizing the network structure and quantizing the model after training. Initially, Ghost convolution is introduced to reduce the complexity of the backbone, thereby achieving the extraction of key information of the interested region at a lower cost. To further reduce the model complexity and calculation, a new lightweight detection head is proposed to achieve the best balance between accuracy and efficiency. Subsequently, we introduce quantization techniques to map the model's floating-point weights and activation values into lower bit-width fixed-point representations, reducing computational demands and memory footprint, ultimately accelerating the model's inference. Finally, we draw inspiration from GPU parallel programming principles to expedite the pre-processing and post-processing stages of the algorithm by doing parallel processing. The approach is evaluated with public and challenging dataset RailSem19 and tested on Jetson Nano. Experimental results demonstrate that our enhanced algorithm achieves an accuracy level of 83.3% while achieving a real-time inference rate of 25 frames per second when the input size is 480x480, thereby effectively meeting the requirements for real-time and high-efficiency operation.
We present an implementation of interval analysis and mixed monotone interval reachability analysis as function transforms in Python, fully composable with the computational framework JAX. The resulting toolbox inherits several key features from JAX, including computational efficiency through Just-In-Time Compilation, GPU acceleration for quick parallelized computations, and Automatic Differentiability. We demonstrate the toolbox's performance on several case studies, including a reachability problem on a vehicle model controlled by a neural network, and a robust closed-loop optimal control problem for a swinging pendulum.
This study introduces an innovative approach aimed at the efficient pruning of neural networks, with a particular focus on their deployment on edge devices. Our method involves the integration of the Lottery Ticket Hypothesis (LTH) with the Knowledge Distillation (KD) framework, resulting in the formulation of three distinct pruning models. These models have been developed to address scalability issue in recommender systems, whereby the complexities of deep learning models have hindered their practical deployment. With judicious application of the pruning techniques, we effectively curtail the power consumption and model dimensions without compromising on accuracy. Empirical evaluation has been performed using two real world datasets from diverse domains against two baselines. Gratifyingly, our approaches yielded a GPU computation-power reduction of up to 66.67%. Notably, our study contributes to the field of recommendation system by pioneering the application of LTH and KD.
In this paper we build a case for providing job completion time predictions to cloud users, similar to the delivery date of a package or arrival time of a booked ride. Our analysis reveals that providing predictability can come at the expense of performance and fairness. Existing cloud scheduling systems optimize for extreme points in the trade-off space, making them either extremely unpredictable or impractical. To address this challenge, we present PCS, a new scheduling framework that aims to provide predictability while balancing other traditional objectives. The key idea behind PCS is to use Weighted-Fair-Queueing (WFQ) and find a suitable configuration of different WFQ parameters (e.g., class weights) that meets specific goals for predictability. It uses a simulation-aided search strategy, to efficiently discover WFQ configurations that lie on the Pareto front of the trade-off space between these objectives. We implement and evaluate PCS in the context of DNN job scheduling on GPUs. Our evaluation, on a small scale GPU testbed and larger-scale simulations, shows that PCS can provide accurate completion time estimates while marginally compromising on performance and fairness.
In this paper, we present dSASA (differentiable SASA), an exact geometric method to calculate solvent accessible surface area (SASA) analytically along with atomic derivatives on GPUs. The atoms in a molecule are first assigned to tetrahedra in groups of four atoms by Delaunay tetrahedrization adapted for efficient GPU implementation and the SASA values for atoms and molecules are calculated based on the tetrahedrization information and inclusion-exclusion method. The SASA values from the numerical icosahedral-based method can be reproduced with more than 98% accuracy for both proteins and RNAs. Having been implemented on GPUs and incorporated into the software Amber, we can apply dSASA to implicit solvent molecular dynamics simulations with inclusion of this nonpolar term. The current GPU version of GB/SA simulations has been accelerated up to nearly 20-fold compared to the CPU version, outperforming LCPO, a commonly used, fast algorithm for calculating SASA, as the system size increases. While we focus on the accuracy of the SASA calculations for proteins and nucleic acids, we also demonstrate stable GB/SA MD mini-protein simulations.
Visual defect detection plays an important role in intelligent industry. Patch based methods consider visual images as a collection of image patches according to positions, which have stronger discriminative ability for small defects in products, e.g. scratches on pills. However, the nearest neighbor search for the query image and the stored patches will occupy $O(n)$ complexity in terms of time and space requirements, posing strict challenges for deployment in edge environments. In this paper, we propose an alternative approach to the distance calculation of image patches via collaborative representation models. Starting from the nearest neighbor distance with $L_0$ constraint, we relax the constraint to $L_2$ constraint and solve the distance quickly in close-formed without actually accessing the original stored collection of image patches. Furthermore, we point out that the main computational burden of this close-formed solution can be pre-computed by high-performance server before deployment. Consequently, the distance calculation on edge devices only requires a simple matrix multiplication, which is extremely lightweight and GPU-friendly. Performance on real industrial scenarios demonstrates that compared to the existing state-of-the-art methods, this distance achieves several hundred times improvement in computational efficiency with slight performance drop, while greatly reducing memory overhead.
Large Language Models (LLMs) face limitations due to the high demand on GPU memory and computational resources when handling long contexts. While sparsify the Key-Value (KV) cache of transformer model is a typical strategy to alleviate resource usage, it unavoidably results in the loss of information. We introduce Lossless Compressed Memory Attention (LoMA), a novel approach that enables lossless compression of the KV cache, thereby reducing the memory and computational demands during autoregressive generation. LoMA incorporates a specialized training or fine-tuning precedure alongside an autoregressive generation algorithm optimized for the compressed context. Our method compresses the KV cache after every $tc$ generated tokens with a compression ratio of $c$ and a target compressed length $t$, and this process occurs within a single inference pass without dependency on auxiliary models. We engineered an efficient training scheme involving specific inputs, attention masks, and position identifiers to instill this compression capability. Experimental validation has demonstrated that LoMA significantly reducing computational consumption and memory usage through achieving lossless KV cache compression.
Recently the state space models (SSMs) with efficient hardware-aware designs, i.e., the Mamba deep learning model, have shown great potential for long sequence modeling. Meanwhile building efficient and generic vision backbones purely upon SSMs is an appealing direction. However, representing visual data is challenging for SSMs due to the position-sensitivity of visual data and the requirement of global context for visual understanding. In this paper, we show that the reliance on self-attention for visual representation learning is not necessary and propose a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models. On ImageNet classification, COCO object detection, and ADE20k semantic segmentation tasks, Vim achieves higher performance compared to well-established vision transformers like DeiT, while also demonstrating significantly improved computation & memory efficiency. For example, Vim is 2.8$\times$ faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on images with a resolution of 1248$\times$1248. The results demonstrate that Vim is capable of overcoming the computation & memory constraints on performing Transformer-style understanding for high-resolution images and it has great potential to be the next-generation backbone for vision foundation models. Code is available at https://github.com/hustvl/Vim.
Accurate and robust localization and mapping are essential components for most autonomous robots. In this paper, we propose a SLAM system for building globally consistent maps, called PIN-SLAM, that is based on an elastic and compact point-based implicit neural map representation. Taking range measurements as input, our approach alternates between incremental learning of the local implicit signed distance field and the pose estimation given the current local map using a correspondence-free, point-to-implicit model registration. Our implicit map is based on sparse optimizable neural points, which are inherently elastic and deformable with the global pose adjustment when closing a loop. Loops are also detected using the neural point features. Extensive experiments validate that PIN-SLAM is robust to various environments and versatile to different range sensors such as LiDAR and RGB-D cameras. PIN-SLAM achieves pose estimation accuracy better or on par with the state-of-the-art LiDAR odometry or SLAM systems and outperforms the recent neural implicit SLAM approaches while maintaining a more consistent, and highly compact implicit map that can be reconstructed as accurate and complete meshes. Finally, thanks to the voxel hashing for efficient neural points indexing and the fast implicit map-based registration without closest point association, PIN-SLAM can run at the sensor frame rate on a moderate GPU. Codes will be available at: https://github.com/PRBonn/PIN_SLAM.
Deep neural network (DNN) inference is increasingly being executed on mobile and embedded platforms due to several key advantages in latency, privacy and always-on availability. However, due to limited computing resources, efficient DNN deployment on mobile and embedded platforms is challenging. Although many hardware accelerators and static model compression methods were proposed by previous works, at system runtime, multiple applications are typically executed concurrently and compete for hardware resources. This raises two main challenges: Runtime Hardware Availability and Runtime Application Variability. Previous works have addressed these challenges through either dynamic neural networks that contain sub-networks with different performance trade-offs or runtime hardware resource management. In this thesis, we proposed a combined method, a system was developed for DNN performance trade-off management, combining the runtime trade-off opportunities in both algorithms and hardware to meet dynamically changing application performance targets and hardware constraints in real time. We co-designed novel Dynamic Super-Networks to maximise runtime system-level performance and energy efficiency on heterogeneous hardware platforms. Compared with SOTA, our experimental results using ImageNet on the GPU of Jetson Xavier NX show our model is 2.4x faster for similar ImageNet Top-1 accuracy, or 5.1% higher accuracy at similar latency. We also designed a hierarchical runtime resource manager that tunes both dynamic neural networks and DVFS at runtime. Compared with the Linux DVFS governor schedutil, our runtime approach achieves up to a 19% energy reduction and a 9% latency reduction in single model deployment scenario, and an 89% energy reduction and a 23% latency reduction in a two concurrent model deployment scenario.
Homogeneous diffusion inpainting can reconstruct missing image areas with high quality from a sparse subset of known pixels, provided that their location as well as their gray or color values are well optimized. This property is exploited in inpainting-based image compression, which is a promising alternative to classical transform-based codecs such as JPEG and JPEG2000. However, optimizing the inpainting data is a challenging task. Current approaches are either fairly slow or do not produce high quality results. As a remedy we propose fast spatial and tonal optimization algorithms for homogeneous diffusion inpainting that efficiently utilize GPU parallelism, with a careful adaptation of some of the most successful numerical concepts. We propose a densification strategy using ideas from error-map dithering combined with a Delaunay triangulation for the spatial optimization. For the tonal optimization we design a domain decomposition solver that solves the corresponding normal equations in a matrix-free fashion and supplement it with a Voronoi-based initialization strategy. With our proposed methods we are able to generate high quality inpainting masks for homogeneous diffusion and optimized tonal values in a runtime that outperforms prior state-of-the-art by a wide margin.
In recent years inpainting-based compression methods have been shown to be a viable alternative to classical codecs such as JPEG and JPEG2000. Unlike transform-based codecs, which store coefficients in the transform domain, inpainting-based approaches store a small subset of the original image pixels and reconstruct the image from those by using a suitable inpainting operator. A good candidate for such an inpainting operator is homogeneous diffusion inpainting, as it is simple, theoretically well-motivated, and can achieve good reconstruction quality for optimized data. However, a major challenge has been to design fast solvers for homogeneous diffusion inpainting that scale to 4K image resolution ($3840 \times 2160$ pixels) and are real-time capable. We overcome this with a careful adaptation and fusion of two of the most efficient concept from numerical analysis: multigrid and domain decomposition. Our domain decomposition algorithm efficiently utilizes GPU parallelism by solving inpainting problems on small overlapping blocks. Unlike simple block decomposition strategies such as the ones in JPEG, our approach yields block artifact-free reconstructions. Furthermore, embedding domain decomposition in a full multigrid scheme provides global interactions and allows us to achieve optimal convergence by reducing both low- and high-frequency errors at the same rate. We are able to achieve 4K color image reconstruction at more than $60$ frames per second even from very sparse data - something which has been previously unfeasible.
Speech bandwidth extension (BWE) refers to widening the frequency bandwidth range of speech signals, enhancing the speech quality towards brighter and fuller. This paper proposes a generative adversarial network (GAN) based BWE model with parallel prediction of Amplitude and Phase spectra, named AP-BWE, which achieves both high-quality and efficient wideband speech waveform generation. The proposed AP-BWE generator is entirely based on convolutional neural networks (CNNs). It features a dual-stream architecture with mutual interaction, where the amplitude stream and the phase stream communicate with each other and respectively extend the high-frequency components from the input narrowband amplitude and phase spectra. To improve the naturalness of the extended speech signals, we employ a multi-period discriminator at the waveform level and design a pair of multi-resolution amplitude and phase discriminators at the spectral level, respectively. Experimental results demonstrate that our proposed AP-BWE achieves state-of-the-art performance in terms of speech quality for BWE tasks targeting sampling rates of both 16 kHz and 48 kHz. In terms of generation efficiency, due to the all-convolutional architecture and all-frame-level operations, the proposed AP-BWE can generate 48 kHz waveform samples 292.3 times faster than real-time on a single RTX 4090 GPU and 18.1 times faster than real-time on a single CPU. Notably, to our knowledge, AP-BWE is the first to achieve the direct extension of the high-frequency phase spectrum, which is beneficial for improving the effectiveness of existing BWE methods.