Search

Small language models (SLMs), despite their widespread adoption in modern smart devices, have received significantly less academic attention compared to their large language model (LLM) counterparts, which are predominantly deployed in data centers and cloud environments. While researchers continue to improve the capabilities of LLMs in the pursuit of artificial general intelligence, SLM research aims to make machine intelligence more accessible, affordable, and efficient for everyday tasks. Focusing on transformer-based, decoder-only language models with 100M-5B parameters, we survey 59 state-of-the-art open-source SLMs, analyzing their technical innovations across three axes: architectures, training datasets, and training algorithms. In addition, we evaluate their capabilities in various domains, including commonsense reasoning, in-context learning, mathematics, and coding. To gain further insight into their on-device runtime costs, we benchmark their inference latency and memory footprints. Through in-depth analysis of our benchmarking data, we offer valuable insights to advance research in this field.

Air-ground robots (AGRs) are widely used in surveillance and disaster response due to their exceptional mobility and versatility (i.e., flying and driving). Current AGR navigation systems perform well in static occlusion-prone environments (e.g., indoors) by using 3D semantic occupancy networks to predict occlusions for complete local mapping and then computing Euclidean Signed Distance Field (ESDF) for path planning. However, these systems face challenges in dynamic, severe occlusion scenes (e.g., crowds) due to limitations in perception networks' low prediction accuracy and path planners' high computation overhead. In this paper, we propose OMEGA, which contains OccMamba with an Efficient AGR-Planner to address the above-mentioned problems. OccMamba adopts a novel architecture that separates semantic and occupancy prediction into independent branches, incorporating two mamba blocks within these branches. These blocks efficiently extract semantic and geometric features in 3D environments with linear complexity, ensuring that the network can learn long-distance dependencies to improve prediction accuracy. Semantic and geometric features are combined within the Bird's Eye View (BEV) space to minimise computational overhead during feature fusion. The resulting semantic occupancy map is then seamlessly integrated into the local map, providing occlusion awareness of the dynamic environment. Our AGR-Planner utilizes this local map and employs kinodynamic A* search and gradient-based trajectory optimization to guarantee planning is ESDF-free and energy-efficient. Extensive experiments demonstrate that OccMamba outperforms the state-of-the-art 3D semantic occupancy network with 25.0% mIoU. End-to-end navigation experiments in dynamic scenes verify OMEGA's efficiency, achieving a 96% average planning success rate. Code and video are available at https://jmwang0117.github.io/OMEGA/., Comment: OccMamba is Coming!

The era of large deep learning models has given rise to advanced training strategies such as 3D parallelism and the ZeRO series. These strategies enable various (re-)configurable execution plans for a training job, which exhibit remarkably different requirements of multiple resource types. Existing cluster scheduling systems, however, treat such reconfigurable training jobs as black boxes: they rely on users to choose execution plans statically, and then make resource allocations without awareness of the chosen plans and their resource requirements. This approach results in mismatches between execution plans and resources, making both training performance and cluster utilization far from optimal. We introduce Rubick, a cluster scheduling system for deep learning training that exploits the reconfigurability to improve job performance and cluster efficiency. Rubick incorporates the job execution planning as a new dimension in cluster scheduling, by continuously reconfiguring jobs' execution plans and tuning multi-resource allocations across jobs jointly. Such a co-optimization is navigated by a performance model that understands the diverse resource requirements and performance characteristics of different jobs and execution plans. Rubick exploits such a model to make performance-aware scheduling decisions to maximize cluster throughput while providing performance guarantees to individual jobs. Evaluations on a 64-GPU high-performance training cluster show that Rubick improves average job completion time and makespan by up to 3.2x and 1.4x, respectively, compared against state-of-the-art systems.

The rapid advancements in machine learning techniques have led to significant achievements in various real-world robotic tasks. These tasks heavily rely on fast and energy-efficient inference of deep neural network (DNN) models when deployed on robots. To enhance inference performance, distributed inference has emerged as a promising approach, parallelizing inference across multiple powerful GPU devices in modern data centers using techniques such as data parallelism, tensor parallelism, and pipeline parallelism. However, when deployed on real-world robots, existing parallel methods fail to provide low inference latency and meet the energy requirements due to the limited bandwidth of robotic IoT. We present Hybrid-Parallel, a high-performance distributed inference system optimized for robotic IoT. Hybrid-Parallel employs a fine-grained approach to parallelize inference at the granularity of local operators within DNN layers (i.e., operators that can be computed independently with the partial input, such as the convolution kernel in the convolution layer). By doing so, Hybrid-Parallel enables different operators of different layers to be computed and transmitted concurrently, and overlap the computation and transmission phases within the same inference task. The evaluation demonstrate that Hybrid-Parallel reduces inference time by 14.9% ~41.1% and energy consumption per inference by up to 35.3% compared to the state-of-the-art baselines.

Deep Neural Network (DNN) inference on serverless functions is gaining prominence due to its potential for substantial budget savings. Existing works on serverless DNN inference solely optimize batching requests from one application with a single Service Level Objective (SLO) on CPU functions. However, production serverless DNN inference traces indicate that the request arrival rate of applications is surprisingly low, which inevitably causes a long batching time and SLO violations. Hence, there is an urgent need for batching multiple DNN inference requests with diverse SLOs (i.e., multi-SLO DNN inference) in serverless platforms. Moreover, the potential performance and cost benefits of deploying heterogeneous (i.e., CPU and GPU) functions for DNN inference have received scant attention. In this paper, we present HarmonyBatch, a cost-efficient resource provisioning framework designed to achieve predictable performance for multi-SLO DNN inference with heterogeneous serverless functions. Specifically, we construct an analytical performance and cost model of DNN inference on both CPU and GPU functions, by explicitly considering the GPU time-slicing scheduling mechanism and request arrival rate distribution. Based on such a model, we devise a two-stage merging strategy in HarmonyBatch to judiciously batch the multi-SLO DNN inference requests into application groups. It aims to minimize the budget of function provisioning for each application group while guaranteeing diverse performance SLOs of inference applications. We have implemented a prototype of HarmonyBatch on Alibaba Cloud Function Compute. Extensive prototype experiments with representative DNN inference workloads demonstrate that HarmonyBatch can provide predictable performance to serverless DNN inference workloads while reducing the monetary cost by up to 82.9% compared to the state-of-the-art methods., Comment: 10 pages, 14 figures, accepted by IWQOS24

Next-generation mobile networks are expected to facilitate fast AI model downloading to end users. By caching models on edge servers, mobile networks can deliver models to end users with low latency, resulting in a paradigm called edge model caching. In this paper, we develop a novel model placement scheme, called parameter-sharing model caching (TrimCaching). TrimCaching exploits the key observation that a wide range of AI models, such as convolutional neural networks or large language models, can share a significant proportion of parameter blocks containing reusable knowledge, thereby improving storage efficiency. To this end, we formulate a parameter-sharing model placement problem to maximize the cache hit ratio in multi-edge wireless networks by balancing the fundamental tradeoff between storage efficiency and service latency. We show that the formulated problem is a submodular maximization problem with submodular constraints, for which no polynomial-time approximation algorithm exists. To overcome this challenge, we study an important special case, where a small fixed number of parameter blocks are shared across models, which often holds in practice. In such a case, a polynomial-time algorithm with $\left(1-\epsilon\right)/2$-approximation guarantee is developed. Subsequently, we address the original problem for the general case by developing a greedy algorithm. Simulation results demonstrate that the proposed TrimCaching framework significantly improves the cache hit ratio compared with state-of-the-art content caching without exploiting shared parameters in AI models., Comment: 11 pages, 7 figures. This paper has been accepted by ICDCS 2024. The extended version of this paper is at arXiv:2404.14204

Next-generation mobile networks are expected to facilitate fast AI model downloading to end users. By caching models on edge servers, mobile networks can deliver models to end users with low latency, resulting in a paradigm called edge model caching. In this paper, we develop a novel model placement scheme, called parameter-sharing model caching (TrimCaching). TrimCaching exploits the key observation that a wide range of AI models, such as convolutional neural networks or large language models, can share a significant proportion of parameter blocks containing reusable knowledge, thereby improving storage efficiency. To this end, we formulate a parameter-sharing model placement problem to maximize the cache hit ratio in multi-edge wireless networks by balancing the fundamental tradeoff between storage efficiency and service latency. We show that the formulated problem is a submodular maximization problem with submodular constraints, for which no polynomial-time approximation algorithm exists. To overcome this challenge, we study an important special case, where a small fixed number of parameter blocks are shared across models, which often holds in practice. In such a case, a polynomial-time algorithm with $\left(1-\epsilon\right)/2$-approximation guarantee is developed. Subsequently, we address the original problem for the general case by developing a greedy algorithm. Simulation results demonstrate that the proposed TrimCaching framework significantly improves the cache hit ratio compared with state-of-the-art content caching without exploiting shared parameters in AI models., Comment: 15 pages, 11 figures. Part of this work has been accepted by ICDCS 2024

Deep neural networks (DNNs) have been widely adopted for various mobile inference tasks, yet their ever-increasing computational demands are hindering their deployment on resource-constrained mobile devices. Hybrid deep learning partitions a DNN into two parts and deploys them across the mobile device and a server, aiming to reduce inference latency or prolong battery life of mobile devices. However, such partitioning produces (non-uniform) DNN fragments which are hard to serve efficiently on the server.This paper presents Graft -- an efficient inference serving system for hybrid deep learning with latency service-level objective (SLO) guarantees. Our main insight is to mitigate the non-uniformity by a core concept called DNN re-alignment, allowing multiple heterogeneous DNN fragments to be restructured to share layers. To fully exploit the potential of DNN re-alignment, Graft employs fine-grained GPU resource sharing. Based on that, we propose efficient algorithms for merging, grouping, and re-aligning DNN fragments to maximize request batching opportunities, minimizing resource consumption while guaranteeing the inference latency SLO. We implement a Graft prototype and perform extensive experiments with five types of widely used DNNs and real-world network traces. Our results show that Graft improves resource efficiency by up to 70% compared with the state-of-the-art inference serving systems.

GPUs have become the \emph{defacto} hardware devices for accelerating Deep Neural Network (DNN) inference workloads. However, the conventional \emph{sequential execution mode of DNN operators} in mainstream deep learning frameworks cannot fully utilize GPU resources, even with the operator fusion enabled, due to the increasing complexity of model structures and a greater diversity of operators. Moreover, the \emph{inadequate operator launch order} in parallelized execution scenarios can lead to GPU resource wastage and unexpected performance interference among operators. In this paper, we propose \emph{Opara}, a resource- and interference-aware DNN \underline{Op}erator \underline{para}llel scheduling framework to accelerate DNN inference on GPUs. Specifically, \emph{Opara} first employs \texttt{CUDA Streams} and \texttt{CUDA Graph} to \emph{parallelize} the execution of multiple operators automatically. To further expedite DNN inference, \emph{Opara} leverages the resource demands of operators to judiciously adjust the operator launch order on GPUs, overlapping the execution of compute-intensive and memory-intensive operators. We implement and open source a prototype of \emph{Opara} based on PyTorch in a \emph{non-intrusive} manner. Extensive prototype experiments with representative DNN and Transformer-based models demonstrate that \emph{Opara} outperforms the default sequential \texttt{CUDA Graph} in PyTorch and the state-of-the-art operator parallelism systems by up to $1.68\times$ and $1.29\times$, respectively, yet with acceptable runtime overhead., Comment: 9 pages,10 figures

The metaverse has recently gained increasing attention from the public. It builds up a virtual world where we can live as a new role regardless of the role we play in the physical world. However, building and operating this virtual world will generate an extraordinary amount of carbon emissions for computing, communicating, displaying, and so on. This inevitably hinders the realization of carbon neutrality as a priority of our society, adding heavy burden to our earth. In this survey, we first present a green viewpoint of the metaverse by investigating the carbon issues in its three core layers, namely the infrastructure layer, the interaction layer, and the economy layer, and estimate their carbon footprints in the near future. Next, we analyze a range of current and emerging applicable green techniques for the purpose of reducing energy usage and carbon emissions of the metaverse, and discuss their limitations in supporting metaverse workloads. Then, in view of these limitations, we discuss important implications and bring forth several insights and future directions to make each metaverse layer greener. After that, we investigate green solutions from the governance perspective, including both public policies in the physical world and regulation of users in the virtual world, and propose an indicator Carbon Utility (CU) to quantify the service quality brought by an user activity per unit of carbon emissions. Finally, we identify an issue for the metaverse as a whole and summarize three directions: (1) a comprehensive consideration of necessary performance metrics, (2) a comprehensive consideration of involved layers and multiple internal components, and (3) a new assessing, recording, and regulating mechanism on carbon footprints of user activities. Our proposed quantitative indicator CU would be helpful in regulating user activities in the metaverse world., Comment: 24 pages

GPUs are essential to accelerating the latency-sensitive deep neural network (DNN) inference workloads in cloud datacenters. To fully utilize GPU resources, spatial sharing of GPUs among co-located DNN inference workloads becomes increasingly compelling. However, GPU sharing inevitably brings severe performance interference among co-located inference workloads, as motivated by an empirical measurement study of DNN inference on EC2 GPU instances. While existing works on guaranteeing inference performance service level objectives (SLOs) focus on either temporal sharing of GPUs or reactive GPU resource scaling and inference migration techniques, how to proactively mitigate such severe performance interference has received comparatively little attention. In this paper, we propose iGniter, an interference-aware GPU resource provisioning framework for cost-efficiently achieving predictable DNN inference in the cloud. iGniter is comprised of two key components: (1) a lightweight DNN inference performance model, which leverages the system and workload metrics that are practically accessible to capture the performance interference; (2) A cost-efficient GPU resource provisioning strategy that jointly optimizes the GPU resource allocation and adaptive batching based on our inference performance model, with the aim of achieving predictable performance of DNN inference workloads. We implement a prototype of iGniter based on the NVIDIA Triton inference server hosted on EC2 GPU instances. Extensive prototype experiments on four representative DNN models and datasets demonstrate that iGniter can guarantee the performance SLOs of DNN inference workloads with practically acceptable runtime overhead, while saving the monetary cost by up to 25% in comparison to the state-of-the-art GPU resource provisioning strategies., Comment: 16 pages, 21 figures, submitted to IEEE Transactions on Parallel and Distributed Systems

On edge devices, data scarcity occurs as a common problem where transfer learning serves as a widely-suggested remedy. Nevertheless, transfer learning imposes a heavy computation burden to resource-constrained edge devices. Existing task allocation works usually assume all submitted tasks are equally important, leading to inefficient resource allocation at a task level when directly applied in Multi-task Transfer Learning (MTL). To address these issues, we first reveal that it is crucial to measure the impact of tasks on overall decision performance improvement and quantify \emph{task importance}. We then show that task allocation with task importance for MTL (TATIM) is a variant of the NP-complete Knapsack problem, where the complicated computation to solve this problem needs to be conducted repeatedly under varying contexts. To solve TATIM with high computational efficiency, we propose a Data-driven Cooperative Task Allocation (DCTA) approach. Finally, we evaluate the performance of DCTA by not only a trace-driven simulation, but also a new comprehensive real-world AIOps case study that bridges model and practice via a new architecture and main components design within the AIOps system. Extensive experiments show that our DCTA reduces 3.24 times of processing time, and saves 48.4\% energy consumption compared with the state-of-the-art when solving TATIM., Comment: 15 pages, published in IEEE TRANSACTIONS ON Parallel and Distributed Systems, VOL. 31, NO. 6, JUNE 2020

A new adaptive shaping method that can generate arbitrary optical waveforms with folded-type or fan-type birefringent variable shapers is proposed in this paper. Automatic arbitrary laser temporal shaping of picosecond and femtosecond lasers in shapers with material dispersion, nonidentical birefringent retarders are also studied and results are presented demonstrating the power of the method., Comment: 27 pages

This paper reports the study and demonstration of a new variable temporal shaping method capable of generating linearly polarized picosecond laser pulses with arbitrary predefined shapes, which are highly desired by various applications including low emittance high brightness electron bunch generation in photocathode guns. It is found that both high transmittance and high stability of the shaped pulses can be achieved simultaneously when birefringent stages (BSs) are set at specific phase delay. Such variable temporal shaping technique may lead to new opportunities for many potential applications over a wide range of laser wavelengths, pulse repetition rates, time structures and power levels, etc. In addition, a new double-pass variable temporal shaping method is also proposed and could significantly simplify the shaper structure and reduce the cost., Comment: 5 pages, 10 figures; added laser transmittance results in Appendix section, modification of abstract

This paper describes the demonstration of linearly polarized picosecond pulse shaping with variable profiles including symmetric and non-symmetric intensity distributions. Important characteristics such as stability and transmission were studied, resulting in highly reliable performance of this fan-type birefringent shaping system. This variable temporal shaping technique is applicable over a wide range of laser parameters and may lead to new opportunities for many potential applications. A new double-pass variable temporal shaping method that significantly reduces the required crystal quantity is also proposed in this paper., Comment: 8 pages, 6 figures. arXiv admin note: text overlap with arXiv:1810.01717

We report the study and demonstration of a new laser pulse shaping system capable of generating linearly polarized picosecond laser pulses with variable temporal profiles including symmetric intensity distributions such as parabolic, flattop, elliptical, triangular, as well as non-symmetric distributions, which are highly desired by various applications. It is found that both high transmittance and high stability of the shaped pulse can be achieved simultaneously when crystals are set at a specific phase delay through fine control of the crystal temperature. Although multi-crystal pulse stacking with different configurations were reported before particularly for flattop pulse generation, this new configuration leads to new opportunities for many potential applications over a wide range of laser wavelengths, pulse repetition rate, time structures and power levels. A practical double-pass temporal shaping configuration that significantly reduces the number of crystals is also proposed in this paper as a result of this work., Comment: 10 pages, 6 figures

Graphs are an essential data structure that can represent the structure of social networks. Many online companies, in order to provide intelligent and personalized services for their users, aim to comprehensively analyze a significant amount of graph data with different features. One example is k-core decomposition which captures the degree of connectedness in social graphs. The main purpose of this report is to explore a distributed algorithm for k-core decomposition on Apache Giraph. Namely, we would like to determine whether a cluster-based, Giraph implementation of k-core decomposition that we provide is more efficient than a single-machine, disk-based implementation on GraphChi for large networks. In this report, we describe (a) the programming model of Giraph and GraphChi, (b) the specific implementation of k-core decomposition with Giraph, and (c) the result comparison between Giraph and GraphChi. By analyzing the results, we conclude that Giraph is faster than GraphChi when dealing with large data. However, since worker nodes need time to communicate with each other, Giraph is not very efficient for small data.