arXiv subjects

Yuedong Xu

Publications and source records attributed to Yuedong Xu.

At least 19 recordsLinked to original sources

arXiv2026

DualDecoder: Accelerate Long Context LLM Inference by Predictive Prefetch

Long-context inference is becoming a fundamental capability for modern LLM serving, especially driven by emerging agentic applications. Yet it faces a severe memory wall that the KV cache scales proportionally with increasing context length and request concurrency. Existing sparse KV cache methods offload most KV entries to host memory and retrieve only the critical KV entries needed by each decoding step. However, they commonly introduce substantial auxiliary states in GPU memory for KV retrieval management. Our measurements show that these often-overlooked auxiliary states introduce significant memory overhead and become a new bottleneck under high-concurrency workloads. In this paper, we present DualDecoder, a lightweight serving system for long-context LLM inference that enables efficient sparse KV cache retrieval from host memory. Our key insight is that the critical KV entries required for decoding the next token can be accurately predicted from the preceding speculated token. This predictability enables KV retrieval to be proactively prefetched and overlapped with decoding computation, effectively eliminating the GPU memory overhead of auxiliary states. To achieve this prefetching efficiently, DualDecoder leverages a novel dual-token decoding pipeline that accurately identifies critical KV entries with negligible computational overhead, and designs a layer-aware transfer schedule to overlap KV prefetching with model computation and a layer-scoped memory manager to reduce the GPU runtime buffer. Experimental results show that DualDecoder improves decoding throughput by up to 2.62$\times$ over state-of-the-art systems while preserving decoding latency and model quality.

cs.DC↗

arXiv2026

Pragma-VL: Towards a Pragmatic Arbitration of Safety and Helpfulness in MLLMs

Multimodal Large Language Models (MLLMs) pose critical safety challenges, as they are susceptible not only to adversarial attacks such as jailbreaking but also to inadvertently generating harmful content for benign users. While internal safety alignment via Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) is a primary mitigation strategy, current methods often face a safety-utility trade-off: they either refuse benign queries out of excessive caution or overlook latent risks in cross-modal interactions. To resolve this, we introduce Pragma-VL, an end-to-end alignment algorithm that enables MLLMs to pragmatically arbitrate between safety and helpfulness. First, we enhance visual risk perception with a novel cold-start SFT stage. This is achieved by applying risk-aware clustering to the visual encoder and using an interleaved dataset of risk descriptions and high-quality data. Second, we introduce a theoretically-guaranteed reward model that leverages synergistic learning. We train it with a novel data augmentation method that assigns dynamic weights based on the queries, enabling contextual arbitration between safety and helpfulness. Extensive experiments show that Pragma-VL effectively balances safety and helpfulness, outperforming baselines by 5% to 20% on most multimodal safety benchmarks while preserving its general capabilities in areas such as mathematics and knowledge reasoning.

cs.LG↗

arXiv2025

Determining Blockchain Transaction Timing and Fee with Observable Mempools

Transaction fee plays an important role in determining the priority of transaction processing in public blockchain systems. Owing to the observability of unconfirmed transactions, a strategic user can postpone his transaction broadcasting time and set a fee as low as possible by prying into his mempool that stores them. However, the stochastic mining interval may cause the delayed transaction to miss the next valid block. Meanwhile, a new feature (i.e. fee bumping) emerges that allows each user to increase his transaction fee before confirmation, making the fee setting more challenging. In this paper, we investigate a novel transaction policy from the perspective of a single strategic user that determines the broadcasting time and the transaction fee simultaneously. Two representative scenarios are considered, in which a number of coexisting ordinary users are mempool-oblivious that set their fees according to certain distribution, and are semi-strategic that check their mempools at a Poisson rate and update their fees. In the former, we compute the optimal broadcasting time and transaction fee that adapts to the arbitrary distribution of mining interval. When the block interval is exponentially distributed in Bitcoin-like PoW systems, the strategic user needs to broadcast his transaction immediately after its creation. And when the block interval is fixed in Ethereum-like PoS systems, he finds it profitable to wait until the last moment before block generation. In the latter, we formulate a continuous-time Markov chain to characterize the dynamics of mempool states, and derive the optimal fee adjusting frequency of the strategic user when the block interval is exponentially distributed. In both theory and simulations, we show that this strategic user should immediately increase his fee whenever it falls behind the minimum fee of being included.

cs.GT↗

arXiv2025

Diving into 3D Parallelism with Heterogeneous Spot Instance GPUs: Design and Implications

The rapid growth of large language models (LLMs) and the continuous release of new GPU products have significantly increased the demand for distributed training across heterogeneous GPU environments. In this paper, we present a comprehensive analysis of the challenges involved in implementing 3D parallelism in such environments, addressing critical issues such as the need for symmetric tensor parallelism, efficient gradient synchronization in asymmetric pipeline parallelism, and the trade-offs between memory utilization and computational efficiency. Building upon these insights, we introduce AutoHet, a novel system that automatically identifies the optimal parallelism plan for distributed training on heterogeneous GPUs. AutoHet supports asymmetric 3D parallelism structures and facilitates fine-grained workload distribution. We propose a theoretical model that frames the device grouping and load balancing as an optimization problem to minimize per-iteration training time, thus effectively balancing computing power and memory usage across GPUs with diverse capabilities. To enable elastic training upon spot instance preemption, AutoHet presents an efficient recovery strategy that prioritizes to retrieve training states from local nodes, and only downloads the missing checkpoints from the cloud storage. Our extensive evaluation, conducted on three large-scale models and utilizing combinations of three different GPU types, demonstrates that AutoHet outperforms existing DNN training systems, achieving up to a 1.79$\times$ speedup in training throughput compared with Megatron-LM and Whale, and a 4.38$\times$ speedup of recovery speed compared to a spot instance baseline.

cs.DC↗

arXiv2025

Deadline-Aware Online Scheduling for LLM Fine-Tuning with Spot Market Predictions

As foundation models grow in size, fine-tuning them becomes increasingly expensive. While GPU spot instances offer a low-cost alternative to on-demand resources, their volatile prices and availability make deadline-aware scheduling particularly challenging. We tackle this difficulty by using a mix of spot and on-demand instances. Distinctively, we show the predictability of prices and availability in a spot instance market, the power of prediction in enabling cost-efficient scheduling and its sensitivity to estimation errors. An integer programming problem is formulated to capture the use of mixed instances under both the price and availability dynamics. We propose an online allocation algorithm with prediction based on the committed horizon control approach that leverages a \emph{commitment level} to enforce the partial sequence of decisions. When this prediction becomes inaccurate, we further present a complementary online algorithm without predictions. An online policy selection algorithm is developed that learns the best policy from a pool constructed by varying the parameters of both algorithms. We prove that the prediction-based algorithm achieves tighter performance bounds as prediction error decreases, while the policy selection algorithm possesses a regret bound of $\mathcal{O}(\sqrt{T})$. Experimental results demonstrate that our online framework can adaptively select the best policy under varying spot market dynamics and prediction quality, consistently outperforming baselines and improving utility by up to 54.8\%.

cs.DC↗

arXiv2025

A Modular and Scalable Simulator for Connected-UAVs Communication in 5G Networks

Cellular-connected UAV systems have enabled a wide range of low-altitude aerial services. However, these systems still face many challenges, such as frequent handovers and the inefficiency of traditional transport protocols. To better study these issues, we develop a modular and scalable simulation platform specifically designed for UAVs communication leveraging the research ecology in wireless communication of MATLAB. The platform supports flexible 5G NR node deployment, customizable UAVs mobility models, and multi-network-interface extensions. It also supports multiple transport protocols including TCP, UDP, QUIC, etc., allowing to investigate how different transport protocols affect UAVs communication performance. In addition, the platform includes a handover management module, enabling the evaluation of both traditional and learning-based handover strategies. Our platform can serve as a testbed for the development and evaluation of advanced transmission strategies in cellular-connected UAV systems.

cs.NI↗

arXiv2025

Differentially Private Federated Low Rank Adaptation Beyond Fixed-Matrix

Large language models (LLMs) typically require fine-tuning for domain-specific tasks, and LoRA offers a computationally efficient approach by training low-rank adapters. LoRA is also communication-efficient for federated LLMs when multiple users collaboratively fine-tune a global LLM model without sharing their proprietary raw data. However, even the transmission of local adapters between a server and clients risks serious privacy leakage. Applying differential privacy (DP) to federated LoRA encounters a dilemma: adding noise to both adapters amplifies synthetic noise on the model, while fixing one adapter impairs the learnability of fine-tuning. In this paper, we propose FedASK (Differentially Private Federated Low Rank Adaptation with Double Sketching) , a novel federated LoRA framework to enable effective updating of both low-rank adapters with robust differential privacy. Inspired by randomized SVD, our key idea is a two-stage sketching pipeline. This pipeline first aggregates carefully sketched, privacy-preserving local updates, and then reconstructs the global matrices on the server to facilitate effective updating of both adapters. We theoretically prove FedASK's differential privacy guarantee and its exact aggregation property. Comprehensive experiments demonstrate that FedASK consistently outperforms baseline methods across a variety of privacy settings and data distributions.

cs.CR↗

arXiv2024

A Communication and Computation Efficient Fully First-order Method for Decentralized Bilevel Optimization

Bilevel optimization, crucial for hyperparameter tuning, meta-learning and reinforcement learning, remains less explored in the decentralized learning paradigm, such as decentralized federated learning (DFL). Typically, decentralized bilevel methods rely on both gradients and Hessian matrices to approximate hypergradients of upper-level models. However, acquiring and sharing the second-order oracle is compute and communication intensive. % and sharing this information incurs heavy communication overhead. To overcome these challenges, this paper introduces a fully first-order decentralized method for decentralized Bilevel optimization, $\text{C}^2$DFB which is both compute- and communicate-efficient. In $\text{C}^2$DFB, each learning node optimizes a min-min-max problem to approximate hypergradient by exclusively using gradients information. To reduce the traffic load at the inner-loop of solving the lower-level problem, $\text{C}^2$DFB incorporates a lightweight communication protocol for efficiently transmitting compressed residuals of local parameters. % during the inner loops. Rigorous theoretical analysis ensures its convergence % of the algorithm, indicating a first-order oracle calls of $\tilde{\mathcal{O}}(ε^{-4})$. Experiments on hyperparameter tuning and hyper-representation tasks validate the superiority of $\text{C}^2$DFB across various typologies and heterogeneous data distributions.

cs.LG↗

arXiv2024

AxiomVision: Accuracy-Guaranteed Adaptive Visual Model Selection for Perspective-Aware Video Analytics

The rapid evolution of multimedia and computer vision technologies requires adaptive visual model deployment strategies to effectively handle diverse tasks and varying environments. This work introduces AxiomVision, a novel framework that can guarantee accuracy by leveraging edge computing to dynamically select the most efficient visual models for video analytics under diverse scenarios. Utilizing a tiered edge-cloud architecture, AxiomVision enables the deployment of a broad spectrum of visual models, from lightweight to complex DNNs, that can be tailored to specific scenarios while considering camera source impacts. In addition, AxiomVision provides three core innovations: (1) a dynamic visual model selection mechanism utilizing continual online learning, (2) an efficient online method that efficiently takes into account the influence of the camera's perspective, and (3) a topology-driven grouping approach that accelerates the model selection process. With rigorous theoretical guarantees, these advancements provide a scalable and effective solution for visual tasks inherent to multimedia systems, such as object detection, classification, and counting. Empirically, AxiomVision achieves a 25.7\% improvement in accuracy.

cs.MM↗

arXiv2024

A Queueing Theoretic Perspective on Low-Latency LLM Inference with Variable Token Length

Large language models (LLMs) propel the prosperity of interactive AI applications showcased by ChatGPT that demand timely response of inference services. However, LLM inference is computation intensive and memory intensive, and improper parameter configuration at LLM platforms may exacerbate the inference time. In this paper, we analyze the impact of LLM output token distribution on the inference queueing delay, where the max-token clipping and the batched inference are considered. By formulating an M/G/1 model, we observe that enforcing a maximum output token limit on a very small fraction of inference requests can significantly reduce the queueing delay, and our model facilitates the selection of the optimal limit. For the batch inference, we model the service process as a bulk queue in which the batch processing time is affected by the batch size and the maximum token size inside this batch jointly. The queueing delays of the batching of all buffered requests (dynamic batching), the batching of constant number of requests (fixed batching), and the batching without intra-batch waiting (elastic batching) are derived. Experimental results show that our mathematical models coincide with the event-driven simulations well.

cs.NI↗

arXiv2024

Hierarchical Reinforcement Learning Empowered Task Offloading in V2I Networks

Edge computing plays an essential role in the vehicle-to-infrastructure (V2I) networks, where vehicles offload their intensive computation tasks to the road-side units for saving energy and reduce the latency. This paper designs the optimal task offloading policy to address the concerns involving processing delay, energy consumption and edge computing cost. Each computation task consisting of some interdependent sub-tasks is characterized as a directed acyclic graph (DAG). In such dynamic networks, a novel hierarchical Offloading scheme is proposed by leveraging deep reinforcement learning (DRL). The inter-dependencies among the DAGs of the computation tasks are extracted using a graph neural network with attention mechanism. A parameterized DRL algorithm is developed to deal with the hierarchical action space containing both discrete and continuous actions. Simulation results with a real-world car speed dataset demonstrate that the proposed scheme can effectively reduce the system overhead.

cs.NI↗

arXiv2024

Optimizing Layerwise Microservice Management in Heterogeneous Wireless Networks

Small cells with edge computing are densely deployed in 5G mobile networks to provide high throughput communication and low-latency computation. The flexibility of edge computation is empowered by the deployment of lightweight container-based microservices. In this paper, we take the first step toward optimizing the microservice management in small-cell networks. The prominent feature is that each microservice consists of multiple image layers and different microservices may share some basic layers, thus bringing deep coupling in their placement and service provision. Our objective is to minimize the expected total latency of microservice requests under the storage, communication and computing constraints of the sparsely interconnected small cell nodes. We formulate a binary quadratic program (BQP) with the multi-dimensional strategy of the image layer placement, the access selection and the task assignment. The BQP problem is then transformed into an ILP problem, and is solved by use of a novel sphere-box alternating direction multipliers method (ADMM) with reasonable complexity $O(q^{4})$, where $q$ is the number of variables in the transformed problem. Trace-driven experiments show that the gap between our proposed algorithm and the optimal is reduced by 35$\%$ compared with benchmark algorithms.

cs.NI↗

arXiv2023

Minerva: Decentralized Collaborative Query Processing over InterPlanetary File System

Data silos create barriers in accessing and utilizing data dispersed over networks. Directly sharing data easily suffers from the long downloading time, the single point failure and the untraceable data usage. In this paper, we present Minerva, a peer-to-peer cross-cluster data query system based on InterPlanetary File System (IPFS). Minerva makes use of the distributed Hash table (DHT) lookup to pinpoint the locations that store content chunks. We theoretically model the DHT query delay and introduce the fat Merkle tree structure as well as the DHT caching to reduce it. We design the query plan for read and write operations on top of Apache Drill that enables the collaborative query with decentralized workers. We conduct comprehensive experiments on Minerva, and the results show that Minerva achieves up to $2.08 \times$ query performance acceleration compared to the original IPFS data query, and could complete data analysis queries on the Internet-like environments within an average latency of $0.615$ second. With collaborative query, Minerva could perform up to $1.39 \times$ performance acceleration than centralized query with raw data shipment.

cs.DB↗

arXiv2023

Blockchain Mining with Multiple Selfish Miners

This paper studies a fundamental problem regarding the security of blockchain PoW consensus on how the existence of multiple misbehaving miners influences the profitability of selfish mining. Each selfish miner (or attacker interchangeably) maintains a private chain and makes it public opportunistically for acquiring more rewards incommensurate to his Hash power. We first establish a general Markov chain model to characterize the state transition of public and private chains for Basic Selfish Mining (BSM), and derive the stationary profitable threshold of Hash power in closed-form. It reduces from 25% for a single attacker to below 21.48% for two symmetric attackers theoretically, and further reduces to around 10% with eight symmetric attackers experimentally. We next explore the profitable threshold when one of the attackers performs strategic mining based on Partially Observable Markov Decision Process (POMDP) that only half of the attributes pertinent to a mining state are observable to him. An online algorithm is presented to compute the nearly optimal policy efficiently despite the large state space and high dimensional belief space. The strategic attacker mines selfishly and more agilely than BSM attacker when his Hash power is relatively high, and mines honestly otherwise, thus leading to a much lower profitable threshold. Last, we formulate a simple model of absolute mining revenue that yields an interesting observation: selfish mining is never profitable at the first difficulty adjustment period, but replying on the reimbursement of stationary selfish mining gains in the future periods. The delay till being profitable of an attacker increases with the decrease of his Hash power, making blockchain miners more cautious on performing selfish mining.

cs.CR↗

arXiv2022

Multi-dimensional Network Security Game: How do attacker and defender battle on parallel targets?

In this paper, we consider a new network security game wherein an attacker and a defender are battling over "multiple" targets. This type of game is appropriate to model many current network security conflicts such as Internet phishing, mobile malware or network intrusions. In such attacks, the attacker and the defender need to decide how to allocate resources on each target so as to maximize his utility within his resource limit. We model such a multi-dimensional network security game as a constrained non-zero sum game. Two security breaching models, the product-form and the proportion-form, are considered. For each breaching model, we prove the existence of a unique Nash equilibrium (NE) based on Rosen's theorem and propose efficient algorithms to find the NE when the games are strictly concave. Furthermore, we show the existence of multiple NEs in the product-form breaching model when the strict concavity does not hold. Our study sheds light on the strategic behaviors of the attacker and the defender, in particular, on how they allocate resources to the targets which have different weights, and how their utilities as well as strategies are influenced by the resource constraints.

cs.GT↗

arXiv2022

Decentralized Stochastic Proximal Gradient Descent with Variance Reduction over Time-varying Networks

In decentralized learning, a network of nodes cooperate to minimize an overall objective function that is usually the finite-sum of their local objectives, and incorporates a non-smooth regularization term for the better generalization ability. Decentralized stochastic proximal gradient (DSPG) method is commonly used to train this type of learning models, while the convergence rate is retarded by the variance of stochastic gradients. In this paper, we propose a novel algorithm, namely DPSVRG, to accelerate the decentralized training by leveraging the variance reduction technique. The basic idea is to introduce an estimator in each node, which tracks the local full gradient periodically, to correct the stochastic gradient at each iteration. By transforming our decentralized algorithm into a centralized inexact proximal gradient algorithm with variance reduction, and controlling the bounds of error sequences, we prove that DPSVRG converges at the rate of $O(1/T)$ for general convex objectives plus a non-smooth term with $T$ as the number of iterations, while DSPG converges at the rate $O(\frac{1}{\sqrt{T}})$. Our experiments on different applications, network topologies and learning models demonstrate that DPSVRG converges much faster than DSPG, and the loss function of DPSVRG decreases smoothly along with the training epochs.

cs.LG↗

arXiv2021

On The Robustness of Price-Anticipating Kelly Mechanism

The price-anticipating Kelly mechanism (PAKM) is one of the most extensively used strategies to allocate divisible resources for strategic users in communication networks and computing systems. The users are deemed as selfish and also benign, each of which maximizes his individual utility of the allocated resources minus his payment to the network operator. However, in many applications a user can use his payment to reduce the utilities of his opponents, thus playing a misbehaving role. It remains mysterious to what extent the misbehaving user can damage or influence the performance of benign users and the network operator. In this work, we formulate a non-cooperative game consisting of a finite amount of benign users and one misbehaving user. The maliciousness of this misbehaving user is captured by his willingness to pay to trade for unit degradation in the utilities of benign users. The network operator allocates resources to all the users via the price-anticipating Kelly mechanism. We present six important performance metrics with regard to the total utility and the total net utility of benign users, and the revenue of network operator under three different scenarios: with and without the misbehaving user, and the maximum. We quantify the robustness of PAKM against the misbehaving actions by deriving the upper and lower bounds of these metrics. With new approaches, all the theoretical bounds are applicable to an arbitrary population of benign users. Our study reveals two important insights: i) the performance bounds are very sensitive to the misbehaving user's willingness to pay at certain ranges; ii) the network operator acquires more revenues in the presence of the misbehaving user which might disincentivize his countermeasures against the misbehaving actions.

cs.SI↗

arXiv2021

CRISLoc: Reconstructable CSI Fingerprintingfor Indoor Smartphone Localization

Channel state information (CSI) based fingerprinting for WIFI indoor localization has attracted lots of attention very recently.The frequency diverse and temporally stable CSI better represents the location dependent channel characteristics than the coarsereceived signal strength (RSS). However, the acquisition of CSI requires the cooperation of access points (APs) and involves only dataframes, which imposes restrictions on real-world deployment. In this paper, we present CRISLoc, the first CSI fingerprinting basedlocalization prototype system using ubiquitous smartphones. CRISLoc operates in a completely passive mode, overhearing thepackets on-the-fly for his own CSI acquisition. The smartphone CSI is sanitized via calibrating the distortion enforced by WiFi amplifiercircuits. CRISLoc tackles the challenge of altered APs with a joint clustering and outlier detection method to find them. A novel transferlearning approach is proposed to reconstruct the high-dimensional CSI fingerprint database on the basis of the outdated fingerprintsand a few fresh measurements, and an enhanced KNN approach is proposed to pinpoint the location of a smartphone. Our studyreveals important properties about the stability and sensitivity of smartphone CSI that has not been reported previously. Experimentalresults show that CRISLoc can achieve a mean error of around 0.29m in a6m times 8mresearch laboratory. The mean error increases by 5.4 cm and 8.6 cm upon the movement of one and two APs, which validates the robustness of CRISLoc against environment changes.

cs.NI↗