Searcharxiv⌕ Search

arXiv subjects

Zonghang Li

Publications and source records attributed to Zonghang Li.

At least 19 recordsLinked to original sources
arXiv2026

COPUS: Co-adaptive Parallelism and Batch Size Selection in Large Language Model Training

Training large language models requires jointly configuring two interdependent aspects of the system: the global batch size, which governs statistical efficiency, and the 3D parallelism strategy, which governs hardware throughput. Existing approaches make these decisions independently: optimization work adapts the batch size to track the evolving critical batch size while keeping parallelism fixed, and systems work selects the fastest parallelism for a given fixed batch size without anticipating that the optimal batch size could change. We show that these decisions are tightly coupled: the throughput-optimal parallelism strategy may shift as the global batch size changes, so any method that fixes one while adapting the other operates with a suboptimal configuration for part of the training run. We present COPUS, a system that adaptively tunes the global batch size, parallelism strategy, and micro-batch size as training evolves. COPUS is guided by Goodput, the product of throughput and statistical efficiency, which models both hardware and statistical effects jointly and directly measures useful convergence per unit of wall-clock time. The system combines online gradient noise scale estimation under 3D parallelism with throughput-aware evaluation of candidate configurations, and supports efficient reconfiguration of both batch size and parallelism during training. We evaluate COPUS on LLM pre-training workloads across 1-4 nodes of 8xH100 and 8xMI210 GPUs and model sizes from 3B to 32B parameters, demonstrating average time-to-convergence speedups of 3.9-8.0% over the fastest baseline across four configurations, with peak gains up to 11.1%, including system overheads.

cs.DC↗
arXiv2026

LAPS: A Length-Aware-Prefill LLM Serving System

LAPS identifies and disaggregates requests with different prompt lengths in LLM serving to reduce TTFT latency. While recent systems have decoupled the prefill and decode stages to improve throughput, they still rely on unified scheduling policies that fail to adapt to heterogeneous workload characteristics. We observe that prompt-length variations lead to distinct performance bottlenecks, motivating an adaptive scheduling strategy. LAPS disaggregates multi-turn long-prefill requests from short-prefill ones and introduces a length-aware smart batching mechanism for short-prefill workloads. It adopts a dual-queue design that supports temporal disaggregation on a single prefill instance or spatial disaggregation across multiple instances. For short-prefill batches, a batch waiting window and CUDA Graph-based clustering mitigate interference from heterogeneous computation, reducing batching delay and lowering average latency. In real multi-turn workloads, LAPS reduces prefill latency by over 30\% compared to vanilla SGLang under prefill-decode disaggregation, and further decreases SLO violations by 28\% in multi-instance deployments with vanilla data-parallel configuration. Compared to the SGLang router with load balancing, it further lowers SLO violations by 12\% in multi-GPU settings. Under high concurrency and mixed-request scenarios, LAPS improves request throughput by 35\% serving Qwen2.5-32B model for prefill instance, demonstrating its effectiveness in optimizing heterogeneous LLM serving workloads.

cs.DC↗
arXiv2025

Tequila: Trapping-free Ternary Quantization for Large Language Models

Quantization techniques are essential for the deployment of Large Language Models (LLMs) on edge devices. However, prevailing methods often rely on mixed-precision multiplication that lacks efficient hardware support, making it not feasible. Ternary weight quantization addresses this by constraining weights to {-1, 0, 1}, replacing expensive multiplications with hardware-efficient additions. However, such aggressive compression leads to significant accuracy degradation, even after costly quantization-aware training with massive data. We identify the core issue as deadzone trapping: a large number of weights are trapped at the deadzone boundary. This occurs because these weights receive only noisy, uninformative gradients, preventing stable escape from the deadzone and severely impeding model capacity and optimization. To address this issue, we propose Tequila, a trapping-free quantization optimization method that reactivates deadzone-trapped weights by repurposing them as dynamic biases. This allows the repurposed weights to provide a continuous signal in the forward pass and, critically, receive direct, meaningful gradient signals during backpropagation, thereby enhancing model capacity and optimization with nearly zero inference overhead. Extensive evaluations demonstrate that Tequila outperforms state-of-the-art (SOTA) ternary quantization methods across five benchmarks. Specifically, on the ARC benchmark, it achieves >4% accuracy gain over the SOTA baseline, nearly matching full-precision performance (within <1% gap) with a 3.0x inference speedup. Consequently, Tequila offers a highly practical and efficient implementation for the deployment of advanced LLMs in resource-constrained environments. The code is available at https://github.com/Tencent/AngelSlim.

cs.LG↗
arXiv2025

Learning In Chaos: Efficient Autoscaling and Self-Healing for Multi-Party Distributed Training

Node and link churn in multi-party, cross-region clusters over wide-area networks (WANs) often disrupts distributed training. However, checkpoint-based recovery and cloud-centric autoscaling react slowly and assume centralized control, which is misaligned with the self-governed setup where institutions can freely join and leave. This paper proposes Chaos, a multi-party distributed training system with self-healing and autoscaling, enabling robust and elastic training under churn. It speeds up autoscaling via multi-neighbor state replication and model sharding. We formalize the sharding and assignment as a MINLP that captures WAN heterogeneity, and reduce it to a tractable MILP by analyzing its monotonicity on a divisibility chain. By establishing an equivalence, we derive a greedy algorithm that follows optimality rules and yields the optimal solution in polynomial time. Chaos uses a cluster monitor to track resource and topology changes, and handles scaling events through peer negotiation protocols, enabling fully self-governed autoscaling among institutions. Experiments show that Chaos has substantially lower scale-out delay than Pollux, Elan, and Autoscaling, and handles scale-in, connect-link, and disconnect-link events within 20ms. It also delivers the lowest idle time, showing superior resource use and scalability as the cluster grows.

cs.DC↗
arXiv2025

Prima.cpp: Fast 30-70B LLM Inference on Heterogeneous and Low-Resource Home Clusters

On-device inference offers privacy, offline use, and instant response, but consumer hardware restricts large language models (LLMs) to low throughput and capability. To overcome this challenge, we present prima.cpp, a distributed on-device inference system that runs 30-70B LLMs on consumer home clusters with mixed CPUs/GPUs, insufficient RAM/VRAM, slow disks, Wi-Fi links, and heterogeneous OSs. We introduce pipelined-ring parallelism (PRP) to overlap disk I/O with compute and communication, and address the prefetch-release conflict in mmap-based offloading. We further propose Halda, a heterogeneity-aware scheduler that co-optimizes per-device CPU/GPU workloads and device selection under RAM/VRAM constraints. On four consumer home devices, a 70B model reaches 674 ms/token TPOT with <6% memory pressure, and a 32B model with speculative decoding achieves 26 tokens/s. Compared with llama.cpp, exo, and dllama, our proposed prima.cpp achieves 5-17 times lower TPOT, supports fine-grained model sizes from 8B to 70B, ensures broader cross-OS and quantization compatibility, and remains OOM-free, while also being Wi-Fi tolerant, privacy-preserving, and hardware-independent. The code is available at https://github.com/OpenCPIL/prima.cpp.

cs.DC↗
arXiv2024

TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices

Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

cs.DC↗
arXiv2024

Log2graphs: An Unsupervised Framework for Log Anomaly Detection with Efficient Feature Extraction

In the era of rapid Internet development, log data has become indispensable for recording the operations of computer devices and software. These data provide valuable insights into system behavior and necessitate thorough analysis. Recent advances in text analysis have enabled deep learning to achieve significant breakthroughs in log anomaly detection. However, the high cost of manual annotation and the dynamic nature of usage scenarios present major challenges to effective log analysis. This study proposes a novel log feature extraction model called DualGCN-LogAE, designed to adapt to various scenarios. It leverages the expressive power of large models for log content analysis and the capability of graph structures to encapsulate correlations between logs. It retains key log information while integrating the causal relationships between logs to achieve effective feature extraction. Additionally, we introduce Log2graphs, an unsupervised log anomaly detection method based on the feature extractor. By employing graph clustering algorithms for log anomaly detection, Log2graphs enables the identification of abnormal logs without the need for labeled data. We comprehensively evaluate the feature extraction capability of DualGCN-LogAE and the anomaly detection performance of Log2graphs using public log datasets across five different scenarios. Our evaluation metrics include detection accuracy and graph clustering quality scores. Experimental results demonstrate that the log features extracted by DualGCN-LogAE outperform those obtained by other methods on classic classifiers. Moreover, Log2graphs surpasses existing unsupervised log detection methods, providing a robust tool for advancing log anomaly detection research.

cs.CR↗
arXiv2024

AMSNet: Netlist Dataset for AMS Circuits

Today's analog/mixed-signal (AMS) integrated circuit (IC) designs demand substantial manual intervention. The advent of multimodal large language models (MLLMs) has unveiled significant potential across various fields, suggesting their applicability in streamlining large-scale AMS IC design as well. A bottleneck in employing MLLMs for automatic AMS circuit generation is the absence of a comprehensive dataset delineating the schematic-netlist relationship. We therefore design an automatic technique for converting schematics into netlists, and create dataset AMSNet, encompassing transistor-level schematics and corresponding SPICE format netlists. With a growing size, AMSNet can significantly facilitate exploration of MLLM applications in AMS circuit design. We have made an initial set of netlists public, and will make both our netlist generation tool and the full dataset available upon publishing of this paper.

cs.CV↗
arXiv2024

Accelerating Geo-distributed Machine Learning with Network-Aware Adaptive Tree and Auxiliary Route

Distributed machine learning is becoming increasingly popular for geo-distributed data analytics, facilitating the collaborative analysis of data scattered across data centers in different regions. This paradigm eliminates the need for centralizing sensitive raw data in one location but faces the significant challenge of high parameter synchronization delays, which stems from the constraints of bandwidth-limited, heterogeneous, and fluctuating wide-area networks. Prior research has focused on optimizing the synchronization topology, evolving from starlike to tree-based structures. However, these solutions typically depend on regular tree structures and lack an adequate topology metric, resulting in limited improvements. This paper proposes NetStorm, an adaptive and highly efficient communication scheduler designed to speed up parameter synchronization across geo-distributed data centers. First, it establishes an effective metric for optimizing a multi-root FAPT synchronization topology. Second, a network awareness module is developed to acquire network knowledge, aiding in topology decisions. Third, a multipath auxiliary transmission mechanism is introduced to enhance network awareness and facilitate multipath transmissions. Lastly, we design policy consistency protocols to guarantee seamless updates of transmission policies. Empirical results demonstrate that NetStorm significantly outperforms distributed training systems like MXNET, MLNET, and TSEngine, with a speedup of 6.5~9.2 times over MXNET.

cs.DC↗
arXiv2023

Enhancing Deep Reinforcement Learning: A Tutorial on Generative Diffusion Models in Network Optimization

Generative Diffusion Models (GDMs) have emerged as a transformative force in the realm of Generative Artificial Intelligence (GenAI), demonstrating their versatility and efficacy across various applications. The ability to model complex data distributions and generate high-quality samples has made GDMs particularly effective in tasks such as image generation and reinforcement learning. Furthermore, their iterative nature, which involves a series of noise addition and denoising steps, is a powerful and unique approach to learning and generating data. This paper serves as a comprehensive tutorial on applying GDMs in network optimization tasks. We delve into the strengths of GDMs, emphasizing their wide applicability across various domains, such as vision, text, and audio generation. We detail how GDMs can be effectively harnessed to solve complex optimization problems inherent in networks. The paper first provides a basic background of GDMs and their applications in network optimization. This is followed by a series of case studies, showcasing the integration of GDMs with Deep Reinforcement Learning (DRL), incentive mechanism design, Semantic Communications (SemCom), Internet of Vehicles (IoV) networks, etc. These case studies underscore the practicality and efficacy of GDMs in real-world scenarios, offering insights into network design. We conclude with a discussion on potential future directions for GDM research and applications, providing major insights into how they can continue to shape the future of network optimization.

cs.NI↗
arXiv2023

Effective Intrusion Detection in Highly Imbalanced IoT Networks with Lightweight S2CGAN-IDS

Since the advent of the Internet of Things (IoT), exchanging vast amounts of information has increased the number of security threats in networks. As a result, intrusion detection based on deep learning (DL) has been developed to achieve high throughput and high precision. Unlike general deep learning-based scenarios, IoT networks contain benign traffic far more than abnormal traffic, with some rare attacks. However, most existing studies have been focused on sacrificing the detection rate of the majority class in order to improve the detection rate of the minority class in class-imbalanced IoT networks. Although this way can reduce the false negative rate of minority classes, it both wastes resources and reduces the credibility of the intrusion detection systems. To address this issue, we propose a lightweight framework named S2CGAN-IDS. The proposed framework leverages the distribution characteristics of network traffic to expand the number of minority categories in both data space and feature space, resulting in a substantial increase in the detection rate of minority categories while simultaneously ensuring the detection precision of majority categories. To reduce the impact of sparsity on the experiments, the CICIDS2017 numeric dataset is utilized to demonstrate the effectiveness of the proposed method. The experimental results indicate that our proposed approach outperforms the superior method in both Precision and Recall, particularly with a 10.2% improvement in the F1-score.

cs.CR↗
arXiv2023

ESIA: An Efficient and Stable Identity Authentication for Internet of Vehicles

Decentralized, tamper-proof blockchain is regarded as a solution to a challenging authentication issue in the Internet of Vehicles (IoVs). However, the consensus time and communication overhead of blockchain increase significantly as the number of vehicles connected to the blockchain. To address this issue, vehicular fog computing has been introduced to improve efficiency. However, existing studies ignore several key factors such as the number of vehicles in the fog computing system, which can impact the consensus communication overhead. Meanwhile, there is no comprehensive study on the stability of vehicular fog composition. The vehicle movement will lead to dynamic changes in fog. If the composition of vehicular fog is unstable, the blockchain formed by this fog computing system will be unstable, which can affect the consensus efficiency. With the above considerations, we propose an efficient and stable identity authentication (ESIA) empowered by hierarchical blockchain and fog computing. By grouping vehicles efficiently, ESIA has low communication complexity and achieves high stability. Moreover, to enhance the consensus security of the hierarchical blockchain, the consensus process is from the bottom layer to the up layer (bottom-up), which we call B2UHChain. Through theoretical analysis and simulation verification, our scheme achieves the design goals of high efficiency and stability while significantly improving the IoV scalability to the power of 1.5 (^1.5) under similar security to a single-layer blockchain. In addition, ESIA has less communication and computation overhead, lower latency, and higher throughput than other baseline authentication schemes.

cs.CR↗
arXiv2023

Diffusion-based Reinforcement Learning for Edge-enabled AI-Generated Content Services

As Metaverse emerges as the next-generation Internet paradigm, the ability to efficiently generate content is paramount. AIGenerated Content (AIGC) emerges as a key solution, yet the resource intensive nature of large Generative AI (GAI) models presents challenges. To address this issue, we introduce an AIGC-as-a-Service (AaaS) architecture, which deploys AIGC models in wireless edge networks to ensure broad AIGC services accessibility for Metaverse users. Nonetheless, an important aspect of providing personalized user experiences requires carefully selecting AIGC Service Providers (ASPs) capable of effectively executing user tasks, which is complicated by environmental uncertainty and variability. Addressing this gap in current research, we introduce the AI-Generated Optimal Decision (AGOD) algorithm, a diffusion model-based approach for generating the optimal ASP selection decisions. Integrating AGOD with Deep Reinforcement Learning (DRL), we develop the Deep Diffusion Soft Actor-Critic (D2SAC) algorithm, enhancing the efficiency and effectiveness of ASP selection. Our comprehensive experiments demonstrate that D2SAC outperforms seven leading DRL algorithms. Furthermore, the proposed AGOD algorithm has the potential for extension to various optimization problems in wireless networks, positioning it as a promising approach for future research on AIGC-driven services. The implementation of our proposed method is available at: https://github.com/Lizonghang/AGOD.

cs.NI↗
arXiv2023

Enabling AI-Generated Content (AIGC) Services in Wireless Edge Networks

Artificial Intelligence-Generated Content (AIGC) refers to the use of AI to automate the information creation process while fulfilling the personalized requirements of users. However, due to the instability of AIGC models, e.g., the stochastic nature of diffusion models, the quality and accuracy of the generated content can vary significantly. In wireless edge networks, the transmission of incorrectly generated content may unnecessarily consume network resources. Thus, a dynamic AIGC service provider (ASP) selection scheme is required to enable users to connect to the most suited ASP, improving the users' satisfaction and quality of generated content. In this article, we first review the AIGC techniques and their applications in wireless networks. We then present the AIGC-as-a-service (AaaS) concept and discuss the challenges in deploying AaaS at the edge networks. Yet, it is essential to have performance metrics to evaluate the accuracy of AIGC services. Thus, we introduce several image-based perceived quality evaluation metrics. Then, we propose a general and effective model to illustrate the relationship between computational resources and user-perceived quality evaluation metrics. To achieve efficient AaaS and maximize the quality of generated content in wireless edge networks, we propose a deep reinforcement learning-enabled algorithm for optimal ASP selection. Simulation results show that the proposed algorithm can provide a higher quality of generated content to users and achieve fewer crashed tasks by comparing with four benchmarks, i.e., overloading-avoidance, random, round-robin policies, and the upper-bound schemes.

cs.AI↗
arXiv2022

HFedMS: Heterogeneous Federated Learning with Memorable Data Semantics in Industrial Metaverse

Federated Learning (FL), as a rapidly evolving privacy-preserving collaborative machine learning paradigm, is a promising approach to enable edge intelligence in the emerging Industrial Metaverse. Even though many successful use cases have proved the feasibility of FL in theory, in the industrial practice of Metaverse, the problems of non-independent and identically distributed (non-i.i.d.) data, learning forgetting caused by streaming industrial data, and scarce communication bandwidth remain key barriers to realize practical FL. Facing the above three challenges simultaneously, this paper presents a high-performance and efficient system named HFEDMS for incorporating practical FL into Industrial Metaverse. HFEDMS reduces data heterogeneity through dynamic grouping and training mode conversion (Dynamic Sequential-to-Parallel Training, STP). Then, it compensates for the forgotten knowledge by fusing compressed historical data semantics and calibrates classifier parameters (Semantic Compression and Compensation, SCC). Finally, the network parameters of the feature extractor and classifier are synchronized in different frequencies (Layer-wiseAlternative Synchronization Protocol, LASP) to reduce communication costs. These techniques make FL more adaptable to the heterogeneous streaming data continuously generated by industrial equipment, and are also more efficient in communication than traditional methods (e.g., Federated Averaging). Extensive experiments have been conducted on the streamed non-i.i.d. FEMNIST dataset using 368 simulated devices. Numerical results show that HFEDMS improves the classification accuracy by at least 6.4% compared with 8 benchmarks and saves both the overall runtime and transfer bytes by up to 98%, proving its superiority in precision and efficiency.

cs.LG↗
arXiv2022

Personalized Saliency in Task-Oriented Semantic Communications: Image Transmission and Performance Analysis

Semantic communication, as a promising technology, has emerged to break through the Shannon limit, which is envisioned as the key enabler and fundamental paradigm for future 6G networks and applications, e.g., smart healthcare. In this paper, we focus on UAV image-sensing-driven task-oriented semantic communications scenarios. The majority of existing work has focused on designing advanced algorithms for high-performance semantic communication. However, the challenges, such as energy-hungry and efficiency-limited image retrieval manner, and semantic encoding without considering user personality, have not been explored yet. These challenges have hindered the widespread adoption of semantic communication. To address the above challenges, at the semantic level, we first design an energy-efficient task-oriented semantic communication framework with a triple-based {\color{black}scene graph} for image information. We then design a new personalized semantic encoder based on user interests to meet the requirements of personalized saliency. Moreover, at the communication level, we study the effects of dynamic wireless fading channels on semantic transmission mathematically and thus design an optimal multi-user resource allocation scheme by using game theory. Numerical results based on real-world datasets clearly indicate that the proposed framework and schemes significantly enhance the personalization and anti-interference performance of semantic communication, and are also efficient to improve the communication quality of semantic communication services.

eess.IV↗
arXiv2022

Encoded Gradients Aggregation against Gradient Leakage in Federated Learning

Federated learning enables isolated clients to train a shared model collaboratively by aggregating the locally-computed gradient updates. However, privacy information could be leaked from uploaded gradients and be exposed to malicious attackers or an honest-but-curious server. Although the additive homomorphic encryption technique guarantees the security of this process, it brings unacceptable computation and communication burdens to FL participants. To mitigate this cost of secure aggregation and maintain the learning performance, we propose a new framework called Encoded Gradient Aggregation (\emph{EGA}). In detail, EGA first encodes local gradient updates into an encoded domain with injected noises in each client before the aggregation in the server. Then, the encoded gradients aggregation results can be recovered for the global model update via a decoding function. This scheme could prevent the raw gradients of a single client from exposing on the internet and keep them unknown to the server. EGA could provide optimization and communication benefits under different noise levels and defend against gradient leakage. We further provide a theoretical analysis of the approximation error and its impacts on federated optimization. Moreover, EGA is compatible with the most federated optimization algorithms. We conduct intensive experiments to evaluate EGA in real-world federated settings, and the results have demonstrated its efficacy.

cs.CR↗
arXiv2022

Cross-Silo Heterogeneous Model Federated Multitask Learning

Federated learning (FL) is a machine learning technique that enables participants to collaboratively train high-quality models without exchanging their private data. Participants utilizing cross-silo federated learning (CS-FL) settings are independent organizations with different task needs, and they are concerned not only with data privacy but also with independently training their unique models due to intellectual property considerations. Most existing FL methods are incapable of satisfying the above scenarios. In this study, we present a novel federated learning method CoFED based on unlabeled data pseudolabeling via a process known as cotraining. CoFED is a federated learning method that is compatible with heterogeneous models, tasks, and training processes. The experimental results suggest that the proposed method outperforms competing ones. This is especially true for non-independent and identically distributed settings and heterogeneous models, where the proposed method achieves a 35% performance improvement.

cs.LG↗