arXiv subjects

Zhibin Yu

Publications and source records attributed to Zhibin Yu.

At least 19 recordsLinked to original sources

arXiv2026

ASAP: A Disaggregated and Asynchronous Inference System for MoE Prefill

Mixture-of-Experts (MoE) models have become the de facto standard for scaling large language models. To maintain computational efficiency, modern MoE serving systems typically employ a hybrid parallelism strategy, combining Data Parallelism (DP) for attention stages with Expert Parallelism (EP) for MoE stages. However, this design necessitates frequent global synchronization barriers between attention DP groups and experts. In online serving, significant variance in request arrival rates and sequence lengths inherently leads to DP imbalance, causing severe synchronization stalls that degrade Time-to-First-Token (TTFT) and system throughput. We present ASAP, an asynchronous inference system specifically designed to accelerate the prefill phase of MoE models. ASAP disaggregates the attention and MoE stages and implements a fully asynchronous execution pipeline. This is achieved through a suite of specialized asynchronous communication primitives and four coordinated optimizations across request scheduling and model execution, which collectively dismantle global synchronization barriers. We implement and evaluate ASAP on CloudMatrix384 super-nodes, demonstrating that it improves SLO-compliant prefill throughput by 90% compared to state-of-the-art synchronous serving solutions.

cs.DC↗

arXiv2026

S-HPLB: Efficient LLM Attention Serving via Sparsity-Aware Head Parallelism Load Balance

With the increasing volumes of Large Language Models (LLMs) and the expanding context lengths, attention computation has become a key performance bottleneck in LLM serving. For fast attention computation, recent practices often parallelize the attention heads on multiple GPUs, and also widely adopt attention sparsification to reduce the computation amount -- which selectively computes a subset of attention pairs under a preset sparsity budget. In this paper, we notice that attention heads of an LLM model often exhibit heterogeneous-yet-stable sparsity elasticities, which motivates us to enforce head-adaptive sparsity budgets to attain better efficiency while preserving high inference quality. Yet, from the system aspect, with heterogeneous sparsity levels, attention computation time on different heads would be inconsistent, yielding cross-GPU resource bubbles under head-parallel deployment. To further minimize such bubbles, we propose a novel attention deployment strategy called Sparsity-aware Head-Parallel Load Balance (S-HPLB). Experiments on long-context benchmark show that, S-HPLB can achieve a $2.88\times$ improvement in average attention computation latency without quality degradation.

cs.DC↗

arXiv2026

Meta-PINNs: Meta-Learning Enhanced Physics-Informed Machine Learning Framework for Turbomachinery Flow Predictions under Varying Operation Conditions

Coupling physics with machine learning models has shown great potential for solving fluid dynamics problems governed by partial differential equations. However, conventional methods, such as physics-informed neural networks, often suffer from slow convergence, unstable training, and limited generalization across different flow conditions. To overcome these challenges, this study proposes a novel meta-learning en- hanced physics-informed neural networks (Meta-PINNs) framework, which integrates a meta-optimization strategy into the training process. The approach allows the model to automatically adapt its learning process to varying physical regimes, thereby substantially improving both training efficiency and predictive robustness. The proposed Meta-PINNs model is evaluated on two representative flow problems: (1) unsteady flow around a circular cylinder at multiple inlet Reynolds numbers, and (2) steady turbulent flow within a compressor cascade passage at various angles of attack. In both cases, the extrapolation performance of the developed framework is comprehensively tested by predicting the flow fields at Reynolds numbers and angles of attack that are not included in the training set. The results demonstrate that Meta-PINNs achieve a 1-2 order-of-magnitude improvement in accuracy over vanilla physics-informed neural networks and standard neural networks, while reducing computational cost by up to 95.7 % and 92.1 %, respectively. It successfully captures the sequential patterns of key flow features such as pressure and velocity distributions under unseen conditions. Thus, the findings confirm that the Meta-PINNs framework offers a notable improvement in convergence and generalization over existing machine learning approaches, providing a promising pathway toward smart simulations of complex turbomachinery flows.

physics.flu-dyn↗

arXiv2026

Device-Centric ISAC for Exposure Control via Opportunistic Virtual Aperture Sensing

Regulatory limits on Maximum Permissible Exposure (MPE) require handheld devices to reduce transmit power when operated near the user's body. Current proximity sensors provide only binary detection, triggering conservative power back-off that degrades link quality. If the device could measure its distance from the body, transmit power could be adjusted proportionally, improving throughput while maintaining compliance. This paper develops a device-centric integrated sensing and communication (ISAC) method for the device to measure this distance. The uplink communication waveform is exploited for sensing, and the natural motion of the user's hand creates a virtual aperture that provides the angular resolution necessary for localization. Virtual aperture processing requires precise knowledge of the device trajectory, which in this scenario is opportunistic and unknown. One can exploit onboard inertial sensors to estimate the device trajectory; however, the inertial sensors accuracy is not sufficient. To address this, we develop an autofocus algorithm based on extended Kalman filtering that jointly tracks the trajectory and compensates residual errors using phase observations from strong scatterers. The Bayesian Cram\'er-Rao bound for localization is derived under correlated inertial errors. Numerical results at 28GHz demonstrate centimeter-level accuracy with realistic sensor parameters.

eess.SP↗

arXiv2025

Harli: SLO-Aware Co-location of LLM Inference and PEFT-based Finetuning on Model-as-a-Service Platforms

Large language models (LLMs) are increasingly deployed under the Model-as-a-Service (MaaS) paradigm. To meet stringent quality-of-service (QoS) requirements, existing LLM serving systems disaggregate the prefill and decode phases of inference. However, decode instances often experience low GPU utilization due to their memory-bound nature and insufficient batching in dynamic workloads, leaving compute resources underutilized. We introduce Harli, a serving system that improves GPU utilization by co-locating parameter-efficient finetuning (PEFT) tasks with LLM decode instances. PEFT tasks are compute-bound and memory-efficient, making them ideal candidates for safe co-location. Specifically, Harli addresses key challenges--limited memory and unpredictable interference--using three components: a unified memory allocator for runtime memory reuse, a two-stage latency predictor for decode latency modeling, and a QoS-guaranteed throughput-maximizing scheduler for throughput maximization. Experimental results show that Harli improves the finetune throughput by 46.2% on average (up to 92.0%) over state-of-the-art serving systems, while maintaining strict QoS guarantees for inference decode.

cs.DC↗

arXiv2025

Indoor Positioning Based on Active Radar Sensing and Passive Reflectors: Reflector Placement Optimization

We extend our work on a novel indoor positioning system (IPS) for autonomous mobile robots (AMRs) based on radar sensing of local, passive radar reflectors. Through the combination of simple reflectors and a single-channel frequency modulated continuous wave (FMCW) radar, high positioning accuracy at low system cost can be achieved. Further, a multi-objective (MO) particle swarm optimization (PSO) algorithm is presented that optimizes the 2D placement of radar reflectors in complex room settings.

cs.RO↗

arXiv2025

Serving Large Language Models on Huawei CloudMatrix384

The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.

cs.DC↗

arXiv2025

Discrete Codebook Design for Self-interference Suppression in mmWave ISAC

This paper presents discrete codebook synthesis methods for self-interference (SI) suppression in a mmWave device, designed to support FD ISAC. We formulate a SINR maximization problem that optimizes the RX and TX codewords, aimed at suppressing the near-field SI signal while maintaining the beamforming gain in the far-field sensing directions. The formulation considers the practical constraints of discrete RX and TX codebooks with quantized phase settings, as well as a TX beamforming gain requirement in the specified communication direction. Under an alternating optimization framework, the RX and TX codewords are iteratively optimized, with one fixed while the other is optimized. When the TX codeword is fixed, we show that the RX codeword optimization problem can be formulated as an integer quadratic fractional programming (IQFP) problem. Using Dinkelbach's algorithm, we transform the problem into a sequence of subproblems in which the numerator and the denominator of the objective function are decoupled. These subproblems, subject to discrete constraints, are then efficiently solved by the spherical search (SS) method. This overall approach is referred to as FP-SS. When the RX codeword is fixed, the TX codeword optimization problem can similarly be formulated as an IQFP problem, whereas an additional TX beamforming constraint for communication needs to be considered. The problem is solved through Dinkelbach's transformation followed by the constrained spherical search (CSS), and we refer to this approach as FP-CSS. Finally, we integrate the FP-SS and FP-CSS methods into a joint RX-TX codebook design approach. Simulations show that, the proposed FP-SS and FP-CSS achieve the same SI suppression performance as the corresponding exhaustive search method, but with much lower complexity. Furthermore, the alternating optimization framework achieved even better SI suppression performance.

eess.SP↗

arXiv2025

Machine Learning-based Regional Cooling Demand Prediction with Optimised Dataset Partitioning

In the context of global warming, even relatively cooler countries like the UK are experiencing a rise in cooling demand, particularly in southern regions such as London. This growing demand, especially during the summer months, presents significant challenges for energy management systems. Accurately predicting cooling demand in urban domestic buildings is essential for maintaining energy efficiency. This study introduces a generalised framework for developing high-resolution Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks using physical model-based summer cooling demand data. To maximise the predictive capability and generalisation ability of the models under limited data scenarios, four distinct data partitioning strategies were implemented, including the extrapolation, month-based interpolation, global interpolation, and day-based interpolation. Bayesian Optimisation (BO) was then applied to fine-tune the hyper-parameters, substantially improving the framework predictive accuracy. Results show that the day-based interpolation GRU model demonstrated the best performance due to its ability to retain both the data randomness and the time sequence continuity characteristics. This optimal model achieves a Root Mean Squared Error (RMSE) of 2.22%, a Mean Absolute Error (MAE) of 0.87%, and a coefficient of determination (R square) of 0.9386 on the test set. The generalisation ability of this framework was further evaluated by forecasting.

physics.soc-ph↗

arXiv2024

UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement

Learning-based underwater image enhancement (UIE) methods have made great progress. However, the lack of large-scale and high-quality paired training samples has become the main bottleneck hindering the development of UIE. The inter-frame information in underwater videos can accelerate or optimize the UIE process. Thus, we constructed the first large-scale high-resolution underwater video enhancement benchmark (UVEB) to promote the development of underwater vision.It contains 1,308 pairs of video sequences and more than 453,000 high-resolution with 38\% Ultra-High-Definition (UHD) 4K frame pairs. UVEB comes from multiple countries, containing various scenes and video degradation types to adapt to diverse and complex underwater environments. We also propose the first supervised underwater video enhancement method, UVE-Net. UVE-Net converts the current frame information into convolutional kernels and passes them to adjacent frames for efficient inter-frame information exchange. By fully utilizing the redundant degraded information of underwater videos, UVE-Net completes video enhancement better. Experiments show the effective network design and good performance of UVE-Net.

cs.CV↗

arXiv2023

Indoor Positioning based on Active Radar Sensing and Passive Reflectors: Concepts & Initial Results

To navigate reliably in indoor environments, an industrial autonomous vehicle must know its position. However, current indoor vehicle positioning technologies either lack accuracy, usability or are too expensive. Thus, we propose a novel concept called local reference point assisted active radar positioning, which is able to overcome these drawbacks. It is based on distributing passive retroreflectors in the indoor environment such that each position of the vehicle can be identified by a unique reflection characteristic regarding the reflectors. To observe these characteristics, the autonomous vehicle is equipped with an active radar system. On one hand, this paper presents the basic idea and concept of our new approach towards indoor vehicle positioning and especially focuses on the crucial placement of the reflectors. On the other hand, it also provides a proof of concept by conducting a full system simulation including the placement of the local reference points, the radar-based distance estimation and the comparison of two different positioning methods. It successfully demonstrates the feasibility of our proposed approach.

cs.RO↗

arXiv2022

Channel Pruned YOLOv5-based Deep Learning Approach for Rapid and Accurate Outdoor Obstacles Detection

One-stage algorithm have been widely used in target detection systems that need to be trained with massive data. Most of them perform well both in real-time and accuracy. However, due to their convolutional structure, they need more computing power and greater memory consumption. Hence, we applied pruning strategy to target detection networks to reduce the number of parameters and the size of model. To demonstrate the practicality of the pruning method, we select the YOLOv5 model for experiments and provide a data set of outdoor obstacles to show the effect of model. In this specific data set, in the best circumstances, the volume of the network model is reduced by 49.7% compared with the original model, and the reasoning time is reduced by 52.5%. Meanwhile, it also uses data processing methods to compensate for the drop in accuracy caused by pruning.

cs.CV↗

arXiv2022

LOCAT: Low-Overhead Online Configuration Auto-Tuning of Spark SQL Applications

Spark SQL has been widely deployed in industry but it is challenging to tune its performance. Recent studies try to employ machine learning (ML) to solve this problem, but suffer from two drawbacks. First, it takes a long time (high overhead) to collect training samples. Second, the optimal configuration for one input data size of the same application might not be optimal for others. To address these issues, we propose a novel Bayesian Optimization (BO) based approach named LOCAT to automatically tune the configurations of Spark SQL applications online. LOCAT innovates three techniques. The first technique, named QCSA, eliminates the configuration-insensitive queries by Query Configuration Sensitivity Analysis (QCSA) when collecting training samples. The second technique, dubbed DAGP, is a Datasize-Aware Gaussian Process (DAGP) which models the performance of an application as a distribution of functions of configuration parameters as well as input data size. The third technique, called IICP, Identifies Important Configuration Parameters (IICP) with respect to performance and only tunes the important ones. As such, LOCAT can tune the configurations of a Spark SQL application with low overhead and adapt to different input data sizes. We employ Spark SQL applications from benchmark suites TPC-DS, TPC-H, and HiBench running on two significantly different clusters, a four-node ARM cluster and an eight-node x86 cluster, to evaluate LOCAT. The experimental results on the ARM cluster show that LOCAT accelerates the optimization procedures of the state-of-the-art approaches by at least 4.1x and up to 9.7x; moreover, LOCAT improves the application performance by at least 1.9x and up to 2.4x. On the x86 cluster, LOCAT shows similar results to those on the ARM cluster.

cs.DC↗

arXiv2021

Detecting Mitosis against Domain Shift using a Fused Detector and Deep Ensemble Classification Model for MIDOG Challenge

Mitotic figure count is an important marker of tumor proliferation and has been shown to be associated with patients' prognosis. Deep learning based mitotic figure detection methods have been utilized to automatically locate the cell in mitosis using hematoxylin \& eosin (H\&E) stained images. However, the model performance deteriorates due to the large variation of color tone and intensity in H\&E images. In this work, we proposed a two stage mitotic figure detection framework by fusing a detector and a deep ensemble classification model. To alleviate the impact of color variation in H\&E images, we utilize both stain normalization and data augmentation, aiding model to learn color irrelevant features. The proposed model obtains an F1 score of 0.7550 on the preliminary testing set released by the MIDOG challenge.

cs.CV↗

arXiv2019

EncryptGAN: Image Steganography with Domain Transform

We propose an image steganographic algorithm called EncryptGAN, which disguises private image communication in an open communication channel. The insight is that content transform between two very different domains (e.g., face to flower) allows one to hide image messages in one domain (face) and communicate using its counterpart in another domain (flower). The key ingredient in our method, unlike related approaches, is a specially trained network to extract transformed images from both domains and use them as the public and private keys. We ensure the image communication remain secret except for the intended recipient even when the content transformation networks are exposed. To communicate, one directly pastes the `message' image onto a larger public key image (face). Depending on the location and content of the message image, the `disguise' image (flower) alters its appearance and shape while maintaining its overall objectiveness (flower). The recipient decodes the alternated image to uncover the original image message using its message image key. We implement the entire procedure as a constrained Cycle-GAN, where the public and the private key generating network is used as an additional constraint to the cycle consistency. Comprehensive experimental results show our EncryptGAN outperforms the state-of-arts in terms of both encryption and security measures.

cs.MM↗

arXiv2019

ReshapeGAN: Object Reshaping by Providing A Single Reference Image

The aim of this work is learning to reshape the object in an input image to an arbitrary new shape, by just simply providing a single reference image with an object instance in the desired shape. We propose a new Generative Adversarial Network (GAN) architecture for such an object reshaping problem, named ReshapeGAN. The network can be tailored for handling all kinds of problem settings, including both within-domain (or single-dataset) reshaping and cross-domain (typically across mutiple datasets) reshaping, with paired or unpaired training data. The appearance of the input object is preserved in all cases, and thus it is still identifiable after reshaping, which has never been achieved as far as we are aware. We present the tailored models of the proposed ReshapeGAN for all the problem settings, and have them tested on 8 kinds of reshaping tasks with 13 different datasets, demonstrating the ability of ReshapeGAN on generating convincing and superior results for object reshaping. To the best of our knowledge, we are the first to be able to make one GAN framework work on all such object reshaping tasks, especially the cross-domain tasks on handling multiple diverse datasets. We present here both ablation studies on our proposed ReshapeGAN models and comparisons with the state-of-the-art models when they are made comparable, using all kinds of applicable metrics that we are aware of.

cs.CV↗

arXiv2019

One-Shot Image-to-Image Translation via Part-Global Learning with a Multi-adversarial Framework

It is well known that humans can learn and recognize objects effectively from several limited image samples. However, learning from just a few images is still a tremendous challenge for existing main-stream deep neural networks. Inspired by analogical reasoning in the human mind, a feasible strategy is to translate the abundant images of a rich source domain to enrich the relevant yet different target domain with insufficient image data. To achieve this goal, we propose a novel, effective multi-adversarial framework (MA) based on part-global learning, which accomplishes one-shot cross-domain image-to-image translation. In specific, we first devise a part-global adversarial training scheme to provide an efficient way for feature extraction and prevent discriminators being over-fitted. Then, a multi-adversarial mechanism is employed to enhance the image-to-image translation ability to unearth the high-level semantic representation. Moreover, a balanced adversarial loss function is presented, which aims to balance the training data and stabilize the training process. Extensive experiments demonstrate that the proposed approach can obtain impressive results on various datasets between two extremely imbalanced image domains and outperform state-of-the-art methods on one-shot image-to-image translation.

cs.CV↗

arXiv2019

Generative Adversarial Network with Multi-Branch Discriminator for Cross-Species Image-to-Image Translation

Current approaches have made great progress on image-to-image translation tasks benefiting from the success of image synthesis methods especially generative adversarial networks (GANs). However, existing methods are limited to handling translation tasks between two species while keeping the content matching on the semantic level. A more challenging task would be the translation among more than two species. To explore this new area, we propose a simple yet effective structure of a multi-branch discriminator for enhancing an arbitrary generative adversarial architecture (GAN), named GAN-MBD. It takes advantage of the boosting strategy to break a common discriminator into several smaller ones with fewer parameters, which can enhance the generation and synthesis abilities of GANs efficiently and effectively. Comprehensive experiments show that the proposed multi-branch discriminator can dramatically improve the performance of popular GANs on cross-species image-to-image translation tasks while reducing the number of parameters for computation. The code and some datasets are attached as supplementary materials for reference.

cs.CV↗