arXiv subjects

Wenli Zhou

Publications and source records attributed to Wenli Zhou.

At least 19 recordsLinked to original sources

arXiv2025

Huawei Cloud Model-as-a-Service on the CloudMatrix384 SuperPod

Scaled-out MoE LLMs and scaled-up SuperPods create new systems challenges for production Model-as-a-Service (MaaS), requiring disaggregation, low-latency communication, and decentralized serving. This report presents xDeepServe, the production serving system behind Huawei Cloud's MaaS offering on CloudMatrix384, a 48-server SuperPod with 384 Ascend 910C chips connected by a high-bandwidth UB fabric and global shared memory. It serves models including DeepSeek, Kimi, GLM, Qwen, and MiniMax, among others. xDeepServe is built around Transformerless, a disaggregated execution architecture that decomposes transformer inference into modular units -- attention, feedforward, and MoE -- and supports disaggregated Prefill-Decode and MoE-Attention deployments. To enable disaggregation, we develop XCCL, a memory-semantic communication layer providing microsecond-level point-to-point and scalable all-to-all primitives, and we extend FlowServe with decentralized DP groups and techniques to mitigate stragglers and synchronization variance. In a peak decoding configuration, xDeepServe reaches 2400 tokens/s per Ascend 910C chip at ~50ms time-per-output-token (TPOT).

cs.DC↗

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

Scalable Reinforcement Learning for Virtual Machine Scheduling

Recent advancements in reinforcement learning (RL) have shown promise for optimizing virtual machine scheduling (VMS) in small-scale clusters. The utilization of RL to large-scale cloud computing scenarios remains notably constrained. This paper introduces a scalable RL framework, called Cluster Value Decomposition Reinforcement Learning (CVD-RL), to surmount the scalability hurdles inherent in large-scale VMS. The CVD-RL framework innovatively combines a decomposition operator with a look-ahead operator to adeptly manage representation complexities, while complemented by a Top-$k$ filter operator that refines exploration efficiency. Different from existing approaches limited to clusters of $10$ or fewer physical machines (PMs), CVD-RL extends its applicability to environments encompassing up to $50$ PMs. Furthermore, the CVD-RL framework demonstrates generalization capabilities that surpass contemporary SOTA methodologies across a variety of scenarios in empirical studies. This breakthrough not only showcases the framework's exceptional scalability and performance but also represents a significant leap in the application of RL for VMS within complex, large-scale cloud infrastructures. The code is available at https://anonymous.4open.science/r/marl4sche-D0FE.

cs.LG↗

arXiv2024

Linear stability analysis of the Couette flow for the 2D Euler-Poisson system

This paper is concerned with the linear stability analysis for the Couette flow of the Euler-Poisson system for both ionic fluid and electronic fluid in the domain $\bb{T}\times\bb{R}$. We establish the upper and lower bounds of the linearized solutions of the Euler-Poisson system near Couette flow. In particular, the inviscid damping for the solenoidal component of the velocity is obtained.

math.AP↗

arXiv2023

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision

It is very challenging for various visual tasks such as image fusion, pedestrian detection and image-to-image translation in low light conditions due to the loss of effective target areas. In this case, infrared and visible images can be used together to provide both rich detail information and effective target areas. In this paper, we present LLVIP, a visible-infrared paired dataset for low-light vision. This dataset contains 30976 images, or 15488 pairs, most of which were taken at very dark scenes, and all of the images are strictly aligned in time and space. Pedestrians in the dataset are labeled. We compare the dataset with other visible-infrared datasets and evaluate the performance of some popular visual algorithms including image fusion, pedestrian detection and image-to-image translation on the dataset. The experimental results demonstrate the complementary effect of fusion on image information, and find the deficiency of existing algorithms of the three visual tasks in very low-light conditions. We believe the LLVIP dataset will contribute to the community of computer vision by promoting image fusion, pedestrian detection and image-to-image translation in very low-light applications. The dataset is being released in https://bupt-ai-cz.github.io/LLVIP. Raw data is also provided for further research such as image registration.

cs.CV↗

arXiv2022

ReAssigner: A Plug-and-Play Virtual Machine Scheduling Intensifier for Heterogeneous Requests

With the rapid development of cloud computing, virtual machine scheduling has become one of the most important but challenging issues for the cloud computing community, especially for practical heterogeneous request sequences. By analyzing the impact of request heterogeneity on some popular heuristic schedulers, it can be found that existing scheduling algorithms can not handle the request heterogeneity properly and efficiently. In this paper, a plug-and-play virtual machine scheduling intensifier, called Resource Assigner (ReAssigner), is proposed to enhance the scheduling efficiency of any given scheduler for heterogeneous requests. The key idea of ReAssigner is to pre-assign roles to physical resources and let resources of the same role form a virtual cluster to handle homogeneous requests. ReAssigner can cooperate with arbitrary schedulers by restricting their scheduling space to virtual clusters. With evaluations on the real dataset from Huawei Cloud, the proposed ReAssigner achieves significant scheduling performance improvement compared with some state-of-the-art scheduling methods.

cs.DC↗

arXiv2022

Rigorous derivation of the full primitive equations by the scaled Boussinesq equations with rotation

The primitive equations of large-scale oceanic dynamics form a fundamental model in geophysical flows. It is well-known that the primitive equations can be formally derived by the hydrostatic approximation. On the other hand, the mathematically rigorous derivation of the primitive equations without coupling with the temperature is also known. In this paper, we generalize the above result from the mathematical point of view. More precisely, we prove that the scaled Boussinesq equations with rotation converge to the full primitive equations in a strong sense, globally in time, with the convergence rate $O(\varepsilon)$, as the aspect ratio $\varepsilon$ goes to zero.

math.AP↗

arXiv2022

Gleo-Det: Deep Convolution Feature-Guided Detector with Local Entropy Optimization for Salient Points

Feature detection is an important procedure for image matching, where unsupervised feature detection methods are the detection approaches that have been mostly studied recently, including the ones that are based on repeatability requirement to define loss functions, and the ones that attempt to use descriptor matching to drive the optimization of the pipelines. For the former type, mean square error (MSE) is usually used which cannot provide strong constraint for training and can make the model easy to be stuck into the collapsed solution. For the later one, due to the down sampling operation and the expansion of receptive fields, the details can be lost for local descriptors can be lost, making the constraint not fine enough. Considering the issues above, we propose to combine both ideas, which including three aspects. 1) We propose to achieve fine constraint based on the requirement of repeatability while coarse constraint with guidance of deep convolution features. 2) To address the issue that optimization with MSE is limited, entropy-based cost function is utilized, both soft cross-entropy and self-information. 3) With the guidance of convolution features, we define the cost function from both positive and negative sides. Finally, we study the effect of each modification proposed and experiments demonstrate that our method achieves competitive results over the state-of-the-art approaches.

cs.CV↗

arXiv2022

The hydrostatic approximation of the Boussinesq equations with rotation in a thin domain

In this paper, we improve the global existence result in [9] slightly. More precisely, the global existence of strong solutions to the primitive equations with only horizontal viscosity and diffusivity is obtained under the assumption of initial data $(v_0,T_0) \in H^1$ with $\partial_z v_0 \in L^4$. Moreover, we prove that the scaled Boussinesq equations with rotation strongly converge to the primitive equations with only horizontal viscosity and diffusivity, in the cases of $H^1$ initial data, $H^1$ initial data with additional regularity $\partial_z v_0 \in L^4$ and $H^2$ initial data, respectively, as the aspect ration parameter $λ$ goes to zero, and the rate of convergence is of the order $O(λ^{η/2})$ with $η=\min\{2,β-2,γ-2\}(2<β,γ<\infty)$. The convergence result implies a rigorous justification of the hydrostatic approximation.

math.AP↗

arXiv2022

On the rigorous mathematical derivation for the viscous primitive equations with density stratification

In this paper, we rigorously derive the governed equations describing the motion of stable stratified fluid, from the mathematical point of view. Specially, we prove that the scaled Boussinesq equations strongly converge to the viscous primitive equations with density stratification as the aspect ration parameter goes to zero, and the rate of convergence is of the same order as the aspect ratio parameter. Moreover, in order to obtain this convergence result, we also establish the global well-posedness of strong solutions to the viscous primitive equations with density stratification.

math.AP↗

arXiv2021

VMAgent: Scheduling Simulator for Reinforcement Learning

A novel simulator called VMAgent is introduced to help RL researchers better explore new methods, especially for virtual machine scheduling. VMAgent is inspired by practical virtual machine (VM) scheduling tasks and provides an efficient simulation platform that can reflect the real situations of cloud computing. Three scenarios (fading, recovering, and expansion) are concluded from practical cloud computing and corresponds to many reinforcement learning challenges (high dimensional state and action spaces, high non-stationarity, and life-long demand). VMAgent provides flexible configurations for RL researchers to design their customized scheduling environments considering different problem features. From the VM scheduling perspective, VMAgent also helps to explore better learning-based scheduling solutions.

cs.LG↗

arXiv2021

Meta Self-Learning for Multi-Source Domain Adaptation: A Benchmark

In recent years, deep learning-based methods have shown promising results in computer vision area. However, a common deep learning model requires a large amount of labeled data, which is labor-intensive to collect and label. What's more, the model can be ruined due to the domain shift between training data and testing data. Text recognition is a broadly studied field in computer vision and suffers from the same problems noted above due to the diversity of fonts and complicated backgrounds. In this paper, we focus on the text recognition problem and mainly make three contributions toward these problems. First, we collect a multi-source domain adaptation dataset for text recognition, including five different domains with over five million images, which is the first multi-domain text recognition dataset to our best knowledge. Secondly, we propose a new method called Meta Self-Learning, which combines the self-learning method with the meta-learning paradigm and achieves a better recognition result under the scene of multi-domain adaptation. Thirdly, extensive experiments are conducted on the dataset to provide a benchmark and also show the effectiveness of our method. The code of our work and dataset are available soon at https://bupt-ai-cz.github.io/Meta-SelfLearning/.

cs.CV↗

arXiv2016

Learning-Based Coexistence in Two-Tier Heterogeneous Networks with Cognitive Small Cells

We study the coexistence problem in a two-tier heterogeneous network (HetNet) with cognitive small cells. In particular, we consider an underlay HetNet, where the cognitive small base station (C-SBS) is allowed to use the frequency bands of the macro cell with an access probability (AP) as long as the C-SBS satisfies a preset interference probability (IP) constraint at macro users (MUs). To enhance the AP (or transmission opportunity) of the C-SBS, we propose a learning-based algorithm for the C-SBS and exploit the distance information between the macro base station (MBS) and MUs. Generally, the signal from the MBS to a specific MU contains the distance information between the MBS to the MU. We enable the C-SBS to analyze the MBS signal on a target frequency band, and learn the distance information between the MBS and the corresponding MU. With the learnt distance information, we calculate the upper bound of the probability that the C-SBS may interfere with the MU, and design an AP with a closed-form expression under the IP constraint. Numerical results indicate that the proposed algorithm outperforms the existing methods up to $60\%$ AP (or transmission opportunity).

cs.IT↗

arXiv2016

Primary Channel Gain Estimation for Spectrum Sharing in Cognitive Radio Networks

In cognitive radio networks, the channel gain between primary transceivers, namely, primary channel gain, is crucial for a cognitive transmitter (CT) to control the transmit power and achieve spectrum sharing. Conventionally, the primary channel gain is estimated in the primary system and thus unavailable at the CT. To deal with this issue, two estimators are proposed by enabling the CT to sense primary signals. In particular, by adopting the maximum likelihood (ML) criterion to analyze the received primary signals, a ML estimator is first developed. After demonstrating the high computational complexity of the ML estimator, a median based (MB) estimator with proved low complexity is then proposed. Furthermore, the estimation accuracy of the MB estimation is theoretically characterized. By comparing the ML estimator and the MB estimator from the aspects of the computational complexity as well as the estimation accuracy, both advantages and disadvantages of two estimators are revealed. Numerical results show that the estimation errors of the ML estimator and the MB estimator can be as small as $0.6$ dB and $0.7$ dB, respectively.

cs.IT↗

arXiv2016

Estimating the Distance Between Macro Base Station and Users in Heterogeneous Networks

In underlay heterogeneous networks (HetNets), the distance between a macro base station (MBS) and a macro user (MU) is crucial for a small-cell based station (SBS) to control the interference to the MU and achieve the coexistence. To obtain the distance between the MBS and the MU, the SBS needs a backhaul link from the macro system, such that the macro system is able to transmit the information of the distance to the SBS through the backhaul link. However, there may not exist any backhaul link from the macro system to the SBS in practical situations. Thus, it is challenging for the SBS to obtain the distance. To deal with this issue, we propose a median based (MB) estimator for the SBS to obtain the distance between the MBS and the MU without any backhaul link. Numerical results show that the estimation error of the MB estimator can be as small as $4\%$.

cs.IT↗

arXiv2009

Note on the 4- and 5-leaf powers

Motivated by the problem of reconstructing evolutionary history, Nishimura et al. defined $k$-leaf powers as the class of graphs $G=(V,E)$ which has a $k$-leaf root $T$, i.e., $T$ is a tree such that the vertices of $G$ are exactly the leaves of $T$ and two vertices in $V$ are adjacent in $G$ if and only if their distance in $T$ is at most $k$. It is known that leaf powers are chordal graphs. Brandstädt and Le proved that every $k$-leaf power is a $(k+2)$-leaf power and every 3-leaf power is a $k$-leaf power for $k\geq 3$. They asked whether a $k$-leaf power is also a $(k+1)$-leaf power for any $k\geq 4$. Fellows et al. gave an example of a 4-leaf power which is not a 5-leaf power. It is interesting to find all the graphs which have both 4-leaf roots and 5-leaf roots. In this paper, we prove that, if $G$ is a 4-leaf power with $L(G)\neq \emptyset$, then $G$ is also a 5-leaf power, where $L(G)$ denotes the set of leaves of $G$.

math.CO↗

arXiv2009

Sharp bounds for the generalized connectivity $κ_3(G)$

Let $G$ be a nontrivial connected graph of order $n$ and let $k$ be an integer with $2\leq k\leq n$. For a set $S$ of $k$ vertices of $G$, let $κ(S)$ denote the maximum number $\ell$ of edge-disjoint trees $T_1,T_2,...,T_\ell$ in $G$ such that $V(T_i)\cap V(T_j)=S$ for every pair $i,j$ of distinct integers with $1\leq i,j\leq \ell$. A collection $\{T_1,T_2,...,T_\ell\}$ of trees in $G$ with this property is called an internally disjoint set of trees connecting $S$. Chartrand et al. generalized the concept of connectivity as follows: The $k$-$connectivity$, denoted by $κ_k(G)$, of $G$ is defined by $κ_k(G)=$min$\{κ(S)\}$, where the minimum is taken over all $k$-subsets $S$ of $V(G)$. Thus $κ_2(G)=κ(G)$, where $κ(G)$ is the connectivity of $G$. In general, the investigation of $κ_k(G)$ is very difficult. We therefore focus on the investigation on $κ_3(G)$ in this paper. We study the relation between the connectivity and the 3-connectivity of a graph. First we give sharp upper and lower bounds of $κ_3(G)$ for general graphs $G$, and construct two kinds of graphs which attain the upper and lower bound, respectively. We then show that if $G$ is a connected planar graph, then $κ(G)-1 \leq κ_3(G)\leq κ(G)$, and give some classes of graphs which attain the bounds. In the end we show that the problem whether $κ(G)=κ_3(G)$ for a planar graph $G$ can be solved in polynomial time.

math.CO↗

arXiv2007

Complexity of the conditional colorability of graphs

For an integer $r>0$, a conditional $(k,r)$-coloring of a graph $G$ is a proper $k$-coloring of the vertices of $G$ such that every vertex $v$ of degree $d(v)$ in $G$ is adjacent to vertices with at least $min\{r, d(v)\}$ different colors. The smallest integer $k$ for which a graph $G$ has a conditional $(k,r)$-coloring is called the $r$th order conditional chromatic number, denoted by $χ_r(G)$. It is easy to see that the conditional coloring is a generalization of the traditional vertex coloring for which $r=1$. In this paper, we consider the complexity of the conditional colorings of graphs. The main result is that the conditional $(3,2)$-colorability is $NP$-complete for triangle-free graphs with maximum degree at most 3, which is different from the old result that the traditional 3-colorability is polynomial solvable for graphs with maximum degree at most 3. This also implies that it is $NP$-complete to determine if a graph of maximum degree 3 is $(3,2)$- or $(4,2)$-colorable. Also we have proved that some old complexity results for traditional colorings still hold for the conditional colorings.

cs.DM↗