arXiv subjects

Xiaowen Chu

Publications and source records attributed to Xiaowen Chu.

At least 73 records · Page 4Linked to original sources

arXiv2025

FlowKV: Enhancing Multi-Turn Conversational Coherence in LLMs via Isolated Key-Value Cache Management

Large Language Models (LLMs) are increasingly deployed in multi-turn conversational applications, where the management of the Key-Value (KV) Cache presents a significant bottleneck. The linear growth of the KV Cache with dialogue history imposes substantial computational costs, and existing eviction strategies often degrade performance by repeatedly compressing early conversational context, leading to information loss and context forgetting. This paper introduces FlowKV, a novel \textbf{multi-turn isolation mechanism} for KV Cache management, which can be applied to any KV Cache compression method without training. FlowKV's core innovation is a multi-turn isolation mechanism that preserves the accumulated compressed KV cache from past turns. Compression is then strategically applied only to the newly generated KV pairs of the latest completed turn, effectively preventing the re-compression of older context and thereby mitigating catastrophic forgetting. Our results demonstrate that FlowKV consistently and significantly outperforms baseline strategies in maintaining instruction-following accuracy and user preference retention from 10.90\% to 75.40\%, particularly in later conversational turns.

cs.CL↗

arXiv2025

ElasWave: An Elastic-Native System for Scalable Hybrid-Parallel Training

Large-scale LLM pretraining now runs across $10^5$--$10^6$ accelerators, making failures routine and elasticity mandatory. We posit that an elastic-native training system must jointly deliver (i) parameter consistency, (ii) low mean time to recovery (MTTR), (iii) high post-change throughput, and (iv) computation consistency. No prior system achieves all four simultaneously. To achieve these goals, we present ElasWave, which delivers per-step fault tolerance via multi-dimensional scheduling across graph, dataflow, DVFS, and RNG. ElasWave reshapes and reshards micro-batches while preserving the global batch size and gradient scale. It performs online pipeline resharding with asynchronous parameter migration and interleaves ZeRO partitions, reducing parameter recovery processes to disjoint rank-to-rank transfers. It further leverages DVFS to absorb pipeline bubbles and reshards RNG to keep computation consistency. Together, a dynamic communicator enables in-place communication group edits, while per-step in-memory snapshots support online verification and redistribution. We evaluate ElasWave on 96 NPUs and benchmark it against state-of-the-art baselines: throughput improves by $1.35\times$ over ReCycle and $1.60\times$ over TorchFT; communicator recovery completes within one second (up to $82\times/3.6\times$ faster than full/partial rebuilds); migration MTTR drops by as much as $51\%$; and convergence deviation is reduced by approximately $78\%$.

cs.DC↗

arXiv2025

SGMAGNet: A Baseline Model for 3D Cloud Phase Structure Reconstruction on a New Passive Active Satellite Benchmark

Cloud phase profiles are critical for numerical weather prediction (NWP), as they directly affect radiative transfer and precipitation processes. In this study, we present a benchmark dataset and a baseline framework for transforming multimodal satellite observations into detailed 3D cloud phase structures, aiming toward operational cloud phase profile retrieval and future integration with NWP systems to improve cloud microphysics parameterization. The multimodal observations consist of (1) high--spatiotemporal--resolution, multi-band visible (VIS) and thermal infrared (TIR) imagery from geostationary satellites, and (2) accurate vertical cloud phase profiles from spaceborne lidar (CALIOP\slash CALIPSO) and radar (CPR\slash CloudSat). The dataset consists of synchronized image--profile pairs across diverse cloud regimes, defining a supervised learning task: given VIS/TIR patches, predict the corresponding 3D cloud phase structure. We adopt SGMAGNet as the main model and compare it with several baseline architectures, including UNet variants and SegNet, all designed to capture multi-scale spatial patterns. Model performance is evaluated using standard classification metrics, including Precision, Recall, F1-score, and IoU. The results demonstrate that SGMAGNet achieves superior performance in cloud phase reconstruction, particularly in complex multi-layer and boundary transition regions. Quantitatively, SGMAGNet attains a Precision of 0.922, Recall of 0.858, F1-score of 0.763, and an IoU of 0.617, significantly outperforming all baselines across these key metrics.

cs.CV↗

arXiv2025

Dissecting the NVIDIA Hopper Architecture through Microbenchmarking and Multiple Level Analysis

This study presents a comprehensive multi-level analysis of the NVIDIA Hopper GPU architecture, focusing on its performance characteristics and novel features. We benchmark Hopper's memory subsystem, highlighting improvements in the L2 partitioned cache and global memory access compared to Ampere and Ada Lovelace. The evaluation of Hopper's fourth-generation tensor cores reveals the benefits of FP8 precision and asynchronous wgmma instructions for matrix operations. Additionally, we investigate the performance of DPX instructions for dynamic programming, distributed shared memory (DSM) for inter-SM communication, and the Tensor Memory Accelerator (TMA) for asynchronous data movement. Through multi-level evaluation, we discover that the Hopper architecture demonstrates significant acceleration potential in real-world applications. For instance, the asynchronous programming model supported by TMA achieves a 1.5x speedup in matrix multiplication, FP8 delivers nearly double the performance of FP16, and DPX instructions accelerate a computational biology algorithm by at least 4.75x. Our findings provide actionable insights for optimizing compute-intensive workloads, from AI training to bioinformatics, on Hopper GPUs.

cs.DC↗

arXiv2025

HierMoE: Accelerating MoE Training with Hierarchical Token Deduplication and Expert Swap

The sparsely activated mixture-of-experts (MoE) transformer has become a common architecture for large language models (LLMs) due to its sparsity, which requires fewer computational demands while easily scaling the model size. In MoE models, each MoE layer requires to dynamically choose tokens to activate particular experts for computation while the activated experts may not be located in the same device or GPU as the token. However, this leads to substantial communication and load imbalances across all GPUs, which obstructs the scalability of distributed systems within a GPU cluster. To this end, we introduce HierMoE to accelerate the training of MoE models by two topology-aware techniques: 1) token deduplication to reduce the communication traffic, and 2) expert swap to balance the workloads among all GPUs. To enable the above two proposed approaches to be more general, we build theoretical models aimed at achieving the best token duplication and expert swap strategy under different model configurations and hardware environments. We implement our prototype HierMoE system atop Megatron-LM and conduct experiments on a 32-GPU cluster with DeepSeek-V3 and Qwen3-30B-A3B models. Experimental results show that our HierMoE achieves $1.55\times$ to $3.32\times$ faster communication and delivers $1.18\times$ to $1.27\times$ faster end-to-end training compared to state-of-the-art MoE training systems, Tutel-2DH, SmartMoE, and Megatron-LM.

cs.DC↗

arXiv2025

The Multi-Round Diagnostic RAG Framework for Emulating Clinical Reasoning

In recent years, accurately and quickly deploying medical large language models (LLMs) has become a trend. Among these, retrieval-augmented generation (RAG) has garnered attention due to rapid deployment and privacy protection. However, the challenge hinder the practical deployment of RAG for medical diagnosis: the semantic gap between colloquial patient descriptions and the professional terminology within medical knowledge bases. We try to address the challenge from the data perspective and the method perspective. First, to address the semantic gap in existing knowledge bases, we construct DiagnosGraph, a generalist knowledge graph covering both modern medicine and Traditional Chinese Medicine. It contains 876 common diseases with the graph of 7,997 nodes and 37,201 triples. To bridge the gap between colloquial patient narratives and academic medical knowledge, DiagnosGraph also introduces $1,908$ medical record by formalizing the patient chief complaint and proposing a medical diagnosis. Second, we introduce the Multi-Round Diagnostic RAG (MRD-RAG) framework. It utilizes a multi-round dialogue to refine diagnostic possibilities, emulating the clinical reasoning of a physician. Experiments conducted on four medical benchmarks, with evaluations by human physicians, demonstrate that MRD-RAG enhances the diagnostic performance of LLMs, highlighting its potential to make automated diagnosis more accurate and human-aligned.

cs.CL↗

arXiv2025

RouteMark: A Fingerprint for Intellectual Property Attribution in Routing-based Model Merging

Model merging via Mixture-of-Experts (MoE) has emerged as a scalable solution for consolidating multiple task-specific models into a unified sparse architecture, where each expert is derived from a model fine-tuned on a distinct task. While effective for multi-task integration, this paradigm introduces a critical yet underexplored challenge: how to attribute and protect the intellectual property (IP) of individual experts after merging. We propose RouteMark, a framework for IP protection in merged MoE models through the design of expert routing fingerprints. Our key insight is that task-specific experts exhibit stable and distinctive routing behaviors under probing inputs. To capture these patterns, we construct expert-level fingerprints using two complementary statistics: the Routing Score Fingerprint (RSF), quantifying the intensity of expert activation, and the Routing Preference Fingerprint (RPF), characterizing the input distribution that preferentially activates each expert. These fingerprints are reproducible, task-discriminative, and lightweight to construct. For attribution and tampering detection, we introduce a similarity-based matching algorithm that compares expert fingerprints between a suspect and a reference (victim) model. Extensive experiments across diverse tasks and CLIP-based MoE architectures show that RouteMark consistently yields high similarity for reused experts and clear separation from unrelated ones. Moreover, it remains robust against both structural tampering (expert replacement, addition, deletion) and parametric tampering (fine-tuning, pruning, permutation), outperforming weight- and activation-based baseliness. Our work lays the foundation for RouteMark as a practical and broadly applicable framework for IP verification in MoE-based model merging.

cs.CR↗

arXiv2025

FinKario: Event-Enhanced Automated Construction of Financial Knowledge Graph

Individual investors are significantly outnumbered and disadvantaged in financial markets, overwhelmed by abundant information and lacking professional analysis. Equity research reports stand out as crucial resources, offering valuable insights. By leveraging these reports, large language models (LLMs) can enhance investors' decision-making capabilities and strengthen financial analysis. However, two key challenges limit their effectiveness: (1) the rapid evolution of market events often outpaces the slow update cycles of existing knowledge bases, (2) the long-form and unstructured nature of financial reports further hinders timely and context-aware integration by LLMs. To address these challenges, we tackle both data and methodological aspects. First, we introduce the Event-Enhanced Automated Construction of Financial Knowledge Graph (FinKario), a dataset comprising over 305,360 entities, 9,625 relational triples, and 19 distinct relation types. FinKario automatically integrates real-time company fundamentals and market events through prompt-driven extraction guided by professional institutional templates, providing structured and accessible financial insights for LLMs. Additionally, we propose a Two-Stage, Graph-Based retrieval strategy (FinKario-RAG), optimizing the retrieval of evolving, large-scale financial knowledge to ensure efficient and precise data access. Extensive experiments show that FinKario with FinKario-RAG achieves superior stock trend prediction accuracy, outperforming financial LLMs by 18.81% and institutional strategies by 17.85% on average in backtesting.

cs.LG↗

arXiv2025

City-VLM: Towards Multidomain Perception Scene Understanding via Multimodal Incomplete Learning

Scene understanding enables intelligent agents to interpret and comprehend their environment. While existing large vision-language models (LVLMs) for scene understanding have primarily focused on indoor household tasks, they face two significant limitations when applied to outdoor large-scale scene understanding. First, outdoor scenarios typically encompass larger-scale environments observed through various sensors from multiple viewpoints (e.g., bird view and terrestrial view), while existing indoor LVLMs mainly analyze single visual modalities within building-scale contexts from humanoid viewpoints. Second, existing LVLMs suffer from missing multidomain perception outdoor data and struggle to effectively integrate 2D and 3D visual information. To address the aforementioned limitations, we build the first multidomain perception outdoor scene understanding dataset, named \textbf{\underline{SVM-City}}, deriving from multi\textbf{\underline{S}}cale scenarios with multi\textbf{\underline{V}}iew and multi\textbf{\underline{M}}odal instruction tuning data. It contains $420$k images and $4, 811$M point clouds with $567$k question-answering pairs from vehicles, low-altitude drones, high-altitude aerial planes, and satellite. To effectively fuse the multimodal data in the absence of one modality, we introduce incomplete multimodal learning to model outdoor scene understanding and design the LVLM named \textbf{\underline{City-VLM}}. Multimodal fusion is realized by constructing a joint probabilistic distribution space rather than implementing directly explicit fusion operations (e.g., concatenation). Experimental results on three typical outdoor scene understanding tasks show City-VLM achieves $18.14 \%$ performance surpassing existing LVLMs in question-answering tasks averagely. Our method demonstrates pragmatic and generalization performance across multiple outdoor scenes.

cs.CV↗

arXiv2025

RA-NeRF: Robust Neural Radiance Field Reconstruction with Accurate Camera Pose Estimation under Complex Trajectories

Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have emerged as powerful tools for 3D reconstruction and SLAM tasks. However, their performance depends heavily on accurate camera pose priors. Existing approaches attempt to address this issue by introducing external constraints but fall short of achieving satisfactory accuracy, particularly when camera trajectories are complex. In this paper, we propose a novel method, RA-NeRF, capable of predicting highly accurate camera poses even with complex camera trajectories. Following the incremental pipeline, RA-NeRF reconstructs the scene using NeRF with photometric consistency and incorporates flow-driven pose regulation to enhance robustness during initialization and localization. Additionally, RA-NeRF employs an implicit pose filter to capture the camera movement pattern and eliminate the noise for pose estimation. To validate our method, we conduct extensive experiments on the Tanks\&Temple dataset for standard evaluation, as well as the NeRFBuster dataset, which presents challenging camera pose trajectories. On both datasets, RA-NeRF achieves state-of-the-art results in both camera pose estimation and visual quality, demonstrating its effectiveness and robustness in scene reconstruction under complex pose trajectories.

cs.CV↗

arXiv2025

Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression

Post-training compression reduces the computational and memory costs of large language models (LLMs), enabling resource-efficient deployment. However, existing compression benchmarks only focus on language modeling (e.g., perplexity) and natural language understanding tasks (e.g., GLUE accuracy), ignoring the agentic capabilities - workflow, tool use/function call, long-context understanding and real-world application. We introduce the Agent Compression Benchmark (ACBench), the first comprehensive benchmark for evaluating how compression impacts LLMs' agentic abilities. ACBench spans (1) 12 tasks across 4 capabilities (e.g., WorfBench for workflow generation, Needle-in-Haystack for long-context retrieval), (2) quantization (GPTQ, AWQ) and pruning (Wanda, SparseGPT), and (3) 15 models, including small (Gemma-2B), standard (Qwen2.5 7B-32B), and distilled reasoning LLMs (DeepSeek-R1-Distill). Our experiments reveal compression tradeoffs: 4-bit quantization preserves workflow generation and tool use (1%-3% drop) but degrades real-world application accuracy by 10%-15%. We introduce ERank, Top-k Ranking Correlation and Energy to systematize analysis. ACBench provides actionable insights for optimizing LLM compression in agentic scenarios. The code can be found in https://github.com/pprp/ACBench.

cs.LG↗

arXiv2025

BurstGPT: A Real-world Workload Dataset to Optimize LLM Serving Systems

Serving systems for Large Language Models (LLMs) are often optimized to improve quality of service (QoS) and throughput. However, due to the lack of open-source LLM serving workloads, these systems are frequently evaluated under unrealistic workload assumptions. Consequently, performance may degrade when systems are deployed in real-world scenarios. This work presents BurstGPT, an LLM serving workload with 10.31 million traces from regional Azure OpenAI GPT services over 213 days. BurstGPT captures LLM serving characteristics from user, model and system perspectives: (1) User request concurrency: burstiness variations of requests in Azure OpenAI GPT services, revealing diversified concurrency patterns in different services and model types. (2) User conversation patterns: counts and intervals within conversations for service optimizations. (3) Model response lengths: auto-regressive serving processes of GPT models, showing statistical relations between requests and their responses. (4) System response failures: failures of conversation and API services, showing intensive resource needs and limited availability of LLM services in Azure. The details of the characteristics can serve multiple purposes in LLM serving optimizations, such as system evaluation and trace provisioning. In our demo evaluation with BurstGPT, frequent variations in BurstGPT reveal declines in efficiency, stability, or reliability in realistic LLM serving. We identify that the generalization of KV cache management, scheduling and disaggregation optimizations can be improved under realistic workload evaluations. BurstGPT is publicly available now at https://github.com/HPMLL/BurstGPT and is widely used to develop prototypes of LLM serving frameworks in the industry.

cs.DC↗

arXiv2025

Learning 6-DoF Fine-grained Grasp Detection Based on Part Affordance Grounding

Robotic grasping is a fundamental ability for a robot to interact with the environment. Current methods focus on how to obtain a stable and reliable grasping pose in object level, while little work has been studied on part (shape)-wise grasping which is related to fine-grained grasping and robotic affordance. Parts can be seen as atomic elements to compose an object, which contains rich semantic knowledge and a strong correlation with affordance. However, lacking a large part-wise 3D robotic dataset limits the development of part representation learning and downstream applications. In this paper, we propose a new large Language-guided SHape grAsPing datasEt (named LangSHAPE) to promote 3D part-level affordance and grasping ability learning. From the perspective of robotic cognition, we design a two-stage fine-grained robotic grasping framework (named LangPartGPD), including a novel 3D part language grounding model and a part-aware grasp pose detection model, in which explicit language input from human or large language models (LLMs) could guide a robot to generate part-level 6-DoF grasping pose with textual explanation. Our method combines the advantages of human-robot collaboration and LLMs' planning ability using explicit language as a symbolic intermediate. To evaluate the effectiveness of our proposed method, we perform 3D part grounding and fine-grained grasp detection experiments on both simulation and physical robot settings, following language instructions across different degrees of textual complexity. Results show our method achieves competitive performance in 3D geometry fine-grained grounding, object affordance inference, and 3D part-aware grasping tasks. Our dataset and code are available on our project website https://sites.google.com/view/lang-shape

cs.RO↗

arXiv2025

Jupiter: Fast and Resource-Efficient Collaborative Inference of Generative LLMs on Edge Devices

Generative large language models (LLMs) have garnered significant attention due to their exceptional capabilities in various AI tasks. Traditionally deployed in cloud datacenters, LLMs are now increasingly moving towards more accessible edge platforms to protect sensitive user data and ensure privacy preservation. The limited computational resources of individual edge devices, however, can result in excessively prolonged inference latency and overwhelmed memory usage. While existing research has explored collaborative edge computing to break the resource wall of individual devices, these solutions yet suffer from massive communication overhead and under-utilization of edge resources. Furthermore, they focus exclusively on optimizing the prefill phase, neglecting the crucial autoregressive decoding phase for generative LLMs. To address that, we propose Jupiter, a fast, scalable, and resource-efficient collaborative edge AI system for generative LLM inference. Jupiter introduces a flexible pipelined architecture as a principle and differentiates its system design according to the differentiated characteristics of the prefill and decoding phases. For prefill phase, Jupiter submits a novel intra-sequence pipeline parallelism and develops a meticulous parallelism planning strategy to maximize resource efficiency; For decoding, Jupiter devises an effective outline-based pipeline parallel decoding mechanism combined with speculative decoding, which further magnifies inference acceleration. Extensive evaluation based on realistic implementation demonstrates that Jupiter remarkably outperforms state-of-the-art approaches under various edge environment setups, achieving up to 26.1x end-to-end latency reduction while rendering on-par generation quality.

cs.DC↗

arXiv2025

SGDRC: Software-Defined Dynamic Resource Control for Concurrent DNN Inference on NVIDIA GPUs

Cloud service providers heavily colocate high-priority, latency-sensitive (LS), and low-priority, best-effort (BE) DNN inference services on the same GPU to improve resource utilization in data centers. Among the critical shared GPU resources, there has been very limited analysis on the dynamic allocation of compute units and VRAM bandwidth, mainly for two reasons: (1) The native GPU resource management solutions are either hardware-specific, or unable to dynamically allocate resources to different tenants, or both; (2) NVIDIA doesn't expose interfaces for VRAM bandwidth allocation, and the software stack and VRAM channel architectures are black-box, both of which limit the software-level resource management. These drive prior work to design either conservative sharing policies detrimental to throughput, or static resource partitioning only applicable to a few GPU models. To bridge this gap, this paper proposes SGDRC, a fully software-defined dynamic VRAM bandwidth and compute unit management solution for concurrent DNN inference services. SGDRC aims at guaranteeing service quality, maximizing the overall throughput, and providing general applicability to NVIDIA GPUs. SGDRC first reveals a general VRAM channel hash mapping architecture of NVIDIA GPUs through comprehensive reverse engineering and eliminates VRAM channel conflicts using software-level cache coloring. SGDRC applies bimodal tensors and tidal SM masking to dynamically allocate VRAM bandwidth and compute units, and guides the allocation of resources based on offline profiling. We evaluate 11 mainstream DNNs with real-world workloads on two NVIDIA GPUs. The results show that compared with the state-of-the-art GPU sharing solutions, SGDRC achieves the highest SLO attainment rates (99.0% on average), and improves overall throughput by up to 1.47x and BE job throughput by up to 2.36x.

cs.DC↗

arXiv2025

The Lottery LLM Hypothesis, Rethinking What Abilities Should LLM Compression Preserve?

Motivated by reducing the computational and storage costs of LLMs, model compression and KV cache compression have attracted much attention from researchers. However, current methods predominantly emphasize maintaining the performance of compressed LLMs, as measured by perplexity or simple accuracy on tasks of common sense knowledge QA and basic arithmetic reasoning. In this blog, we present a brief review of recent advancements in LLMs related to retrieval-augmented generation, multi-step reasoning, external tools, and computational expressivity, all of which substantially enhance LLM performance. Then, we propose a lottery LLM hypothesis suggesting that for a given LLM and task, there exists a smaller lottery LLM capable of producing the same performance as the original LLM with the assistance of multi-step reasoning and external tools. Based on the review of current progress in LLMs, we discuss and summarize the essential capabilities that the lottery LLM and KV cache compression must possess, which are currently overlooked in existing methods.

cs.LG↗

arXiv2025

DreamDDP: Accelerating Data Parallel Distributed LLM Training with Layer-wise Scheduled Partial Synchronization

The growth of large language models (LLMs) increases challenges of accelerating distributed training across multiple GPUs in different data centers. Moreover, concerns about data privacy and data exhaustion have heightened interest in geo-distributed data centers. Communication in geo-distributed data parallel training (DDP) with stochastic gradient descent (S-SGD) is the main bottleneck in low-bandwidth environments. Local SGD mitigates communication overhead by reducing synchronization frequency, and recent studies have successfully applied it to geo-distributedly pre-train LLMs. However, we identify that its model synchronization mechanism prevents overlapping communication and computation, which makes the system lose opportunities to overlap communication and computation. To overcome this limitation, we expand the design space of local SGD by layer-wisely decoupling model synchronization. In each iteration, only some layers are synchronized instead of the entire model after a specific number of iterations. Leveraging this methodology, we introduce DreamDDP, a training framework to accelerate low-bandwidth distributed training with three key innovations: (1) partial local SGD with theoretical assurances of convergence rates comparable to S-SGD; (2) overlapping parameter synchronization with computation without extra GPU memory occupation; (3) identifying and exploiting three properties to schedule the communication and computation to reduce the training time based on fine-grained profiling of layer-wise communication and computation time. Empirical evaluations conducted on 32 GPUs using prominent deep learning models, including ResNet-18, ResNet-50, GPT-2, and Llama-2, demonstrate that DreamDDP enhances the convergence properties of Local SGD (and Adam) and achieves speedups ranging from $1.49\times$ to $3.91\times$ over leading baseline methods.

cs.DC↗

arXiv2025

SphereFusion: Efficient Panorama Depth Estimation via Gated Fusion

Due to the rapid development of panorama cameras, the task of estimating panorama depth has attracted significant attention from the computer vision community, especially in applications such as robot sensing and autonomous driving. However, existing methods relying on different projection formats often encounter challenges, either struggling with distortion and discontinuity in the case of equirectangular, cubemap, and tangent projections, or experiencing a loss of texture details with the spherical projection. To tackle these concerns, we present SphereFusion, an end-to-end framework that combines the strengths of various projection methods. Specifically, SphereFusion initially employs 2D image convolution and mesh operations to extract two distinct types of features from the panorama image in both equirectangular and spherical projection domains. These features are then projected onto the spherical domain, where a gate fusion module selects the most reliable features for fusion. Finally, SphereFusion estimates panorama depth within the spherical domain. Meanwhile, SphereFusion employs a cache strategy to improve the efficiency of mesh operation. Extensive experiments on three public panorama datasets demonstrate that SphereFusion achieves competitive results with other state-of-the-art methods, while presenting the fastest inference speed at only 17 ms on a 512$\times$1024 panorama image.

cs.CV↗