Searcharxiv⌕ Search

arXiv subjects

Dezun Dong

Publications and source records attributed to Dezun Dong.

At least 19 recordsLinked to original sources

MpFA: Hardware-Efficient Train-Free QK4V8 FlashAttention Kernels on Blackwell GPUs

Long-context LLM inference pushes modern GPU serving stacks into an attention-bound regime, where both compute and memory are dominated by the softmax-GEMM pipeline. On NVIDIA Blackwell GPUs, FP4 Tensor Cores offer high matmul throughput, but we find that fully FP4 attention often fails to translate this throughput into end-to-end speedups due to non-matmul costs: online quantization after softmax, tensor/shared-memory data movement, and contention on the softmax path. We present MpFA, a training-free FlashAttention kernel optimized for Blackwell. Guided by hardware characterization, MpFA uses mixed precision: NVFP4 for QK and FP8 for PV (QK4PV8). This preserves low-bit QK throughput while avoiding the conversion and scaling overheads of FP4 PV. To recover accuracy without further stressing the softmax pipeline, MpFA introduces rank-one smoothing compensation implemented as an additional Tensor Core MMA. MpFA further improves performance with a fine-grained asynchronous pipeline, tensor-memory reuse, and adaptive parallel partitioning across prefill and decode. On an NVIDIA SM100 GPU and across 16K-128K contexts, MpFA improves prefill throughput over state-of-the-art BF16/FP8 baselines and increases end-to-end output throughput by 2.81$\times$ over BF16 FA4 across Llama-3.1-8B and Qwen3-14B. Across five benchmark suites and two models, rank-one compensation recovers 62.5% of the accuracy loss with about 2.0% kernel overhead.

cs.DC↗

Rethinking Adaptive Routing on Dragonfly

The Dragonfly topology is widely deployed in leading supercomputers, where adaptive routing is crucial for high-performance interconnection. Despite years of research, a dilemma remains: state-of-the-art adaptive routing algorithms achieve excellent performance but demand hardware budgets far beyond what real router chips can afford -- prohibitive buffers, logically dependent floating-point operations, or exhaustive path-set traversals -- whereas realizable algorithms deliver unsatisfactory performance. Furthermore, existing algorithms fail to exploit the abundant almost-minimal paths in Dragonfly: when only the local channels of a minimal path are congested, they reroute packets to significantly longer global non-minimal paths, wasting global bandwidth and increasing latency. To address both limitations, we analyze intra-group path contention and propose the Decoupled Incremental Adaptive Routing (DIAR) algorithm, which decouples routing into global path selection at the source and incremental local detours onto almost-minimal paths, guaranteeing deadlock freedom without additional virtual channels, escape rings, or turn models. Requiring only a few comparators, shifters, adders, and multiplexers, DIAR reduces buffer overhead by up to four orders of magnitude compared with state-of-the-art algorithms, while achieving comparable or superior performance across all evaluated scenarios: relative to the strongest realizable adaptive routing algorithm in each scenario, DIAR improves saturation throughput by up to 30.0% across synthetic traffic patterns and improves performance by up to 34.9% (13.2% on average) across HPC workloads.

cs.NI↗

PDScope: Breaking the Prefill-Decode Seesaw for Memory-Bound MoE Inference

Local deployment of MoE-based LLMs on CPU-GPU nodes is attractive for cost and privacy reasons, yet limited GPU memory forces most experts into CPU memory. Prefill is then compute-bound and benefits from aggressive GPU loading, while decode is I/O-bound and requires conservative PCIe transfers. These demands cannot be amortized under low batching---typical of local assistants and document tools---forcing phase-agnostic schedulers into a seesaw: favoring prefill's aggressive loading injects unmaskable I/O that stalls decode, while favoring decode's conservative transfers starves prefill's GPU of work. We observe that expert routing is predictable in phase-specific ways, allowing us to decouple the two phases and optimize prediction, scheduling, and execution independently for each phase. PDScope predicts expert activations to guide placement and prefetching, adapts these policies to each phase's bottleneck via cross-layer cost modeling, and compiles decisions into overhead-free pipelines using a static computation graph, cache prefetching, and zero-copy asynchronous I/O. On mainstream MoE models, PDScope achieves 2.5-10.3x higher prefill throughput and 2.1-6.1x higher decode throughput than existing systems, showing that phase-aware co-design unlocks efficient MoE inference under memory constraints and extends naturally to multi-batch settings.

cs.DC↗

HiNa-MoE: High-Performance, Non-Intrusive MoE Inference on CPUs with Matrix Engines

Mixture-of-Experts (MoE) inference is increasingly deployed in local and on-premise environments, where expert parameters often exceed GPU memory capacity. In latency-sensitive, low-concurrency settings, repeatedly staging routed-expert weights from CPU memory to the GPU can be prohibitive, leaving routed-expert feed-forward networks (FFNs) on the critical path of multi-socket CPUs. Existing CPU accelerations often rely on intrusive, hardware- or topology-specific requirements, such as AMX-specific weight layouts or manual NUMA-aware placement. These requirements reduce portability and complicate integration with standard CPU-GPU offloading pipelines. We present HiNa-MoE, a high-performance, non-intrusive operator library for MoE inference on CPUs with Intel AMX. HiNa-MoE (1) exploits AMX with an optimized micro-kernel that keeps expert weights in standard layouts and instead fuses lightweight layout transforms into token gathering and stores; (2) applies NUMA-aware task partitioning under a simple page-interleaved policy without modifying the framework allocator; and (3) converts decode-phase memory matrix-vector operations into small matrix-matrix execution to utilize AMX. Across multiple MoE models, HiNa-MoE achieves up to 3.37x speedup for FFN kernels and up to 2.09x end-to-end inference speedup over state-of-the-art baselines, while remaining plug-and-play with existing frameworks and deployment workflows.

cs.DC↗

Physically Partitioned KVCache Format for CPU--GPU Load Balancing in MoE Inference

Single-GPU long-context inference with Mixture-of-Experts (MoE) models requires spilling the key-value cache (KVCache) to CPU memory. The spilled KV serves two complementary purposes---transferring to the GPU for attention computation, or computing in-place on the CPU---which demand opposing physical states. The optimal split between them varies with workload, yet existing KVCache abstractions offer only storage semantics over a monolithic object of a single physical state, and cannot express dynamic load balancing. We propose InplaceKVCache, the first KVCache abstraction whose format fixes each byte's physical residency at write time, so that the CPU--GPU load balance can be adjusted without moving data after placement. It realizes this as a four-region layout along two dimensions---device affinity and access pattern---turning load balancing into pure scheduling. Built on this abstraction, WriteScope splits CPU--GPU shares along the sequence dimension, and a portable roofline performance model determines the optimal CPU share as sequence length evolves, with online feedback tracking CPU cost drift. On three MoE models (DeepSeek-V2-Lite, Qwen3-30B-A3B, Mixtral-8$\times$7B) with a 32~GB VRAM budget, WriteScope supports end-to-end inference at the 1M-token aggregate scale. In the long-context regime ($\ge$8K), it achieves geometric-mean speedups of $1.5\times$--$2.5\times$ on A100 and $1.4\times$--$1.7\times$ on V100 over four reproduced baselines, while vLLM, SGLang, and KTransformers fail even with a doubled KV budget. A DeepSeek-V4-Flash case study validates composition with native sparse attention.

cs.DC↗

Multi-Plane HyperX: A Low-Latency and Cost-Effective Network for Large-Scale AI and HPC Systems

Multi-plane architectures have become increasingly prevalent in the Fat-Tree networks of AI data centers. By leveraging multiple ports on a single network interface card (NIC) or multiple NICs within a scale-up domain, each port or NIC is allocated to an independent network plane, thereby provisioning the overall system with multiple network planes. However, no prior literature has explored the application of multi-plane technologies to direct networks such as HyperX. This paper investigates the multi-plane HyperX network and demonstrates that, compared to state-of-the-art network topologies like multi-plane Fat-Tree, Dragonfly, and Dragonfly+, the multi-plane HyperX architecture achieves a significantly smaller network diameter and superior cost-effectiveness.

cs.NI↗

LayerScope: Predictive Cross-Layer Scheduling for Efficient Multi-Batch MoE Inference on Legacy Servers

Mixture-of-Experts (MoE) models face memory and PCIe latency bottlenecks when deployed on commodity hardware. Offloading expert weights to CPU memory results in PCIe transfer latency that exceeds GPU computation by several folds. We present PreScope, a prediction-driven expert scheduling system that addresses three key challenges: inaccurate activation prediction, PCIe bandwidth competition, and cross-device scheduling complexity. Our solution includes: 1) Learnable Layer-Aware Predictor (LLaPor) that captures layer-specific expert activation patterns; 2) Prefetch-Aware Cross-Layer Scheduling (PreSched) that generates globally optimal plans balancing prefetching costs and loading overhead; 3) Asynchronous I/O Optimizer (AsyncIO) that decouples I/O from computation, eliminating waiting bubbles. PreScope achieves 141% higher throughput and 74.6% lower latency than state-of-the-art solutions.

cs.LG↗

Nezha: Breaking Multi-Rail Network Barriers for Distributed DNN Training

In distributed deep learning, communication remains a critical bottleneck. While modern hardware advances rapidly, over 60 percent of production HPC systems still rely on legacy infrastructure (V100 GPUs, multi-plane Ethernet/InfiniBand), necessitating communication optimization without hardware upgrades. Existing approaches face three key limitations: (1) static single-rail binding underutilizes multi-rail bandwidth, (2) protocol heterogeneity (TCP-RDMA coexistence) causes synchronization delays, and (3) mainstream libraries (NCCL/MPI) lack cross-protocol coordination. We present Nezha, the first protocol-agnostic system for multi-rail networks. Our contributions include: (1) Hardware-agnostic cross-protocol coordination: A unified abstraction enabling seamless collaboration between in-network computing (SHARP), adaptive RDMA (GLEX), and TCP, achieving 1.7 to 4.3 times lower latency than Gloo. (2) Protocol-aware dynamic load balancing: A hybrid scheduling strategy with cold/hot start state machine for heterogeneous protocols, reducing startup latency for small payloads while enhancing throughput for large transfers. (3) Fault-tolerant multi-rail collaboration: A self-recovery mechanism that reroutes data flows within 200 milliseconds upon single-rail failures, ensuring uninterrupted training. Experiments on 8-node clusters demonstrate Nezha achieves 74 percent and 80 percent higher throughput than MPTCP in homogeneous (TCP-TCP) and heterogeneous (TCP-SHARP) networks, respectively. On 128-node supercomputers, Nezha delivers 2.36 times higher training efficiency than Gloo. By bridging modern DNN communication demands with legacy infrastructure, Nezha proves that systematic multi-rail optimization can unlock the potential of aging clusters.

cs.DC↗

Mitigating Sensitive Information Leakage in LLMs4Code through Machine Unlearning

Large Language Models for Code (LLMs4Code) have achieved strong performance in code generation, but recent studies reveal that they may memorize and leak sensitive information contained in training data, posing serious privacy risks. To address this gap, this work presents the first comprehensive empirical study on applying machine unlearning to mitigate sensitive information leakage in LLMs4Code. We first construct a dedicated benchmark that includes: (i) a synthetic forget set containing diverse forms of personal information, and (ii) a retain set designed to evaluate whether code-generation capability is preserved after unlearning. Using this benchmark, we systematically assess three representative unlearning algorithms (GA, GA+GD, GA+KL) across three widely used open-source LLMs4Code models (AIXCoder-7B, CodeLlama-7B, CodeQwen-7B). Experimental results demonstrate that machine unlearning can substantially reduce direct memorization-based leakage: on average, the direct leak rate drops by more than 50% while retaining about over 91% of the original code-generation performance. Moreover, by analyzing post-unlearning outputs, we uncover a consistent shift from direct to indirect leakage, revealing an underexplored vulnerability that persists even when the target data has been successfully forgotten. Our findings show that machine unlearning is a feasible and effective solution for enhancing privacy protection in LLMs4Code, while also highlighting the need for future techniques capable of mitigating both direct and indirect leakage simultaneously.

cs.CR↗

Cohet: A CXL-Driven Coherent Heterogeneous Computing Framework with Hardware-Calibrated Full-System Simulation

Conventional heterogeneous computing systems built on PCIe interconnects suffer from inefficient fine-grained host-device interactions and complex programming models. In recent years, many proprietary and open cache-coherent interconnect standards have emerged, among which compute express link (CXL) prevails in the open-standard domain after acquiring several competing solutions. Although CXL-based coherent heterogeneous computing holds the potential to fundamentally transform the collaborative computing mode of CPUs and XPUs, research in this direction remains hampered by the scarcity of available CXL-supported platforms, immature software/hardware ecosystems, and unclear application prospects. This paper presents Cohet, the first CXL-driven coherent heterogeneous computing framework. Cohet decouples the compute and memory resources to form unbiased CPU and XPU pools which share a single unified and coherent memory pool. It exposes a standard malloc/mmap interface to both CPU and XPU compute threads, leaving the OS dealing with smart memory allocation and management of heterogeneous resources. To facilitate Cohet research, we also present a full-system cycle-level simulator named SimCXL, which is capable of modeling all CXL sub-protocols and device types. SimCXL has been rigorously calibrated against a real CXL testbed with various CXL memory and accelerators, showing an average simulation error of 3%. Our evaluation reveals that CXL.cache reduces latency by 68% and increases bandwidth by 14.4x compared to DMA transfers at cacheline granularity. Building upon these insights, we demonstrate the benefits of Cohet with two killer apps, which are remote atomic operation (RAO) and remote procedure call (RPC). Compared to PCIe-NIC design, CXL-NIC achieves a 5.5 to 40.2x speedup for RAO offloading and an average speedup of 1.86x for RPC (de)serialization offloading.

cs.AR↗

CXL-DMSim: A Full-System CXL Disaggregated Memory Simulator With Comprehensive Silicon Validation

Compute eXpress Link (CXL) has emerged as a key enabler of memory disaggregation for future heterogeneous computing systems to expand memory on-demand and improve resource utilization. However, CXL is still in its infancy stage and lacks commodity products on the market, thus necessitating a reliable system-level simulation tool for research and development. In this paper, we propose CXL-DMSim, an open-source full-system simulator to simulate CXL disaggregated memory systems with high fidelity at a gem5-comparable simulation speed. CXL-DMSim incorporates a flexible CXL memory expander model along with its associated device driver, and CXL protocol support with CXL.io and CXL.mem. It can operate in both app-managed mode and kernel-managed mode, with the latter using a dedicated NUMA-compatible mechanism. The simulator has been rigorously verified against a real hardware testbed with both FPGAand ASIC-based CXL memory devices, which demonstrates the qualification of CXL-DMSim in simulating the characteristics of various CXL memory devices at an average simulation error of 3.4%. The experimental results using LMbench and STREAM benchmarks suggest that the CXL-FPGA memory exhibits a ~2.88x higher latency than local DDR while the CXL-ASIC latency is ~2.18x; CXL-FPGA achieves 45-69% of local DDR memory bandwidth, whereas the number for CXLASIC is 82-83%. The study also reveals that CXL memory can significantly enhance the performance of memory-intensive applications, improved by 23x at most with limited local memory for Viper key-value database and approximately 60% in memorybandwidth-sensitive scenarios such as MERCI. Moreover, the simulator's observability and expandability are showcased with detailed case-studies, highlighting its great potential for research on future CXL-interconnected hybrid memory pool.

cs.ET↗

Demystifying ARM SME to Optimize General Matrix Multiplications

General Matrix Multiplication (GEMM) is a critical kernel in high-performance computing and deep learning. While modern architectures like ARM's Scalable Matrix Extension (SME) introduce dedicated hardware for matrix operations, existing linear algebra libraries fail to fully exploit its potential, particularly for large matrices. This paper presents MpGEMM, an open-source library that leverages key architectural features of SME to optimize GEMM across multiple precisions. Through a systematic characterization of SME, we derive optimization guidelines that inform our design. MpGEMM employs cache-aware partitioning, efficient data packing with on-the-fly transposition, and specialized micro-kernels that utilize multi-vector loads and all available tile registers. Evaluated on an Apple M4 Pro with real-world workloads from DeepSeek and LLaMA, MpGEMM achieves an average speedup of 1.23x over the vendor-optimized Apple Accelerate library and significantly outperforms other open-source alternatives.

cs.DC↗

OnlineAnySeg: Online Zero-Shot 3D Segmentation by Visual Foundation Model Guided 2D Mask Merging

Online zero-shot 3D instance segmentation of a progressively reconstructed scene is both a critical and challenging task for embodied applications. With the success of visual foundation models (VFMs) in the image domain, leveraging 2D priors to address 3D online segmentation has become a prominent research focus. Since segmentation results provided by 2D priors often require spatial consistency to be lifted into final 3D segmentation, an efficient method for identifying spatial overlap among 2D masks is essential - yet existing methods rarely achieve this in real time, mainly limiting its use to offline approaches. To address this, we propose an efficient method that lifts 2D masks generated by VFMs into a unified 3D instance using a hashing technique. By employing voxel hashing for efficient 3D scene querying, our approach reduces the time complexity of costly spatial overlap queries from $O(n^2)$ to $O(n)$. Accurate spatial associations further enable 3D merging of 2D masks through simple similarity-based filtering in a zero-shot manner, making our approach more robust to incomplete and noisy data. Evaluated on the ScanNet and SceneNN benchmarks, our approach achieves state-of-the-art performance in online, zero-shot 3D instance segmentation with leading efficiency.

cs.CV↗

UNR: Unified Notifiable RMA Library for HPC

Remote Memory Access (RMA) enables direct access to remote memory to achieve high performance for HPC applications. However, most modern parallel programming models lack schemes for the remote process to detect the completion of RMA operations. Many previous works have proposed programming models and extensions to notify the communication peer, but they did not solve the multi-NIC aggregation, portability, hardware-software co-design, and usability problems. In this work, we proposed a Unified Notifiable RMA (UNR) library for HPC to address these challenges. In addition, we demonstrate the best practice of utilizing UNR within a real-world scientific application, PowerLLEL. We deployed UNR across four HPC systems, each with a different interconnect. The results show that PowerLLEL powered by UNR achieves up to a 36% acceleration on 1728 nodes of the Tianhe-Xingyi supercomputing system.

cs.DC↗

Large Language Models are Few-Shot Summarizers: Multi-Intent Comment Generation via In-Context Learning

Code comment generation aims at generating natural language descriptions for a code snippet to facilitate developers' program comprehension activities. Despite being studied for a long time, a bottleneck for existing approaches is that given a code snippet, they can only generate one comment while developers usually need to know information from diverse perspectives such as what is the functionality of this code snippet and how to use it. To tackle this limitation, this study empirically investigates the feasibility of utilizing large language models (LLMs) to generate comments that can fulfill developers' diverse intents. Our intuition is based on the facts that (1) the code and its pairwise comment are used during the pre-training process of LLMs to build the semantic connection between the natural language and programming language, and (2) comments in the real-world projects, which are collected for the pre-training, usually contain different developers' intents. We thus postulate that the LLMs can already understand the code from different perspectives after the pre-training. Indeed, experiments on two large-scale datasets demonstrate the rationale of our insights: by adopting the in-context learning paradigm and giving adequate prompts to the LLM (e.g., providing it with ten or more examples), the LLM can significantly outperform a state-of-the-art supervised learning approach on generating comments with multiple intents. Results also show that customized strategies for constructing the prompts and post-processing strategies for reranking the results can both boost the LLM's performances, which shed light on future research directions for using LLMs to achieve comment generation.

cs.SE↗

CD-SGD: Distributed Stochastic Gradient Descent with Compression and Delay Compensation

Communication overhead is the key challenge for distributed training. Gradient compression is a widely used approach to reduce communication traffic. When combining with parallel communication mechanism method like pipeline, gradient compression technique can greatly alleviate the impact of communication overhead. However, there exists two problems of gradient compression technique to be solved. Firstly, gradient compression brings in extra computation cost, which will delay the next training iteration. Secondly, gradient compression usually leads to the decrease of convergence accuracy.

cs.LG↗

SSD-SSD: Communication sparsification for distributed deep learning training

Intensive communication and synchronization cost for gradients and parameters is the well-known bottleneck of distributed deep learning training. Based on the observations that Synchronous SGD (SSGD) obtains good convergence accuracy while asynchronous SGD (ASGD) delivers a faster raw training speed, we propose Several Steps Delay SGD (SSD-SGD) to combine their merits, aiming at tackling the communication bottleneck via communication sparsification. SSD-SGD explores both global synchronous updates in the parameter servers and asynchronous local updates in the workers in each periodic iteration. The periodic and flexible synchronization makes SSD-SGD achieve good convergence accuracy and fast training speed. To the best of our knowledge, we strike the new balance between synchronization quality and communication sparsification, and improve the trade-off between accuracy and training speed. Specifically, the core components of SSD-SGD include proper warm-up stage, steps delay stage, and our novel algorithm of global gradient for local update (GLU). GLU is critical for local update operations to effectively compensate the delayed local weights. Furthermore, we implement SSD-SGD on MXNet framework and comprehensively evaluate its performance with CIFAR-10 and ImageNet datasets. Experimental results show that SSD-SGD can accelerate distributed training speed under different experimental configurations, by up to 110%, while achieving good convergence accuracy.

cs.DC↗

Communication optimization strategies for distributed deep neural network training: A survey

Recent trends in high-performance computing and deep learning have led to the proliferation of studies on large-scale deep neural network training. However, the frequent communication requirements among computation nodes drastically slows the overall training speeds, which causes bottlenecks in distributed training, particularly in clusters with limited network bandwidths. To mitigate the drawbacks of distributed communications, researchers have proposed various optimization strategies. In this paper, we provide a comprehensive survey of communication strategies from both an algorithm viewpoint and a computer network perspective. Algorithm optimizations focus on reducing the communication volumes used in distributed training, while network optimizations focus on accelerating the communications between distributed devices. At the algorithm level, we describe how to reduce the number of communication rounds and transmitted bits per round. In addition, we elucidate how to overlap computation and communication. At the network level, we discuss the effects caused by network infrastructures, including logical communication schemes and network protocols. Finally, we extrapolate the potential future challenges and new research directions to accelerate communications for distributed deep neural network training.

cs.DC↗