arXiv subjects

Chong Zhu

Publications and source records attributed to Chong Zhu.

6 recordsLinked to original sources

arXiv2026

Gleam: Adaptive Network-Efficient CUDA API Remoting for Cross-Device GPU Sharing over LANs

This paper aims to enable computation- and communication-efficient GPU sharing across devices within local area networks (LANs), facilitating ubiquitous AI inference on heterogeneous personal devices. We achieve distributed task offloading via CUDA API remoting. However, beyond raw computation, network constraints emerge as the primary bottleneck: limited bandwidth, high-frequency API invocations, and cross-task contention significantly hinder performance. To address these challenges, we propose Gleam, a novel and network-efficient framework for task-generic GPU sharing across local-area CUDA devices, with three key contributions. First, we reduce bandwidth overhead in CUDA API remoting through automatic model weight caching, and mitigate accumulated latency from frequent API calls by asynchronous execution. Second, we design a runtime task scheduler that dynamically determines API remoting pairs between LAN clients and servers, explicitly accounting for both network conditions and GPU resource contention under parallel workloads. Finally, we introduce dedicated mechanisms to ensure CUDA context consistency across distributed executions. Extensive experiments on heterogeneous NVIDIA GPUs and diverse AI workloads show Gleam consistently outperforms state-of-the-art baselines, achieving 1.4-24.2 times improvements in API remoting efficiency and up to 1.79 times higher system throughput.

cs.DC↗

arXiv2025

Nexus: Higher-Order Attention Mechanisms in Transformers

Transformers have achieved significant success across various domains, relying on self-attention to capture dependencies. However, the standard first-order attention mechanism is often limited by a low-rank bottleneck, struggling to capture intricate, multi-hop relationships within a single layer. In this paper, we propose the Nexus, a novel architecture designed to enhance representational power through a recursive framework. Unlike standard approaches that use static linear projections for Queries and Keys, Nexus dynamically refines these representations via nested self-attention mechanisms. Specifically, the Query and Key vectors are themselves outputs of inner attention loops, allowing tokens to aggregate global context and model high-order correlations \textit{prior} to the final attention computation. We enforce a parameter-efficient weight-sharing strategy across recursive steps, ensuring that this enhanced expressivity incurs $\mathcal{O}(1)$ additional parameters. We provide theoretical analysis demonstrating that our method breaks the linear bottleneck of standard attention. Empirically, Nexus outperforms standard Transformers on multiple benchmarks.

cs.CL↗

arXiv2025

Loud-loss: A Perceptually Motivated Loss Function for Speech Enhancement Based on Equal-Loudness Contours

The mean squared error (MSE) is a ubiquitous loss function for speech enhancement, but its problem is that the error cannot reflect the auditory perception quality. This is because MSE causes models to over-emphasize low-frequency components which has high energy, leading to the inadequate modeling of perceptually important high-frequency information. To overcome this limitation, we propose a perceptually-weighted loss function grounded in psychoacoustic principles. Specifically, it leverages equal-loudness contours to assign frequency-dependent weights to the reconstruction error, thereby penalizing deviations in a way aligning with human auditory sensitivity. The proposed loss is model-agnostic and flexible, demonstrating strong generality. Experiments on the VoiceBank+DEMAND dataset show that replacing MSE with our loss in a GTCRN model elevates the WB-PESQ score from 2.17 to 2.93-a significant improvement in perceptual quality.

cs.SD↗

arXiv2025

Two-stage Audio-Visual Target Speaker Extraction System for Real-Time Processing On Edge Device

Audio-Visual Target Speaker Extraction (AVTSE) aims to isolate a target speaker's voice in a multi-speaker environment with visual cues as auxiliary. Most of the existing AVTSE methods encode visual and audio features simultaneously, resulting in extremely high computational complexity and making it impractical for real-time processing on edge devices. To tackle this issue, we proposed a two-stage ultra-compact AVTSE system. Specifically, in the first stage, a compact network is employed for voice activity detection (VAD) using visual information. In the second stage, the VAD results are combined with audio inputs to isolate the target speaker's voice. Experiments show that the proposed system effectively suppresses background noise and interfering voices while spending little computational resources.

cs.SD↗

arXiv2025

Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

This work presents Pangu Embedded, an efficient Large Language Model (LLM) reasoner developed on Ascend Neural Processing Units (NPUs), featuring flexible fast and slow thinking capabilities. Pangu Embedded addresses the significant computational costs and inference latency challenges prevalent in existing reasoning-optimized LLMs. We propose a two-stage training framework for its construction. In Stage 1, the model is finetuned via an iterative distillation process, incorporating inter-iteration model merging to effectively aggregate complementary knowledge. This is followed by reinforcement learning on Ascend clusters, optimized by a latency-tolerant scheduler that combines stale synchronous parallelism with prioritized data queues. The RL process is guided by a Multi-source Adaptive Reward System (MARS), which generates dynamic, task-specific reward signals using deterministic metrics and lightweight LLM evaluators for mathematics, coding, and general problem-solving tasks. Stage 2 introduces a dual-system framework, endowing Pangu Embedded with a "fast" mode for routine queries and a deeper "slow" mode for complex inference. This framework offers both manual mode switching for user control and an automatic, complexity-aware mode selection mechanism that dynamically allocates computational resources to balance latency and reasoning depth. Experimental results on benchmarks including AIME 2024, GPQA, and LiveCodeBench demonstrate that Pangu Embedded with 7B parameters, outperforms similar-size models like Qwen3-8B and GLM4-9B. It delivers rapid responses and state-of-the-art reasoning quality within a single, unified model architecture, highlighting a promising direction for developing powerful yet practically deployable LLM reasoners.

cs.CL↗

arXiv2025

TXSQL: Lock Optimizations Towards High Contented Workloads (Extended Version)

Two-phase locking (2PL) is a fundamental and widely used concurrency control protocol. It regulates concurrent access to database data by following a specific sequence of acquiring and releasing locks during transaction execution, thereby ensuring transaction isolation. However, in strict 2PL, transactions must wait for conflicting transactions to commit and release their locks, which reduces concurrency and system throughput. We have observed this issue is exacerbated in high-contented workloads at Tencent, where lock contention can severely degrade system performance. While existing optimizations demonstrate some effectiveness in high-contention scenarios, their performance remains insufficient, as they suffer from lock contention and waiting in hotspot access. This paper presents optimizations in lock management implemented in Tencent's database, TXSQL, with a particular focus on high-contention scenarios. First, we discuss our motivations and the journey toward general lock optimization, which includes lightweight lock management, a copy-free active transaction list, and queue locking mechanisms that effectively enhance concurrency. Second, we introduce a hotspot-aware approach that enables certain highly conflicting transactions to switch to a group locking method, which groups conflicting transactions at a specific hotspot, allowing them to execute serially in an uncommitted state within a conflict group without the need for locking, thereby reducing lock contention. Our evaluation shows that under high-contented workloads, TXSQL achieves performance improvements of up to 6.5x and up to 22.3x compared to state-of-the-art methods and systems, respectively.

cs.DB↗