WZWenyu Zhao

Wenyu Zhao

*This is an algorithmically generated author record based on publicly available information.

Wenyu Zhao has a total of 11 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    XAccel: Constraint-Guided Cross-Accelerator Inference Optimization for Latency, Cost, and Energy Efficiency

    by Wenyu Zhao, Boyuan Wang

    The same machine learning model can be served by GPUs and AI accelerators that differ by more than an order of magnitude in throughput, price, memory capacity, and power, yet the best execution configuration depends jointly on the model, its tail-latency service-level objective (SLO), its memory footprint, and the load it must sustain. This paper presents XAccel, a constraint-guided framework that selects and adapts inference configurations across heterogeneous accelerators under latency, cost, and energy budgets. XAccel (i) builds a unified, SLO-aware performance descriptor from 2,173 audited MLPerf Inference runs (v4.0-v6.0) covering 28 accelerator types and 34 workloadscenario pairs, including p95/p99 latency, time-to-first-token and time-per-output-token percentiles, and wall power reconstructed for 88 runs; (ii) selects configurations by solving an SLO-, memory-, and budget-constrained cost-energy program over measured operating points; (iii) transfers performance to unmeasured accelerators and models through a datasheetinformed elasticity model with hierarchical anchor calibration; and (iv) re-selects fleets under week-long production traces of Azure LLM services. Across 30 workload-SLO settings, choosing the fastest, the cheapest-per-hour, or the incumbent H100 accelerator raises the cost per inference by 56-152% on average relative to XAccel. Zero-shot transfer predicts SLO-bounded throughput with 28% mean absolute percentage error and preserves 94% of selection decisions at 1.6% mean cost regret, and the cost-optimal accelerator for a new model is found after 2.2 measurements instead of 4.5. Under real production load, XAccel lowers weekly cost by 69-82% and energy by 69-80% relative to a peak-provisioned H100 fleet for LLM services, with at most 0.51% SLO-risk minutes.

  • Open Access

    PreflightML: Constraint-Aware Pre-Deployment Validation for Reliable and Portable LLM-Generated Machine Learning Pipelines

    by Wenyu Zhao, Tingjie Chen, Jie Si Yang

    Large language models (LLMs) now write a growing share of machine-learning (ML) pipeline code, yet code that runs correctly on a developer's machine often fails after it is containerized and scheduled on a cluster whose images, library versions, quotas and runtime configuration differ from the development environment. This paper presents PreflightML, a constraint-aware pre-deployment validator for LLM-generated ML pipelines on Kubernetes. PreflightML compiles code, requirements and Job manifests into a unified constraint model spanning dependencies, resources and the execution environment. It extracts API constraints once, from an instrumented run in the development environment, and projects them onto every target image through a retrieval-based, version-aware API knowledge base that also uses the deprecation warnings observed in that run as evidence. Only uncertain cases are escalated to lightweight dryruns, ranging from skeleton probes to row-sampled executions whose peak memory and runtime are extrapolated with a fitted power law. Violations are repaired by API migration, image retargeting and manifest right-sizing, and an online risk model with a failure-signature memory learns from execution feedback. On 13,085 deployments of 2,617 locally correct DS-1000 programs written by six LLMs and humans to five real container images, the static tier detects 50.3% of all failures and 90.3% of dependency and API failures with no false alarm at 8.2 µs per deployment, and learned escalation raises recall to 72.2% while dry-running 15.6% of the deployments. On 203 containerized training-job deployments with resource, configuration and dependency faults, PreflightML detects every configuration and dependency failure at 58% of the cost of canary runs, predicts peak memory within 13.5% mean error, and repairs 39 of 40 detected configuration faults.

  • Open Access

    Budget-Constrained Evidence Distillation and Risk-Controlled Escalation for LLM-Augmented Failure Triage in Large-Scale AI Development Pipelines

    by Wenyu Zhao

    Large-scale artificial-intelligence (AI) development pipelines fail constantly, and each failure produces thousands of log lines that an engineer, or increasingly a large language model (LLM), must read in order to assign a root cause. Feeding whole logs to an LLM is economically impossible at fleet scale, and the truncation heuristics used in practice-keep the first or the last k tokens, or grep for ERROR-discard the decisive evidence surprisingly often. We formalise the resulting bottleneck as budget-constrained evidence distillation: given a token budget B, select the subset of log lines that maximises the diagnostic information handed to a downstream reasoner. We present Ledger, which (i) scores every line with a contrastive template salience learned from run-level pass/fail labels only, augmented by rarity, severity, temporal-locality and (optionally) entity-sequence and parameter-anomaly channels; (ii) selects lines by maximising a monotone submodular objective that combines facility-location coverage with a concave with severity filtering and 14.3% with tail truncation; the LLM was correct on every excerpt that contained the evidence, so the entire gap is attributable to selection. The conformal router captures 100% of never-before-seen failure modes while issuing 35% fewer LLM calls than confidence-threshold routing. We also report two negative results: the parameter-anomaly channel pays off only on the 4.5% of evidence lines whose template also occurs in passing runs, and strict within-group deduplication-a natural design choice-costs 15 points of recall.

  • Top Cited Papers
    Open Access

    DRL-Adapt: Deep Reinforcement Learning for Adaptive Routing Convergence Optimization in Large-Scale Networks

    by Boyuan Wang, Zimeng Wang, Wenyu Zhao, Fengyuan Zhang, Wenbin Shang

    Large-scale fully-meshed networks often experience prolonged routing convergence after failures because routing protocol parameters are statically configured and cannot adapt to changing network conditions. To address this limitation, this paper proposes DRL-Adapt, a deep reinforcement learning-based framework that dynamically tunes routing protocol parameters using network event-stream observations. The framework employs a dueling prioritized deep Q-network agent to jointly adjust Hello/Dead interval timers and routing weights according to real-time network states. We evaluate the proposed method using a high-fidelity simulator constructed from CAIDA topologies under diverse failure scenarios and network scales. Experimental results demonstrate that DRL-Adapt significantly accelerates recovery and reduces control overhead compared with conventional static configurations. On average, the proposed approach decreases fault convergence time by 32.7% and lowers control-plane overhead by 24.3%, while preserving routing stability. Moreover, the learned policy generalizes across networks ranging from 50 to 500 nodes and incurs minimal inference latency ($< $ 15 ms/decision), indicating practicality for practical deployment.


  • Top Cited Papers
    Open Access

    AutoML-Pipeline: A RAG-Enhanced Code Generation Framework With Pre-Validation for Cloud-Native Machine Learning Workflows

    by Wenyu Zhao, Tingjie Chen, Jie Si Yang, Lei Qiu

    The proliferation of cloud-native machine learning platforms has significantly accelerated model development and deployment cycles. However, constructing and maintaining heterogeneous pipeline code spanning multiple languages (Python, YAML, Spark SQL) and cloud-specific configurations remains labor-intensive and error-prone. Existing LLM-based code generation tools lack awareness of runtime constraints and historical execution patterns, frequently producing code with resource misconfigurations or dependency conflicts that fail upon deployment. To address these challenges, we propose AutoML-Pipeline, a closed-loop code generation framework that integrates Retrieval-Augmented Generation (RAG) with reinforcement learning feedback mechanisms. Our approach leverages a knowledge base constructed from successful pipeline execution logs to guide GPT-4 in generating deployment-ready code that adheres to platform-specific constraints. The key innovation lies in a novel Pre-validation Agent that employs simulated execution environments to predict resource consumption and detect dependency conflicts before actual deployment. This agent iteratively refines generated code through a feedback loop informed by predicted execution profiles and dependency graphs. We evaluate our framework on the CodeSearchNet dataset augmented with Azure ML pipeline specifications, demonstrating a 43.7% improvement in first-submission success rate and 31.2% reduction in resource over-provisioning compared to vanilla GPT-4 baselines. Ablation studies confirm that both the RAG retrieval mechanism and pre-validation agent contribute substantially to performance gains. Our work establishes a practical paradigm for integrating large language models with domain-specific runtime intelligence, with potential applications extending to other infrastructure-as-code generation tasks.


  • Top Cited Papers
    Open Access

    CAGR: A Cross-Accelerator Graph Optimization Framework for Efficient Recommender System Inference

    by Zijian Shen, Wenyu Zhao, Boyuan Wang, Zimeng Wang, Wenbin Shang

    Recommender systems have become ubiquitous in modern online services, yet their deployment across diverse hardware accelerators remains challenging due to significant performance variations. Contemporary deep learning recommendation models (DLRMs), such as DeepFM and NGCF, exhibit substantial inference latency differences when executed on NVIDIA GPUs,AMDGPUs, and Google TPUs, primarily due to architectural disparities and vendor-specific optimization strategies. Existing graph optimization frameworks are typically designed for specific hardware backends, lacking the flexibility to generate portable high-performance implementations across heterogeneous accelerators. This paper presents CAGR (Cross-Accelerator Graph Rewriting), a novel framework that achieves performance-portable inference optimization for recommendation models through three key innovations: (1) a hardware-aware graph rewriting engine that dynamically selects optimal operator implementations by analyzing compute-to-memory bandwidth ratios and operator density characteristics; (2) a reinforcement learning-based transformation policy that learns cross-platform optimization strategies without exhaustive search; and (3) a heterogeneous pipeline architecture enabling “optimize once, deploy across supported backends” semantics. We implement CAGR with support for multiple kernel backends including Triton, cuBLAS, MIOpen, and oneDNN, and demonstrate its effectiveness on the Avazu CTR prediction dataset. Experimental results show that CAGR achieves 1.8-3.2× speedup over baseline implementations across NVIDIA V100, AMD MI100, and Google TPU v3 platforms, while reducing optimization time by 67% compared to platform-specific auto-tuning approaches. Furthermore, CAGR maintains 92-96% of reference optimized performance with zero manual intervention, demonstrating practical viability for production recommendation systems requiring multi-vendor deployment.