CV

Ziqing Li

Ph.D. Student in Computer Architecture

d202381502@hust.edu.cn
Wuhan / London, ,

Summary

Ph.D. student at Huazhong University of Science and Technology working on memory systems for long-context LLM inference, with a focus on KV-cache management, storage systems, GPU memory systems, and MLSys runtime support.

Education

  • Ph.D. student in Computer Architecture
    present
    Huazhong University of Science and Technology
  • B.S. in Computer Science and Technology
    2023
    Beijing Jiaotong University

Work Experience

  • Software Engineering Intern, Data Platform / OLAP Engine Team
    2023-11 - 2024-02
    Xiaomi Group
    Mentor: Yaodong Zhang, Senior Software Engineer and Apache Spark committer.
    • Implemented column-level encryption for Parquet files in Apache Spark and Trino.
    • Helped protect sensitive columns independently in internal OLAP pipelines.

Skills

Programming

  • C/C++
  • Python
  • Java
  • Scala
  • MATLAB

Systems

  • Linux systems programming
  • Git
  • GDB
  • SPDK
  • GPUDirect Storage
  • HDFS

Research Areas

  • Long-context LLM inference
  • KV-cache management
  • Storage systems
  • GPU memory systems
  • CXL and memory disaggregation
  • MLSys

Publications

  • OrchKvCache: Orchestrating LLM KV-Cache across GPU-DRAM-SSD with Attention-Aware Hotness Scheduling
    2026
    Targeting SC 2026
    OrchKvCache manuscript / artifact in progress.
  • HALO: Algebraically Identity-Preserving KV Tiering for Memory-Bounded Long-Context Inference
    2026
    Targeting EMNLP 2026 (ARR)
    HALO manuscript / artifact in progress.
  • SEER: Schedulable KV-Cache Eviction with a Probabilistic Sizing Bound for Real-Time LLM Decoding
    2026
    Targeting IEEE RTSS 2026
    SEER manuscript / artifact in progress.
  • Mining Attention Dynamics: Auditing KV-Saliency Prediction in LLMs
    2026
    Targeting IEEE ICDM 2026
    XQP / KVSalienceBench manuscript / artifact in progress.
  • Direction Is Not a Knob: Cross-GPU KV-Cache Handoff Contention on NVLink
    2026
    Targeting IEEE ICCD 2026
    PeerKV manuscript / artifact in progress.

Portfolio

  • OrchKvCache
    2026
    Tiered kv-cache substrate
    Tiered KV-cache management across GPU HBM, host DRAM, NVM, and SSD.
  • HALO
    2026
    Semantic-safe offloading
    Identity-preserving KV offloading through chunked attention and log-sum-exp merge.
  • SEER
    2026
    Slo-aware scheduling
    SLO-aware KV-cache eviction and prefetch under P99 / P99.9 latency targets.
  • XQP
    2026
    Saliency measurement
    Measurement-driven KV-block saliency prediction from LLM attention dynamics.
  • PeerKV
    2026
    Cross-gpu kv handoff
    Cross-GPU KV-cache handoff contention on NVLink.

Interests

  • Memory systems for long-context LLM inference
    KV-cache management, offloading, compression, SLO-aware scheduling, cross-GPU KV handoff
  • Storage systems and heterogeneous IO
    GPU-SSD data paths, GPU interconnects, NVM/SSD tiering, CXL, memory disaggregation