Ziqing Li
I am Ziqing Li (Chester), a Ph.D. student in Computer Science at the School of Computer Science and Technology, Huazhong University of Science and Technology (HUST), advised by Associate Professor Jianxi Chen. I am a member of the Key Laboratory of Information Storage System, Ministry of Education, led by Professor Dan Feng.
My research is in memory systems for long-context LLM inference. I study how to manage the KV-cache (the key/value tensors cached during Transformer decoding) when contexts become too large for HBM (GPU high-bandwidth memory). My work connects storage systems, computer architecture, MLSys, and real-time serving.
Research question
How can long-context LLM serving remain memory-efficient, latency-predictable, and semantically safe when the KV-cache becomes larger than GPU memory?
I approach this question by treating KV-cache as a first-class managed memory object rather than an opaque tensor buffer.
Research stack
My current projects form a coherent stack for long-context LLM memory systems:
| Layer | Question | System |
|---|---|---|
| Tiered substrate | How should KV blocks move across HBM, DRAM, NVM, and SSD? | OrchKvCache |
| Semantic safety | Can offloading preserve full-attention semantics? | HALO |
| SLO control | Can KV placement satisfy P99 / P99.9 serving targets? | SEER |
| Saliency measurement | Which signals actually predict future important KV blocks? | XQP |
| Hardware-aware memory | What happens when KV-cache handoff crosses GPU interconnects? | PeerKV |
More details are collected on the Systems page.
Technical themes
- Tiered KV-cache management: placing KV blocks across GPU HBM, host DRAM, NVM, and SSD according to access hotness and IO cost.
- Identity-preserving offloading: moving cold KV blocks out of GPU memory while preserving full-attention semantics through chunked attention and log-sum-exp merging.
- SLO-aware scheduling: treating KV-cache eviction and prefetch as a soft real-time scheduling problem under P99 / P99.9 latency targets.
- Attention-dynamics prediction: using measured attention traces to predict future salient KV blocks with small, calibrated models.
- Hardware-aware KV movement: studying how GPU memory systems and interconnects shape KV-cache placement, handoff, and contention.
Broader interests
Beyond long-context LLM inference, I am interested in storage systems, heterogeneous IO paths, CXL and memory disaggregation, GPU memory management, and OS/runtime support for ML workloads.
Contact
I am happy to discuss systems research, long-context LLM inference, storage systems, and research collaboration. Please feel free to email me.