AI Infrastructure & Efficient Inference
Training and inference systems, workload modeling, memory optimization, and performance analysis for generative models.
Computational Mathematics · AI Infrastructure
Guorui Zhu
Ph.D. Candidate at Fudan University
I work at the intersection of AI infrastructure, efficient model inference, high-performance computing, numerical algorithms, and quantum computing. My current work focuses on making demanding AI and scientific workloads faster, more scalable, and easier to deploy.
- Current
- Research Intern, Huawei ICT BG · AI Infrastructure
- Education
- School of Mathematical Sciences, Fudan University
- Advisor
- Prof. Yingzhou Li
01
Research Focus
Systems and algorithms for compute-intensive problems.
High-Performance Computing
Distributed training and scientific computing with Ascend, C++, MPI, CUDA, Slurm, and multi-node clusters.
Numerical Algorithms
Fast solvers, parallel PDE methods, numerical linear algebra, and scientific software.
Quantum Computing
Quantum algorithms for basis transformation, orbital optimization, and excited-state calculation.
02
Experience & Research
A record of engineering, research, operations, and quantitative work.
Video Generation Inference Optimization
Profiled image and video generation inference pipelines and modeled their workload behavior, with practical understanding of Diffusion, Flow Matching, and Score Matching. Independently designed and implemented MiniMax-H3 SeqAttn and Stream-Attn, combining sequence chunking, streaming execution, and activation offloading to reduce memory pressure while preserving output quality and inference speed on consumer GPUs.
Kimi K3 Training Systems & Activation Offloading
Independently built the full training software stack for Ascend 910C and 950 systems, including the operating system, drivers, runtime, and distributed-training dependencies. On a single-node A3 system with 8 Ascend 910C cards (16 dies, 8 × 2 × 64 GB), brought up the Kimi K3 pruned-model pretraining workflow with FSDP2. Evaluated activation offloading on an A5/Ascend 950 setup with high-speed interconnects and an external memory box, reading the recomputation and offloading paths in MindSpeed MM and selecting activations at operator granularity. Used MS Insight to inspect forward hotspots and arithmetic intensity, and analyzed HCCS and DMA behavior under compute, I/O, and communication overlap, achieving about 15% end-to-end speedup.
Parallel PDE Solver
Implemented a parallel Poisson–Boltzmann solver in C++ and MPI using the HIF-SI library. Evaluated scaling on the Tianhe-3 supercomputer and commercial cloud platforms, with optimization centered on communication behavior and workload balance.
High-Frequency Futures Quantitative Research
Used statistical and machine learning methods to mine alpha factors from tick-level limit-order-book snapshot data in the Chinese commodity futures market.
Macro Placement Algorithms
Developed and optimized 2D macro placement algorithms for an internal physical-design tool, with code merged into Huawei's internal repository. Also improved Open3DBench algorithms for 3D placement quality and runtime efficiency.
HPC Cluster Administration
Built and maintain a multi-node research cluster, including networking, storage, scientific software, user environments, and Slurm scheduling. Integrated Grafana and agent-assisted workflows for monitoring, anomaly identification, and job analysis.
03
Publications
Quantum algorithms for electronic-structure calculations.
arXiv preprint · 2026
Structure-Aware Compilation for Scalable Neutral-Atom Quantum Computing
D. Dong, F. Zou, H. Chen, Guorui Zhu, and Y. Li. arXiv:2607.01787.
arXivnpj Quantum Information · 2025
Quantum Circuit for Non-Unitary Linear Transformation of Basis Sets
Guorui Zhu, J. Bierman, J. Lu, and Y. Li. 11, 198 (2025).
DOIarXiv preprint · 2025
State-Specific Orbital Optimization for Enhanced Excited-States Calculation on Quantum Computers
Guorui Zhu, J. Bierman, J. Lu, and Y. Li. arXiv:2510.13544.
arXiv04
Recent Writing
Notes on model systems, parallel computing, and numerical methods.
Contact