CV
Education, experience, awards, skills, and patents.
Professional Summary
Machine learning systems researcher focusing on efficient training and inference for large-scale language models. My work studies system-level optimizations including MFU improvement, communication–computation overlap, MoE inference optimization, and hardware-efficient execution for large-scale AI infrastructure.
Currently AI Infrastructure / Training Systems Researcher at AIGCode (蔻町科技). Previously Senior Algorithm Researcher at Huawei Cloud Architecture Innovation Lab.
Publications
-
2026 Zeta: Dual Whitening for Matrix Optimization via Coordinate-Adaptive Preconditioning
arXiv:2606.14187, Preprint
Kaiwen Chen*, Shuhai Zhang*, Zimo Liu, Linxiao Li, Ying Sun, Yuchen Li, Yifan Zhang, Bo Han, Mingkui Tan†, Qiuwu Chen†
-
2026 Deterministic Component Mining for Multi-framework UI2Code Generation
International Conference on Machine Learning (ICML 2026), Poster
Zixiong Yang*, Linxiao Li*, Jiaye Lin, Binrui Wu, Xiaoyu Kang, Jiechao Gao†
-
2026 EcoThink: A Green Adaptive Inference Framework for Sustainable and Accessible Agents
The Web Conference (WWW 2026), Oral
Linxiao Li*, Zhixiang Lu*†
Public Talks
-
Jun 2026 Communication-Aware Scheduling Optimization for Large-Scale MoE Model Training on Ascend NPUs
Frontiers Series · Guangming Laboratory, Academic Seminar
Academic seminar on communication-aware scheduling optimization for large-scale MoE model training on Ascend NPUs.
-
Apr 2026 DeepSeek-V4 Mechanism Deep Dive
Internal Technical Sharing · AIGCode, Technical Talk
Company technical interpretation and discussion of DeepSeek-V4 mechanism details.
-
Dec 2025 Research Methodology, Internship Planning, and Career Development
Graduate Career Workshop · Guangming Laboratory, Invited Talk
Invited talk for graduate students on research methodology, internship planning, and career development.
Industry Research Experience
-
2025 - Shanghai, China
AI Infrastructure / Training Systems Researcher
AIGCode (蔻町科技)
Leading LLM pre-training optimization and training infrastructure. Currently leading pre-training of a 300B+ parameter LLM from scratch.
- Achieved 5x single-operator speedup (MoE Combine + FFN), improving overall MFU by 10% and reducing training cost by >10%
- Tech stack: Triton, TileLang, PyPTO with Tile paradigm programming
- Profiling tools for end-to-end optimization. Building next-generation AI Infra and model architectures.
-
2023 - 2025 Shenzhen, China
Senior Algorithm Researcher Leader / PL (Huawei Fellow Team)
Huawei Technologies Co., Ltd.
Core member of Huawei Cloud Architecture Innovation Lab (Cloud Lab), under CTO Fellow Jiongjiong Gu (Chief Architect). Led a 10+ person joint team with the Guangming Laboratory on intelligent resource scheduling. All annual performance ratings A.
- MoE Memory Offloading: Expert offloading for MoE LLM inference on Ascend NPUs. TTFT <3s, TPOT <100ms. Supports BF16/W8A8 (DeepSeek V3). (HDC 2025, ~50 min)
- NPU User-Space Virtualization: Multi-tenant NPU sharing with <3% overhead, utilization 10%→30%+. (HDC 2025, ~50 min)
- Huawei Cloud Flexus FRCP: iTransformer capacity prediction for X/L instances, deployed at scale.
- QoS MoE Model: MAE 3%, classification accuracy 99.995%. Key support for Agricultural Bank of China HCS project.
-
2021 - 2023 Chengdu, China
Software Development Leader
Huawei Technologies Co., Ltd.
Led 5-person team as Committer. Core developer on Data Workshop distributed messaging system. (incl. internship from Sep 2021)
- Built on Pulsar distributed messaging middleware with self-developed storage engine
- Zero P3 incidents over 6 months; supported major customers at scale
Selected Research Projects
-
Ascend NPU User-Space Virtualization
Designed a user-space virtualization framework for Ascend NPUs to improve hardware utilization in large-scale AI training.
- Achieved <3% performance overhead while increasing NPU utilization from 10% to 30%
-
MoE Inference Memory Offloading
Designed a memory offloading framework for large-scale Mixture-of-Experts inference to reduce GPU memory pressure.
- Achieved <3% overhead while maintaining TTFT <3s and TPOT <100ms in production systems
- Huawei Cloud High-Potential Patent: A MoE Model-Oriented Inference Optimization Method
-
Both projects were deployed in Huawei Cloud and announced at Huawei HDC 2025 (~50 min mark).
Education
-
2018 - 2021 Sydney, Australia
The University of Sydney
Computer Science
- GRE: 334 (Q170, V164, AW 5.5)
- Advisor: Prof. Wei (Wilson) Li
-
2021 - 2021 Shanghai, China
-
2020 - 2021 Shanghai, China
Open Source
-
Codex AutoResearch
Creator · 1,000+ Stars — GitHub. A self-directed iterative research system for Codex that continuously cycles through modify, verify, retain/discard, and repeat.
-
PyPTO
Contributor — GitCode. High-performance programming framework for AI accelerators using tile-based programming (PTO paradigm), compiling tensor graphs to hardware instructions via multi-level IR.
-
PyPTO-AutoGrad
Creator — GitCode. Source-code level automatic differentiation tool for PyPTO kernels, generating readable and editable backward kernels from forward definitions.
-
Modern LLM Notebook
Contributor · 80+ Stars — GitHub, Website. A hands-on PyTorch curriculum for building modern LLM systems from scratch, covering tokenization, attention, MoE, inference, evaluation, and distillation.
-
Pie
Contributor · 170+ Stars — GitHub, Website. A programmable LLM serving system for custom inference logic, stateful agents, and serving-side optimization.
Awards
-
2025 Huawei ICT Software Competition — 14th / 5200+
Huawei
Top 3 in Huawei Cloud product line, 1st in Chengdu Research Institute.
-
2024 Golden Cloud Award (金代码)
Huawei Cloud
Top 2% recognition for outstanding code contributions.
-
2024 HCS Cloud Summit Star (云巅之星)
Huawei Cloud
Top 5% internal technical honor for outstanding contributions.
-
2023 -
2023 Huawei Trustworthy Professional Certification
Huawei
Passed Huawei Trustworthy Professional Level Certification.
-
2022 -
2020 Vice Chancellor's Global Mobility Scholarship
The University of Sydney
Awarded on the basis of academic merit and exchange destination.
Patents
-
2024 A MoE Model-Oriented Inference Optimization Method
Huawei Cloud High-Potential Patent
Invention patent for MoE large model inference memory offloading optimization.
-
2024 A Latency-Aware Serverless Request Scheduling Strategy
Invention patent for intelligent serverless request scheduling.
Academic Service
-
Teaching
Teaching Assistant, The University of Sydney — Operating Systems, Algorithm Design
-
Reviewing
Reviewer, NeurIPS 2026