CV
Curriculum Vitae of Yansheng Mao.
General Information
| Full Name | Yansheng Mao |
| ysmao_lumen@outlook.com | |
| GitHub | LuckyGlass (https://github.com/LuckyGlass) |
Education
-
2026 - Present Ph.D. Student, School of Intelligence Science and Technology
Peking University - Advised by Prof. Muhan Zhang in Μμ Lab (https://mulabpku.com/).
-
2022 - 2026 Undergraduate, School of Electronics Engineering and Computer Science
Peking University - Member of the Zhi Class, a special class for excellent students majoring in artificial intelligence.
- Recipient of the Zhi Class scholarship in 2024 and 2025.
Experience
-
2026 - Present Training Infrastructure Intern (Founding Team Member)
Norizon (无涯智远) - Core member of the training infrastructure team, responsible for building and maintaining the company's large-model training infrastructure; primary author of all of the company's training frameworks.
- Designed and built three generations of internal training frameworks — a PyTorch/FSDP2 stack, a Megatron-Core rewrite scaling to dense and MoE models up to ~35B parameters, and a from-scratch PyTorch-native framework; the latter two were built solo.
- Implemented FSDP2/HSDP, tensor, sequence (Ulysses and context), and MoE expert parallelism, plus custom Triton/DeepGEMM grouped kernels and packed varlen training with ~66% peak-memory reduction.