Yuntian Deng
Academic Performance
2024.9-Now Shanghai Jiao Tong University
Electrical and Computer Engineering
Relevant Coursework: C++, Data Structures, Honors Calculus, Probability Theory, Computer Architecture
Technical Skills:
- Languages: C/C++, Python (NumPy, PyTorch, OpenCV), MATLAB
- Tools: Linux, Slurm, Git, LaTeX, OpenCode, Claude Code, ROS
Experiences
2024.10-Now SJTU Supercomputer Team Xflops Core Member
- Attend global supercomputing competitions, including ASC and SCC.
- Optimize large-scale open-source software by reconstructing operators and tuning multi-thread scheduling.
- Organized the 1st internal high-performance computing competition, “Hello HPC” at SJTU, reaching nearly 1,000 participants.
- Contribute to and maintain the open-source HPC guide https://xflops.sjtu.edu.cn/hpc-start-guide/.
- Independently deployed the DeepSeek-R1-671B model on the SJTU TPU Cluster.
2024.9-Now Research in Embodied AI Algorithm Project Participant
- Filed a patent application for a new learning method based on probabilistic experience retrospection.
- Completed a short-term internship at Galbot. Conducted research on Visual Language Action (VLA) models and World Model Actions (WMA) to improve robot arm performance in complex real-world scenarios.
- Contributed to a paper submitted to the Robotics: Science and Systems (RSS26) conference.
Projects
2025.3-2025.5 Optimization on AlphaFold3 Model, ASC25
- Participated in the optimization of the AlphaFold3 model, designing a distributed inference framework based on MPI. Leveraging profiling results, implemented high-frequency, low-level operators using the LIBXSMM library. Achieved up to 50x inference speedup.
2025.12-2026.5 Optimization on World Model Inference, ASC26
- UnifoLM-WMA-0: Applied quantization, full-graph compilation, dynamic pipeline simplification, multi-GPU pipeline/communication design, and synchronization tuning, achieving a 8x speedup under precision constraints.
- LeWM: Reduced H2D transfers via data load balancing. Removed CPU bottlenecks with an asynchronous CPU-GPU scheduling pipeline, improving both CPU and GPU utilization. Optimized relative-position KV cache for faster LeWM inference; rewrote small-sequence Transformer hotspots in Triton and used Hipblast automatic operator epilogue fusion, achieving a 200x speedup.
- Built an apptainer based agent coding orchestra pipeline for stable and long-term auto-optimization in containerized environments. Improve the context management, subagent calling mechanism.
Awards & Honors
- World Champion Player, ASC Student Supercomputing Challenge (2025)
- First Prize, ASC Student Supercomputing Challenge (2026)
- Best Presentation Prize, ASC Student Supercomputing Challenge (2026)
- National Scholarship and John Wu & Jane Sun Excellence Scholarship, SJTU (2025)
- Silver Medal, University Physics Competition (2024)
- National 1st Prize, 37th China Adolescence Science & Innovation Contest (2023)
Miscellaneous
- Assistant manager in the International Program Office. Supported cross-cultural activities, welcomed exchange students, and helped them better understand Chinese culture, adapt to campus life, and local community resources.
- Passionate about fine art photography, with a focus on urban scenes, campus life, and visual storytelling. Regularly served as a photographer for school events, capturing memorable moments and highlights.
- Enjoy football, badminton, table tennis, and other team or racket sports. Actively participate in various sports competitions.
