GPU Optimization for All-Pairs Shortest Path (Solo Undergraduate Participant; National Champion outperforming top-tier graduate teams in CCF-TCARCH)

• Architected a memory-aware 3-stage blocked Floyd-Warshall pipeline, heavily leveraging shared memory to bypass global memory bandwidth limits.
• Engineered aggressive kernel optimizations (fusion, dual-stream, double buffering, barrier elimination) and resolved I/O bottlenecks by compressing H2D transfers from O(V²) to O(E) via custom fast I/O and device-side graph reconstruction.
• Achieved a 52%+ end-to-end speedup (reducing total runtime from 11.50s to 5.42s on the official hidden datasets) through systematic, rocprof-guided micro-architectural profiling and hyperparameter tuning.

Research on Adaptive Weight Selection for Image Diffusion Models (National 3rd Prize in Challenge Cup, K-LoRA++)

• Extended the training-free object-style fusion framework (CVPR 2025 K-LoRA) by mathematically formulating and evaluating 8 distinct stage-aware scaling schedules across diffusion timesteps.
• Demonstrated through first-order derivative analysis and cross-domain experiments that a constant-derivative linear schedule optimally balances content fidelity and style consistency, effectively preventing structural distortion and visual jitter during the fusion process.
• Facilitated the end-to-end development of the project’s interactive demo and drove its practical deployment for downstream generative applications.

SparseTSF-FFT: A Frequency-Domain Optimized Framework for Time Series Forecasting

• Spearheaded the PyTorch-to-MindSpore migration of SparseTSF. Engineered custom API mappings to resolve missing operators and fully reconstructed the end-to-end training and evaluation pipelines for seamless cross-framework deployment.
• Architected an FFT-based learnable filtering module to replace local sliding-window convolutions. This frequency-domain approach directly captures global periodicities and long-term trends while strictly preserving the model’s ultra-lightweight sub-1k parameter budget.
• Executed rigorous cross-framework profiling (MSE vs. runtime) and weight-distribution visualizations. Demonstrated superior long-horizon forecasting performance (e.g., at 336 and 720 steps) on ETTh1 and ETTh2 datasets compared to the baseline.