GPU Optimization for All-Pairs Shortest Path

Individual competition project · Aug–Sep 2025 · CCF-TCARCH National Champion

• Designed a memory-aware 3-stage blocked Floyd–Warshall pipeline to reduce bandwidth pressure in dense APSP computation, using shared memory to improve data locality.
• Optimized kernel execution through fusion, dual-stream scheduling, double buffering, and reduced synchronization overhead. Further compressed host-to-device transfer cost from O(V²) to O(E) via custom I/O and device-side graph reconstruction.
• Ranked 1st nationally as an individual competitor, reducing end-to-end runtime by 52.9% on the official hidden datasets (11.50s → 5.42s), equivalent to a 2.12× speedup through profiling-driven optimization and systematic performance tuning.

On-Device AIOS on RK3588 and StarryOS

Team competition project · May–Aug 2026 · Team captain; perception-to-control integration and OS performance optimization.

• Integrated UVC image acquisition, RKNN/NPU inference, object detection, and robotic-arm control into a tennis-ball retrieval robot running on RK3588 and StarryOS.
• Diagnosed RKNPU execution and SMP scheduling, fixing hot-path logging stalls and tasks remaining on a single core. In the project measurements, rknn_run latency fell from 817 to 211 ms.
• Improved the four-thread speedup from 0.96× to 3.45× in the project tests; the team received a National Third Prize in the 2026 Operating System Design Contest.

LLM Inference Serving on Domestic DCUs

Individual competition project · Jul–Aug 2026 · Kernel and serving optimization.

• Profiled Qwen3.5-27B with vLLM for long-context inference on domestic DCUs, separating prefill/decode bottlenecks and analyzing attention and paged KV-cache paths.
• Optimized compute and memory access with FlashAttention, Triton GEMV, and CUDA Graph, and adapted TP=2 tensor parallelism and Custom AllReduce for two-device serving.
• Completed optimization from individual kernels to the serving pipeline, placing 12th nationally and receiving a National Third Prize in the 2026 Xiandao Cup.

Diagnosing and Mitigating Noun-Driven Routing Bias in Multimodal Mixture-of-Experts

Research project · Apr–Jun 2026 · Routing probe design and experimental evaluation.

• Developed CMRD (Controlled MoE Routing Decomposition), a training-free probe using text anchors and controlled pre-router interventions to examine noun-driven visual routing shifts.
• Used same-question controls to approximately separate broad-semantic effects from a queried-specific residual, identifying a shared layer-10 routing hotspot in MoE-LLaVA-Qwen.
• Evaluated Noun-Blind Routing and Two-Stage Verification for object-existence hallucination. On POPE-9000, Two-Stage Verification reduced false positives by about 60% and improved accuracy from 81.7% to 88.2%, while increasing false negatives.

Adaptive Weight Scheduling for Image Diffusion Models (K-LoRA++)

Competition research project · Apr–Oct 2025 · Schedule formulation, evaluation, and demo development; National Third Prize in Challenge Cup.

• Extended the original K-LoRA framework for training-free subject/style LoRA fusion by mathematically formulating and evaluating 8 distinct stage-aware scaling schedules across diffusion timesteps.
• Compared the schedules through first-order derivative analysis and cross-domain experiments; a constant-derivative linear schedule gave the best content/style balance among the evaluated schedules in the tested settings, with less structural distortion and visual jitter.
• 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 Framework for Time Series Forecasting

Course project · Sep–Dec 2024 · Framework migration and model extension.

• Reimplemented and extended SparseTSF from PyTorch to MindSpore, including custom API mappings and reconstruction of the training/evaluation pipeline for cross-framework deployment.
• Designed an FFT-based learnable filtering module to replace local sliding-window convolutions, aiming to better capture long-range periodic structure under a lightweight parameter budget.
• Conducted cross-framework profiling and forecasting evaluations, observing improved long-horizon performance (e.g., at 336 and 720 steps on ETTh1/ETTh2) relative to the baseline in the evaluated settings.

RaceRadar: An End-to-End iOS System for Aggregating and Prioritizing Competition Opportunities

Independent engineering project · Nov 2025–present · App Store release in preparation

• Built a fault-tolerant, multi-source ingestion pipeline that continuously crawls, cleans, deduplicates, and ranks competition announcements, publishing a unified JSON feed through scheduled GitHub Actions workflows.
• Developed a native iOS app in Swift/Xcode with card-based discovery, detailed opportunity pages, deadline-aware prioritization, source/category tagging, and shareable competition posters.
• Designed the system around real student workflows, translating automated data ingestion and ranking into a deployable user-facing information system for discovering higher-quality and time-sensitive opportunities.