During a two-month period, Ziang Li developed and optimized deep learning infrastructure across the kvcache-ai/sglang, yhyang201/sglang, and flashinfer-ai/flashinfer repositories. He engineered a new matrix multiplication kernel and FP32 precision loss mitigation for large-batch model projection, improving stability and performance using CUDA and C++. In yhyang201/sglang, he introduced a CUDA graph-friendly weight binding utility to enhance parameter management during graph reuse. For flashinfer-ai/flashinfer, he implemented MXFP8 quantization pathways for MoE reinforcement learning, including activation-scaling and kernel optimizations. Li’s work demonstrated depth in GPU programming, quantization, and performance optimization, addressing complex challenges in model serving and training.