Perry Zhang contributed to deep learning infrastructure in the IBM/vllm and ROCm/aiter repositories, focusing on GPU programming and parallel computing with Python and CUDA. He implemented ROCm EPLB support for AMD hardware in IBM/vllm, enabling Expert Parallelism Load Balancing validation alongside CUDA and improving compressed tensor methods. In ROCm/aiter, he fixed kernel string formatting for paged MQA logits, enhancing code clarity and correctness. Perry also expanded GPT-OSS 120B support by adding 5D shuffle layout and fused all-reduce RMSNorm for new hidden sizes, validating precision and scalability. His work demonstrated depth in multi-GPU systems and code maintainability.