Zan Zhang developed advanced attention mechanism features for deep learning inference, focusing on performance optimization and hardware compatibility. In the ROCm/composable_kernel repository, Zan implemented chunked prefill support for FlashAttention within the MHA variable-length kernel, addressing compiler issues and adding sequence-length guards to improve reliability for small query workloads. The work involved C++, CUDA, and kernel development, with comprehensive documentation to ensure maintainability. Subsequently, in red-hat-data-services/vllm-cpu, Zan integrated Aiter chunked prefill into the VLLM framework, optimizing attention performance for AMD hardware using Python and PyTorch. These contributions enhanced inference throughput and reduced latency for dynamic deployments.