Worked on a performance refactor for the flashinfer-ai/flashinfer repository, focusing on optimizing attention forward operations in C++ and CUDA. Developed a thread-safe, one-time caching mechanism for CUDA device 0 properties within the CudaDevice (fmha_v2) component, eliminating repeated calls to cudaGetDeviceProperties and reducing initialization overhead by approximately 1.7 milliseconds per fmha_v2_run. This approach improved both startup latency and throughput for attention forward workloads. The implementation used C++11 static initialization to ensure thread safety and maintainability, and included updated tests to validate the new caching behavior while preserving existing functionality, laying groundwork for future per-field caching enhancements.