Worked on the ROCm/aiter repository to deliver two major features focused on performance and scalability. Developed a CUDA All-Reduce gating threshold optimization for Tensor Parallel, updating the logic to incorporate world size into byte-limit calculations and prevent suboptimal scaling. Introduced a parallel MLA metadata generation path, leveraging C++ and CUDA to accelerate decode workflows by distributing work across multiple warps. Ensured correctness and maintainability through comprehensive testing and cross-team collaboration. Improvements included enhanced test infrastructure and formatting for metadata tests. The work demonstrated depth in distributed systems, parallel computing, and performance optimization, addressing both throughput and reliability in production workflows.