Worked on NVIDIA/TransformerEngine to enhance the stability and numerical correctness of distributed GEMM operations, focusing on the GemmRs implementation. Addressed a bug involving local leading dimensions for transposed matrix operations in distributed settings, ensuring correct alignment across processes. Updated the communication type to operate in the output data type, preserving precision and aligning with reduce-scatter semantics and UserBuffers behavior. These changes improved the reliability and traceability of large-scale transformer workloads. The work was implemented in C++ with CUDA, emphasizing matrix operations and performance optimization, and included clear code provenance for maintainability and future reference within the repository.