Worked on enhancing data integrity and reliability in the distributed NCCL AllGather path within the ROCm/FBGEMM repository. Focused on defensive programming by introducing a dtype check to prevent silent data corruption caused by mismatched tensor data types. Developed and integrated a dedicated test to ensure that exceptions are raised when source and destination dtypes do not match, thereby reinforcing the robustness of the AllGather code path. Utilized C++ and Python alongside expertise in distributed systems, GPU computing, and PyTorch to improve test coverage and reliability. The work addressed a critical bug, contributing to safer and more predictable distributed operations.