Worked on stability improvements for the ROCm/tensorflow-upstream repository, focusing on enhancing fault tolerance in distributed split-task workflows. Addressed a critical issue by adding defensive validation to the split_task_done_callback, ensuring that silent failures in split tasks no longer result in crashes from uninitialized output tensors. Applied C++ skills and defensive programming techniques to improve the reliability of the BatchResourceBase::SplitInputTask path, reducing crash scenarios in high-throughput environments. The targeted fix was implemented with minimal code changes, aligning with upstream maintenance practices and emphasizing long-term maintainability. This work improved runtime reliability and reduced the incident surface for end users.