Worked on improving build robustness for the pytorch/pytorch repository by addressing a recurring issue in the ROCm build process. Using Python scripting and build system expertise, implemented checks in build_amd.py to handle missing MSLK tuning cache headers, specifically tuning_cache.cuh and tuning_cache_hip.cuh. This solution prevents FileNotFoundError exceptions when submodules are not fully initialized, ensuring that builds do not fail unexpectedly in varied environments. The change reduced flaky builds and improved reliability across continuous integration pipelines. The work focused on defensive programming and cross-environment compatibility, contributing to a more stable and predictable build process for PyTorch’s ROCm support.