Worked on enhancing PyTorch’s ROCm support by expanding and stabilizing test coverage for distributed data parallelism and tensor operations in the pytorch/pytorch repository. Focused on unskipping and validating ROCm-specific tests, particularly for the DDP backward pass, to ensure reliable cross-architecture functionality and earlier regression detection. Leveraged Python, PyTorch, and distributed systems expertise to address test regressions and reduce flaky failures in continuous integration pipelines. Collaborated with maintainers to deliver safer ROCm-enabled distributed training releases, emphasizing robust CI automation and comprehensive testing. The work improved the reliability and maintainability of ROCm features within PyTorch’s core testing suite.