Santosh Hiremath developed flash attention with key-value caching as a PyTorch custom operation for the ROCm/aiter repository, targeting deep learning workloads on AMD hardware. He registered the op using fake tensors to enable HIPGraph integration, and vectorized the cache update logic to improve performance and compatibility with hipgraph-based execution. Santosh removed .item() calls to support manual graph capture, ensuring the feature aligned with mainline development for future stability. He implemented comprehensive unit tests to validate the new functionality and applied code quality improvements, including formatting and comment cleanup, leveraging Python, CUDA, and PyTorch throughout the development process.