Worked on the ROCm/flash-attention repository to enhance SM100 attention by enabling support for non-contiguous input tensors, addressing a key limitation in handling diverse memory layouts. Refactored normalization utilities into a shared module to improve code maintainability and promote reuse across the backend. Leveraged CUDA, PyTorch, and Python to update the forward kernel, ensuring it reads actual input strides rather than forcing contiguous memory, which eliminated unnecessary memory copies and improved performance. Added regression tests to validate correctness with non-contiguous inputs, resulting in more robust and flexible attention mechanisms suitable for real-world workloads with varied tensor configurations.