Worked on NVIDIA/TransformerEngine to enhance FlashAttention 2 compatibility in PyTorch by addressing mismatches between query and value head dimensions in multi-head attention. Developed utilities in Python for padding and trimming the value tensor, ensuring robust handling when value dimensions exceed those of the query or key. This approach improved the reliability and efficiency of attention-heavy deep learning workloads, reducing edge-case failures and supporting safer production deployment. The contribution focused on tensor operations and deep learning best practices, establishing a more stable foundation for PyTorch models leveraging FlashAttention 2 within the NVIDIA/TransformerEngine repository, and facilitating broader adoption in machine learning applications.