Worked on stabilizing the attention pathway in the AI-Hypercomputer/maxtext repository, focusing on improving the reliability of attention computations for deep learning models. Addressed a numerical instability issue by fixing the scale parameter in the AttentionOp class to a constant value, which enhanced both training and inference stability on cudnn_flash_jax. The solution was implemented in Python, leveraging expertise in NVIDIA CUDA and PyTorch to ensure compatibility with existing machine learning workflows. This targeted bug fix reduced risk and complexity in the attention mechanism, laying groundwork for future optimization and making attention-related code more maintainable and easier to trace for subsequent development.