Worked on the HabanaAI/vllm-hpu-extension repository to enhance softmax stability across diverse hardware platforms. Addressed a hardware limitation by implementing a defensive fallback pathway for cases where index_reduce is unsupported, ensuring softmax operations remain correct and performant. The solution leveraged Python and low-level programming techniques, introducing a mechanism that utilizes wsum_head_amax as an alternative computation path. This approach preserved feature parity and reduced potential downtime, minimizing support overhead for hardware-specific issues. The work demonstrated a focus on hardware acceleration and performance optimization, with clear, traceable changes that maintained the reliability of core functionality in heterogeneous computing environments.