Worked on the huggingface/optimum-habana repository to enhance inference performance for transformer models on Habana Gaudi accelerators. Developed and integrated a FusedSDPA kernel into the Bert self-attention mechanism, replacing the standard scaled dot-product attention within BertSdpaSelfAttention.forward. This targeted optimization focused on non-training scenarios, aiming to deliver improved throughput and reduced latency specific to Habana hardware. The implementation was tracked with a dedicated commit for clear traceability. Leveraged deep learning expertise, performance optimization techniques, and Python to ensure the solution aligned with hardware capabilities. The work demonstrated a focused approach to hardware-aware model optimization within a production codebase.