Qun Yang enhanced the red-hat-data-services/vllm-gaudi repository by addressing a shape-mismatch edge case in RMSNorm, specifically enabling compatibility for 2D input tensors on HPUs. Using Python and deep learning frameworks, Qun implemented a solution that preprocesses 2D inputs by unsqueezing them to 3D for HPUFusedRMSNorm operations, then restores the original shape post-processing. This approach preserved output consistency while broadening hardware and data-input support. The work demonstrated proficiency in custom operations and model optimization, ensuring that RMSNorm behavior aligned with fused operation expectations and reducing the risk of downstream breakages in production machine learning workflows.