Worked on the PaddlePaddle/FastDeploy repository to deliver support for the ERNIE-4.5-21B-A3B model on Iluvatar GPUs, focusing on enhancing deployment reliability and hardware compatibility for large-scale enterprise workloads. The approach involved updating installation documentation and example scripts using Markdown and Shell, as well as refining deployment flows and model paths to ensure seamless onboarding. Python was used to implement changes in deployment artifacts and update expected output logs, reflecting the new model’s performance characteristics. This work improved the visibility and reliability of model deployment for enterprise customers leveraging GPU computing and large language models in production environments.