Developed a configurable parallel residual computation feature for the GPT-NeoX model within the ml-explore/mlx-lm repository, enabling users to toggle between parallel and sequential processing of attention and feedforward network paths. This addition allows for tailored optimization of performance and memory usage depending on hardware constraints, supporting more scalable deployment scenarios. The implementation leveraged deep learning and machine learning principles using PyTorch and Python, focusing on flexibility and efficiency in model execution. The work involved collaborative code contribution and introduced a new setting that empowers users to fine-tune model behavior for diverse workloads without altering core model architecture or logic.