Worked on enhancing the huggingface/optimum-habana repository by integrating PT2E quantization into the main text generation workflow, focusing on performance improvements for Habana accelerators. The approach involved adding configurable arguments to manage quantization, as well as implementing streamlined processes for preparing, saving, and loading quantized models. This enabled more efficient deployment of deep learning models with reduced memory usage and faster inference, while maintaining developer usability. The work was carried out using Python and PyTorch, leveraging expertise in deep learning, HPU optimization, and model quantization to address the specific needs of Habana-backed text generation tasks.