Developed and integrated a 6-bit quantization feature for Llama models within the pytorch/ao repository, focusing on efficient data packing and unpacking to optimize storage and throughput in torchchat workloads. Leveraged C++ and low-level programming techniques to implement quantization utilities and corresponding APIs, enabling more cost-effective inference and training pipelines for large-scale datasets. Updated the benchmarking suite to rigorously evaluate the performance and throughput impact of the new quantization method. Demonstrated expertise in data structures, performance optimization, and cross-repository collaboration, aligning data handling improvements to support scalable, high-throughput machine learning workflows without introducing major bug fixes during the period.