Worked on the JustinTong0323/sglang repository to address a critical issue with Gemma model loading in bitsandbytes formats, focusing on both 8-bit and 4-bit quantization. The solution involved refining weight name checks, introducing helper methods to identify quantized weight files, and updating the mapping and yielding logic to ensure compatibility with the bitsandbytes library. This Python-based work improved the reliability of machine learning model deployment by reducing load-time errors and expanding support for quantized models. The approach demonstrated depth in model loading and quantization, with careful attention to future extensibility and enhanced test coverage for robust engineering outcomes.