Worked on the Lightning-AI/litgpt repository to deliver support for Gemma 3 multimodal models by enabling the loading of only text weights, thereby improving compatibility with text-only configurations. The approach involved updating the model conversion script in Python to handle Hugging Face Transformers checkpoints, introducing explicit logic for multimodal weight mapping, and adding targeted warnings for users. To ensure reliability, an end-to-end test was implemented to validate correct behavior and prevent regressions. These changes enhanced resource efficiency by reducing memory usage and start-up time, while also improving the maintainability and clarity of the PyTorch-based weight loading code paths.