Worked on NVIDIA/TensorRT-LLM to address vocabulary-size mismatches during VILA and NVILA model loading, focusing on improving deployment reliability across varied tokenizers. Developed helper utilities in Python and PyTorch to dynamically resize token embeddings and the language model head, integrating these adjustments directly into the model loading workflow. Enhanced the unit testing setup to streamline validation and reduce preparation time for experiments involving different vocabularies. This targeted bug fix reduced runtime errors and ensured consistent model initialization, contributing to more robust deep learning model configuration and loading processes. The work demonstrated careful attention to detail and practical problem-solving in model deployment.