Worked on the NVIDIA/TensorRT-LLM repository to enhance debugging observability for Qwen model integration by implementing a hidden states capture capability. This involved adding optional parameters to both QwenDecoderLayer and QwenModel, enabling the capture of intermediate representations during inference without disrupting existing APIs. The approach allowed for more efficient debugging and analysis by making internal model states accessible for inspection. Utilized Python and PyTorch to deliver these changes, demonstrating a focus on deep learning model instrumentation. The work addressed the need for faster iteration and troubleshooting in machine learning workflows, contributing a targeted feature to support model development and analysis.