Developed initial Eagle3 support for Deepseek-like models within the kvcache-ai/sglang repository, focusing on enhancing model interpretability and debugging capabilities. Implemented a configurable mechanism in Python and PyTorch to capture auxiliary hidden states at selectable layers, allowing users to specify which internal representations to monitor during processing. This approach enabled more transparent analysis and faster debugging by integrating seamlessly with existing Deepseek-like pipelines. Comprehensive documentation and commit tracing were provided to facilitate future extensions and audits. The work demonstrated depth in deep learning model development, emphasizing maintainability and business value through improved transparency and flexibility in model analysis workflows.