Worked on the kvcache-ai/sglang repository to enhance the Longcat Flash model’s reliability and observability using Python, PyTorch, and deep learning techniques. Addressed CUDA graph instability by introducing a function for QKV latent variable preparation, which reduced execution-time errors and improved inference performance. Developed a feature to capture auxiliary hidden states from specified layers, enabling deeper diagnostics and interpretability of intermediate representations. These targeted improvements increased the throughput and maintainability of the Longcat Flash pipeline, streamlining debugging and supporting data-driven optimizations. The work demonstrated a focused approach to stabilizing complex machine learning systems and improving their operational transparency.