Worked on the ping1jing2/sglang repository to address a critical VRAM memory leak in the overlap scheduling component. Focused on GPU programming and memory management using Python, the solution involved releasing GPU tensors that were no longer needed and clearing references to large tensors to prevent their persistence across iterations, particularly when structured output was used. This approach stabilized memory usage, improved reliability under heavy workloads, and reduced the risk of out-of-memory errors during production runs. The work demonstrated careful attention to resource optimization and maintainability, aligning with broader goals to enhance performance and stability in GPU-accelerated Python applications.