Worked on the sgl-project/sglang repository to enhance the stability of GPU memory reporting within NVIDIA MIG container environments. Addressed a crash scenario by implementing a fallback mechanism using Python, where torch.cuda.mem_get_info() is called if nvidia-smi fails to retrieve GPU memory capacity. This approach ensured that memory information remained accessible, preventing application failures and reducing downtime for containerized GPU workloads. The work focused on bug fixing and GPU computing, specifically targeting the reliability of memory information retrieval in complex deployment scenarios. The solution improved runtime stability for sglang users operating in GPU-enabled containers with NVIDIA MIG configurations.