Worked on stabilizing device mapping and CPU affinity for NUMA systems in the NVIDIA/TensorRT-LLM repository, focusing on resolving a critical bug rather than adding new features. Addressed a mismatch between logical CUDA device indices and physical NVML device indices, ensuring that the correct physical device is referenced when CUDA_VISIBLE_DEVICES is set. This fix improved the accuracy of CPU affinity assignments for CUDA devices, enhancing deployment stability on multi-NUMA systems. The work involved deep understanding of CUDA, GPU programming, and Python, and demonstrated careful attention to low-level system interactions to ensure correctness in complex hardware environments.