Worked on improving autotuning stability in the pytorch/pytorch repository, focusing on backend development and configuration management using Python. Addressed sporadic crashes in Triton-enabled environments by introducing an environment variable that disables Triton device detection, thereby preventing unintended device initialization during autotuning workflows. Implemented a targeted workaround for the mtia double initialization issue within the has_triton logic, directly reducing failure rates in continuous integration and user workloads. These changes enhanced the reliability of model tuning across diverse GPU configurations, enabling more consistent performance optimizations and a smoother developer experience when working with complex autotuning and device management scenarios.