Worked on the huggingface/optimum-habana repository to improve the reliability of Torch Compile workflows for Habana hardware. Addressed a critical configuration issue by implementing an environment guard in Python that asserts PT_HPU_LAZY_MODE is set to 0 whenever torch_compile is enabled. This early-detection mechanism prevents misconfigurations that could otherwise degrade performance or cause runtime errors, directly reducing support overhead for Habana deployments. Focused on environment configuration and performance optimization, the solution was delivered as a targeted bug fix rather than a new feature, demonstrating attention to deployment stability and user experience in specialized machine learning hardware environments.