Worked on enhancing the stability and correctness of Triton-based Inductor code paths within the pytorch/pytorch repository, focusing on GPU programming and kernel development using Python. Addressed a persistent reduction handling issue in combo kernel paths during sorting operations, aligning the logic with non-combo paths to prevent assertion errors and improve reliability. The solution involved mirroring reduction logic across both single-node and multi-node execution, including dynamic guards for runtime conditions. This targeted bug fix improved the maintainability and predictability of Inductor-Triton integration, reducing production failures related to sorting with Triton code generation and strengthening the overall testing process.