Worked on improving the stability of the vllm-project/vllm-gaudi repository by addressing a critical bug in the FusedMoE initialization process. Focused on aligning the HpuMiniMaxM2MoE path with upstream VLLM by removing the reduce_results kwarg, which previously caused worker startup failures on the HPU port. Maintained the MoE output reduction logic to follow TP/EP topology as defined upstream, ensuring correct behavior across different configurations. This approach reduced divergence from upstream code, minimized maintenance risk, and improved compatibility. Utilized Python and deep learning frameworks, demonstrating attention to upstream collaboration, clear commit practices, and a methodical approach to machine learning infrastructure.