Worked on the bytedance-iaas/vllm repository to stabilize distributed FusedMoE parallelism by resolving a critical bug that previously prevented expert parallelism from activating as intended. Addressed the issue by ensuring expert parallelism is enabled only when the product of tensor parallelism and data parallelism sizes exceeds one, which improved both model execution efficiency and hardware utilization in mixture-of-experts workloads. Utilized Python and PyTorch to debug distributed systems and tune performance in a production-like deep learning environment. The work demonstrated a methodical approach to diagnosing and fixing complex parallelism issues, with careful assessment of the impact on distributed model training.