Worked on the sglang repository to enhance the stability of the MiniMaxM2MoE model’s forward pass by addressing a critical bug in top-K expert selection. Using Python and PyTorch, contributed a targeted fix that corrected the misuse of the topk function, ensuring accurate expert routing and consistent output behavior during inference. Collaborated closely with peers to review and implement the solution, which reduced the risk of misrouting and improved the reliability of mixture-of-experts outputs. This work strengthened the dependability of downstream predictions in production machine learning scenarios, reflecting a focused approach to deep learning model maintenance and collaborative engineering.