Worked on the modelscope/ms-swift repository to enhance conversational AI capabilities by addressing multi-turn inference challenges. Using Python and deep learning techniques, resolved a bug that caused repeated inferences in the multi-turn scheduler and improved server-mode processing to streamline user interactions. Introduced a loss mask mechanism for multi-turn reasoning, enabling the model to compute loss at each intermediate step and facilitating more effective learning across reasoning chains. Focused on stabilizing the inference pipeline for both client and server deployments, which reduced edge-case failures and improved latency. Maintained clear documentation and responded to review feedback to support future collaboration and maintainability.