Worked on distributed training stability for the bytedance-iaas/sglang repository, focusing on improving tensor parallelism rank management. Addressed a bug in the initialize_dp_attention function by updating the method for retrieving local_rank, ensuring it is accurately derived from the tp_group object rather than from tp_rank or tp_size. This change, implemented in Python, prevented misrouted attention across processes and enhanced training reliability in multi-node distributed systems. The fix reduced debugging time and improved maintainability for large-scale training jobs, contributing to more consistent model convergence and operational scalability in environments leveraging distributed systems and advanced tensor parallelism techniques.