Worked on the IBM/vllm repository to enhance the reliability and performance of distributed training systems. Focused on backend development and algorithm optimization, the main contribution involved fixing a calculation error in the dcp_local_seq_lens logic, which previously led to inaccurate sequence length distribution across computing nodes. By addressing this bug using Python and distributed computing techniques, the update ensured more accurate per-node sequencing, reducing training variance and improving throughput. This targeted fix strengthened multi-node model stability and increased the reliability of distributed training pipelines, directly addressing potential divergences and supporting more consistent performance in large-scale machine learning workflows.