Worked on backend development and performance optimization for the volcengine/verl repository, focusing on resolving a performance bottleneck during GRPO rollout. Addressed an inefficiency by reducing return sequences from n to 1, which eliminated unnecessary repeated calculations and improved throughput in high-volume deployment scenarios. The solution was implemented in Python and delivered as a targeted, low-risk change, enhancing the reliability and scalability of the rollout process. By preventing CPU-intensive duplicate operations, the update ensured more predictable performance under load. This work demonstrated a methodical approach to identifying and fixing backend performance issues within a production environment using Python.