Worked on the vllm-project/vllm-ascend repository to optimize graph replay synchronization within the backend. Focused on reducing unnecessary overhead by gating synchronization to occur only when the graph mode is set to FULL, this approach improved replay performance and stability, particularly under mixed-mode operations. Leveraged Python for backend development and performance optimization, implementing targeted changes that resulted in a measurable reduction in synchronization latency. Validated all modifications against the vLLM v0.13.0 baseline and main branch to ensure compatibility and minimal regression. Enhanced code clarity and maintainability, providing concise commit messages and improving traceability throughout the synchronization path.