Worked on the kvcache-ai/sglang repository to enhance model weight loading by implementing a synchronization mechanism that disables multi-thread loading during device tensor weight updates. This approach, developed using Python and leveraging skills in concurrency and model optimization, aimed to reduce synchronization overhead and stabilize latency in the weight update path. By ensuring that weight loading occurs synchronously, the solution improved throughput consistency and enabled more predictable runtime behavior in production environments. The work focused on deep learning model infrastructure, addressing the need for deterministic weight updates and contributing to smoother deployment processes without introducing additional bugs or regressions.