Worked on the kvcache-ai/sglang repository to deliver DeepSeek v3.2 Inference Performance Optimization, focusing on improving model inference for long sequences. Implemented context parallelism for the prefill stage, which increased throughput and reduced latency, making the system more efficient for production workloads. The approach involved parallel context processing and careful performance profiling to ensure stability and maintainability. All changes were made in Python using PyTorch, leveraging deep learning and model optimization techniques. No major bugs were reported or fixed during this period, reflecting a targeted and stable optimization effort that preserved compatibility and minimized the code surface area.