Worked on the PaddlePaddle/FastDeploy repository to deliver XPU-specific optimizations for Multi-Turn Processing (MTP) workloads. Developed custom kernels using C++ and CUDA to enhance draft model preprocessing, postprocessing, updating, and speculation-related operations within the XPU environment. This engineering effort focused on improving throughput and reducing latency for MTP tasks, aligning with broader acceleration goals for XPU-enabled deployments. The implementation included new mechanisms for managing hidden states, output, and penalties, laying a technical foundation for future performance gains. All changes were traceable through upstream commits, ensuring maintainability and integration with the existing deep learning inference pipeline.