Contributed to the sglang repositories by enabling Torch compilation in the NPU backend, introducing a feature flag and updating backend components to support optimized model execution. This work, implemented in Python with PyTorch and deep learning techniques, established a foundation for improved inference performance and future backend optimizations. Additionally, addressed a critical memory allocation issue in NPU memory management under PD separation, ensuring correct handling of key-value item lengths when feature flags are enabled. The approach emphasized reliability and maintainability, with collaborative development and peer review, resulting in more robust NPU workflows and reduced risk in production environments for backend systems.