Worked on performance optimization for the vllm-project/vllm-ascend repository, focusing on fusing AddRMSNorm and RMSNorm operations with DynamicMxQuant to accelerate model execution on A5 hardware. Developed two fusion patterns at the compiler level, leveraging deep learning and quantization expertise to reduce operator latency while maintaining model accuracy. The implementation was validated across multiple test matrices using Python and integrated with CANN and FrameworkPTAdapter toolchains. This work emphasized hardware-aware design and cross-repository collaboration, resulting in measurable improvements in execution time and stability. No major bugs were addressed, as the primary focus remained on enhancing performance and reliability through fusion passes.