During March 2026, Bei Wang contributed to the vllm-project/vllm-ascend repository by delivering targeted reliability and performance improvements for vLLM-Ascend integration. Wang addressed a tensor size compatibility issue in the model runner, resolving runtime errors between fc1 and non-single-padding configurations. Leveraging deep learning and model optimization expertise, Wang optimized the DeepSeekOCR2 model’s RelPosAttention and CustomQwen2Decoder components using Python, which reduced inference latency and improved runtime stability. Comprehensive documentation updates were also provided to streamline deployment and evaluation. This work demonstrated a strong grasp of machine learning principles and contributed to more robust and maintainable model deployment workflows.