Wanghuanjun focused on backend reliability in the vllm-project/vllm-ascend repository, addressing a critical bug affecting Multi-Token Prediction (MTP) models. Using Python and leveraging machine learning expertise, Wanghuanjun corrected the layer count retrieval logic to ensure accurate resource allocation for draft MTP models, preventing both under- and over-allocation during speculative decoding. The solution integrated with the model_arch_config_convertor infrastructure, supporting DeepSeek-V3 MTP and Qwen3.5 MTP variants and aligning with upstream vLLM core practices. This work improved deployment stability and resource estimation, demonstrating careful attention to model-specific requirements and collaborative, maintainable engineering in a production backend environment.