Developed end-to-end W4A4 MXFP4 quantization support for Ascend hardware within the vllm-ascend repository, enabling efficient quantized inference for large models with Mixture of Experts (MoE) components. The work involved implementing new dynamic quantization methods and updating core inference operations to support Microscaling FP4 quantization, ensuring compatibility with the main vLLM release. Leveraging Python, PyTorch, and NPU programming, the developer integrated MXFP4 quantization into the MoE runtime, stage parameters, and token dispatching logic. This feature provided a complete quantization path, improving deployment performance and aligning the repository with vLLM v0.18.0 for seamless hardware support.