During April 2026, Fenglin contributed to the vllm-ascend repository by developing end-to-end W4A4 MXFP4 quantization support for Ascend hardware. He implemented core quantization features, including new dynamic linear and fused MoE methods, to enable Microscaling FP4 quantization in large models with MoE components. His work involved updating NPU-specific grouped matrix multiplication operations and integrating parameterized quantization types into the MoE runtime, ensuring compatibility with the main vLLM release. Using Python, PyTorch, and deep learning techniques, Fenglin delivered a robust quantization path that enhances inference performance and deployment flexibility for models running on Ascend devices.