Yangming worked on the vllm-project/vllm-ascend repository, delivering support for the Qwen3.5 Mixture-of-Experts (MoE) model on Ascend devices. He implemented quantization configuration using Python, integrating ModelSlim to optimize model inference and throughput. His work included a Triton kernel fix that addressed operator precedence and memory safety in fused_gdn_gating, preventing out-of-bounds access and improving backend reliability. Yangming provided CI validation guidance to ensure robust deployment of Qwen3.5 MoE configurations. This engineering effort focused on backend enablement, leveraging deep learning and model optimization skills to unlock efficient, memory-safe MoE inference for production workloads on Ascend hardware.