Developed and integrated GLM-4.6V model inference on Ascend NPU for the yhyang201/sglang repository, focusing on expanding on-device deployment capabilities. The work involved implementing preprocessing adjustments for both image and video data, ensuring compatibility with the NPU’s requirements and optimizing the inference pipeline. Leveraging Python and deep learning frameworks, the solution enabled efficient machine learning model execution directly on specialized hardware. The integration required careful handling of image processing and video processing workflows, as well as close attention to NPU-specific constraints. This feature enhanced the repository’s flexibility for deploying GLM workloads across diverse hardware environments without reported bugs.