During July 2025, Zhihe Wang developed ASCEND NPU training support for Qwen3-8B and Qwen3-14B models in the volcengine/verl repository, focusing on scalable large-model training. Leveraging deep learning and model training expertise, Wang implemented Direct Alignment Policy Optimization (DAPO) to optimize performance on ASCEND hardware. He created shell scripts to automate training job configuration, including model paths, data inputs, and hardware-specific parameters, using Shell and RST for scripting and documentation. By validating the workflow and documenting the setup, Wang improved reproducibility and onboarding, delivering a robust foundation for future model iteration and efficient NPU-accelerated machine learning development.