Developed ASCEND NPU training support for Qwen3-8B and Qwen3-14B models in the volcengine/verl repository, focusing on scalable large-model training workflows. Leveraged Direct Alignment Policy Optimization (DAPO) to enable efficient model training on ASCEND hardware, and created shell scripts to automate job configuration, including model paths, data inputs, and hardware-specific performance settings. Documented the setup and parameterization process to enhance reproducibility and streamline onboarding for future contributors. Utilized deep learning and machine learning expertise, along with shell scripting and RST for documentation, to deliver a robust, hardware-optimized training pipeline that prepares the team for further scalable iteration.