Developed and integrated GRPO training script support for Qwen2.5-32B and Qwen3-30B models on the volcengine/verl repository, targeting the Ascend NPU platform. Leveraged bash scripting and automation to create an end-to-end training suite, incorporating both Megatron and vLLM backends to expand training capabilities. The work included adding usage examples and validating workflows with gms8k workloads, ensuring that the scripts enabled rapid experimentation and scalable fine-tuning of large models on NPU hardware. Updated API usage documentation and collaborated on deployment strategies, focusing on machine learning and NPU development to streamline model training and experimentation processes.