Worked on the modelscope/ms-swift repository to deliver a reusable training pathway for the Qwen3 model on NPU hardware, focusing on distributed training efficiency and reproducibility. Developed an example Fully Sharded Data Parallel (FSDP) configuration, a standardized JSON config file for distributed parameters, and a shell-based launcher script to streamline training runs with parameterized options. This approach expanded hardware support for Qwen3, reduced setup time, and improved onboarding for distributed machine learning experiments. Demonstrated skills in NPU optimization, data parallelism, and model training, leveraging JSON and shell scripting to orchestrate and standardize the training process for enhanced resource efficiency.