During August 2025, Setugang developed two core features across PaddlePaddle’s PaddleFormers and ERNIE repositories, focusing on deep learning and model optimization. In PaddleFormers, Setugang introduced tensor parallelism for QLoRA, updating quantization layers, trainer logic, and conversion utilities to enable scalable training and improved hardware utilization. For ERNIE, Setugang expanded fine-tuning capabilities by adding nf4 compute type support, integrating it as a quantization algorithm within FinetuningArguments. These contributions, implemented primarily in Python and YAML, addressed the need for efficient large-scale model training and flexible quantization strategies, reflecting a strong grasp of LLM optimization and full stack development principles.