Worked on enhancing the PaddlePaddle ecosystem by implementing tensor parallelism for QLoRA in the PaddleFormers repository, enabling more scalable and efficient model training. This involved updating quantization layers and coordinating changes across trainer and conversion utilities to support the new parallelization strategy. Additionally, expanded fine-tuning capabilities in the ERNIE repository by adding nf4 compute type support, allowing broader quantization options for model optimization. Leveraged deep learning expertise and full stack development skills, primarily using Python and YAML, to ensure compatibility and stability. The work facilitated larger-scale experiments and improved hardware utilization while maintaining ease of use for end users.