Developed configurable multi-precision training support for the NVIDIA/TransformerEngine repository, enabling flexible selection among FP32, FP16, FP8, MXFP8, and NVFP4 formats within the FSDP script. Leveraged Python and PyTorch to implement robust argument parsing, precise dtype and recipe handling, and clear flag precedence, allowing users to tailor training precision to their needs. Enhanced logging and runtime configuration messages improved transparency and user guidance, while careful initialization across precision branches reduced runtime errors in distributed deep learning environments. This work facilitates faster experimentation with low-precision training, optimizing resource utilization without compromising accuracy or reproducibility in machine learning workflows.