Worked on enhancing the finetuning workflow for large language models in the meta-llama/llama-recipes repository by integrating QLoRA quantization with Fully Sharded Data Parallel (FSDP) training. Focused on improving parameter handling and configuration flow, the work simplified the logic for managing original model parameters under FSDP and Parameter-Efficient Fine-Tuning (PEFT) settings. Adjustments to quantization configurations and logging increased robustness and traceability when combining QLoRA with FSDP. Using Python and deep learning frameworks, the contribution improved the scalability and reliability of distributed model finetuning, addressing edge-case complexities and supporting more flexible experimentation with quantization and distributed training techniques.