Worked on the unslothai/unsloth-zoo repository to enhance model merging workflows by improving device handling for LoRA and GGUF paths. Addressed cross-hardware compatibility by introducing the use of a backend device type and ensuring the active device index is preserved during GGUF merges. This approach reduced manual intervention and post-merge issues, resulting in faster and more reliable merges across different hardware environments. The work demonstrated practical application of PyTorch, Python, and data processing skills, with a focus on robust device management and merging algorithms. All changes were delivered through Git-based collaboration, emphasizing maintainability and workflow acceleration for machine learning projects.