Worked on integrating LoRA support for Cosmos Predict 2.5 within the huggingface/diffusers repository, focusing on aligning the implementation with the official Cosmos repository. Developed an end-to-end workflow that included dataset preprocessing, model training, and evaluation scripts, all implemented in Python and leveraging deep learning and machine learning techniques. Addressed technical challenges such as VAE encoding inconsistencies, text encoder attention, and scheduler device placement, while ensuring compatibility with batch sizes greater than one and upcasting to fp32. Documented dependencies and prepared blog post assets to support communication, resulting in a scalable and maintainable fine-tuning pipeline for Cosmos Predict 2.5.