Worked on performance optimization of the LTX2 processing pipeline in the huggingface/diffusers repository, focusing on improving throughput and GPU resource utilization. Used Python and PyTorch to streamline tensor handling by constructing tensors directly on the GPU, reducing synchronization overhead and minimizing CPU-to-GPU data transfers. Refactored padding logic with vectorized masking to enhance efficiency and updated scheduler logic to eliminate unnecessary data transfer. Emphasized code quality and maintainability through style improvements and documentation. The work addressed deep learning pipeline bottlenecks, resulting in faster inference and more efficient GPU programming, while maintaining a clean and maintainable codebase throughout the process.