Worked on deep learning infrastructure in the ROCm/vllm and jeejeelee/vllm repositories, focusing on both code clarity and performance optimization. Refactored the image input handling for Blip2ForConditionalGeneration by introducing a dedicated parsing and validation method, which improved maintainability and reduced runtime errors in image-driven workflows. Addressed object creation issues to further stabilize the model’s input pipeline. Later, optimized the Vision Transformer (ViT) attention mechanism by removing a synchronization point in the torch sdpa attention backend, resulting in better throughput and lower memory usage. Leveraged Python, PyTorch, and deep learning techniques to deliver targeted, maintainable improvements to model workflows.