Worked on enhancing memory efficiency in the GraniteSpeechConformerAttention module within the jeejeelee/vllm repository. Addressed a memory bottleneck by replacing a resource-intensive operation with an optimized einsum-based approach, leveraging deep learning and PyTorch expertise. This refactoring reduced peak memory usage for the granite_speech component, enabling support for larger input sequences and improved batching without additional memory overhead. The work focused on production scalability and deployment efficiency, demonstrating a strong understanding of performance optimization in machine learning systems. Contributed through careful code review and targeted feature development, with all changes implemented in Python and aligned with best practices for maintainable code.