Contributed to the nv-auto-deploy/TensorRT-LLM repository by developing no-cache attention support within the PyTorch workflow, focusing on enhancing flexibility for large-model inference. This work involved refactoring the attention logic to accommodate diverse mask types and key-value cache interactions, ensuring compatibility with various deployment scenarios. Leveraging skills in C++, Python, and CUDA, the implementation included comprehensive updates to documentation and test coverage to promote maintainability and robustness. By enabling cache-free attention paths, the changes streamlined integration with existing deployment pipelines and improved the reliability of attention mechanisms in production environments, reflecting a thoughtful approach to scalable model deployment challenges.