Developed CPU-based speculative decoding support for draft models in the jeejeelee/vllm repository, enabling accelerated inference on CPU and reducing reliance on GPU resources for select workloads. The implementation integrated seamlessly with the existing draft-model inference path, ensuring compatibility and maintainability across CPU workflows. Leveraging C++, Python, and PyTorch, the work focused on optimizing CPU programming for deep learning and machine learning applications. Adhered to repository contribution standards, including commit-level attribution and metadata requirements, while laying the foundation for future CPU optimizations. This feature broadened deployment options and improved performance flexibility for projects requiring efficient CPU inference capabilities.