Developed and optimized sparse attention compression features for the DeepSeek V4 model in the jeejeelee/vllm repository, focusing on both functionality and performance. Leveraged CUDA, CuTe, and Python to implement new compression kernels and wrappers, enabling efficient sparse attention pathways and improving model scalability. Refactored kernel structures and introduced a specialized C128 block-8 kernel, consolidating compression, normalization, and rope-store operations into a unified pipeline to reduce memory traffic and enhance thread-level data movement. Maintained high standards of code integration and review, ensuring reliable deployment and establishing a robust foundation for further deep learning performance optimizations within the repository.