Developed and integrated a hardware-accelerated video decoding backend for the jeejeelee/vllm repository, leveraging NVIDIA DeepStream and CUDA to enable efficient video processing on NVIDIA GPUs. The work involved incorporating the nvidia-deepstream-videodecode-cu13 library into the existing video pipeline, extending the loading flow to support the new backend, and implementing a shared thread pool to optimize decoding throughput and memory transfers between GPU and CPU. Using Python and GStreamer, the developer established infrastructure for scalable, high-throughput video workloads, addressing performance bottlenecks and laying the groundwork for future DeepStream enhancements without introducing new bugs during the development period.