trtvideo

trtvideo is a specialized runtime designed for headless, file-to-file video processing using TensorRT models. It optimizes the video pipeline by ensuring raw video frames remain in VRAM throughout the entire process—utilizing NVDEC for decoding, CV-CUDA for preprocessing and postprocessing, TensorRT for inference, and NVENC for encoding. This architecture minimizes host-to-device memory transfers, resulting in lower CPU usage and reduced peak VRAM compared to traditional pipelines. Key features include: - GPU-resident path: Integrated pipeline from compressed input to muxed output. - Docker-based deployment: Provided as production and model-tool images for reproducible execution. - Model compatibility: Validated support for super-resolution models like RealESRGAN_x2plus and SPAN. - Tooling: Includes utilities for exporting PyTorch checkpoints to ONNX, preparing static TensorRT engines, and performing compatibility checks. - Performance: Measured throughput advantages in end-to-end FPS and resource efficiency on NVIDIA RTX 3090 and 4090 GPUs.