Vision transformers predict 9 times faster with PyTorch, ONNX and TensorRT