Fix running inference on Bottom-up models with CUDA

Fix running inference on Bottom-up models with CUDA #819