Fork自 PaddlePaddle / Paddle
Fix bug when user set CUDA_VISIBLE_DEVICES be empty and run CPU-only models