Torch mlp support on my local macbook outperforms CUDA T4 on Colab.