I defined a custom DeviceContext (like the one in torch/utils/_device.py) based on torch_function. I find that the context will intercept calling of all torch apis, which decrease the performance of training especially in model with multiple small ops like InceptionV3.
Is there a way to limit the interception？