The conv weight grad from loss.backward has nans, but it only has infs when compute manually

autograd
3

The operation is Conv2d in stem stage in resnet.
The test steps was posted here