🐛 Bug

If there is an exception raised in the backward operation of custom function, PyTorch always raised RuntimeError and does not preserve the exception type.

To Reproduce

import torch

class MyFunc(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        return x

    @staticmethod
    def backward(ctx, gradx):
        raise TypeError("this should be TypeError")

a = torch.tensor(2.).requires_grad_()
a1 = MyFunc.apply(a)
print(torch.autograd.grad(a1, a))

The output:

Traceback (most recent call last):
  File "test.py", line 14, in <module>
    print(torch.autograd.grad(a1, a))
  File "/home/mfkasim/anaconda3/lib/python3.7/site-packages/torch/autograd/__init__.py", line 192, in grad
    inputs, allow_unused)
RuntimeError: this should be TypeError

It becomes RuntimeError despite it's specified as TypeError

Expected behavior

The exception type should be conserved.

Environment

• PyTorch Version (e.g., 1.0): 1.6.0+cpu
• OS (e.g., Linux): Ubuntu
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.7.4
• CUDA/cuDNN version: N/A
• GPU models and configuration: N/A
• Any other relevant information:

Additional context

cc @ezyang @albanD @zou3519 @gqchen @pearu @nikitaved