How to recycle the gradient tensor when calling backward()

1

Hi folks,

I am wondering if there is a good way to recycle the gradient tensor when calling backward(). Here is an example.

import torch
x = torch.rand(4,4)
w = torch.nn.Linear(4,4)
with torch.set_grad_enabled(True):
    y = w(x)
    dy = torch.rand(4,4)
    y.backward(dy)

I am hoping to recycle dy in the middle of y.backward() as soon as it is no longer needed for the chain rule. How can I do that?

Thanks!