What's the difference between autograd.backward and nd.backward()?

Two piece code will get different results, is there any one can give me some tips about the difference between autograd.backward() and mx.nd.backward().

First piece code:

cls_mask = mx.nd.one_hot(cls_id, 1000, on_value=1, off_value=0)
img.attach_grad()
with autograd.record():
       logit = net(img)
       logit = logit * cls_mask
autograd.backward(logit)
grad = img.grad
...

Second piece code:

cls_mask = mx.nd.one_hot(cid, 1000, on_value=1, off_value=0)
img.attach_grad()
with autograd.record():
       logit = net(img)
logit.backward(cls_mask)
grad = img.grad
...

i.e., my question is, why logit.backward(cls_mask) is different from autograd.backward(logit) ??

That’s grateful topic. It’s useful with me