output_nr issue when requires_grad=False

test.py

from torch.library import Library

test_ns = "abc"
lib = Library(test_ns, "FRAGMENT")
lib.define("foo(Tensor(a!) a, Tensor(b!) b) -> (Tensor(a!), Tensor(b!))")

def get_op(name):
    return getattr(getattr(torch.ops, test_ns), name).default

op = get_op("foo")

def foo_impl(a, b):
    return a, b

lib.impl("foo", foo_impl, "CPU")
x = torch.randn(3, requires_grad=False)
y = torch.randn(3, requires_grad=False)

z, w = op(x, y)

print(z.grad_fn, y.grad_fn)
print(z.output_nr, w.output_nr)  # 0, 0

out = (z * 2 + w * 3).sum()
dz, dw = torch.autograd.grad(out, (z, w))

print(dz, dw)  # tensor([5., 5., 5.]) tensor([5., 5., 5.])
# expect tensor([2., 2., 2.]) tensor([3., 3., 3.])