Order of trainable/non-trainable affecting NDArray gradient

[theRoughCode]

I have the following custom autograd function (hugely simplified to demonstrate this error):

class foo(mx.autograd.Function):
    name = "foo"

    def forward(self, *input):
        self.save_for_backward(*input)
        return mx.nd.array([1] * len(input))

    def backward(self, dy):
        return [mx.nd.array([1]) for _ in range(len(self.saved_tensors))]

If I run the following

a = mx.nd.array([0.1])
a.attach_grad()
c = mx.nd.array([0.545])
func = foo()

with autograd.record():
    m = func(a, c)
    m.backward()

I get a.grad = [1]. However, if I swap the order of the parameters (i.e. m = func(c, a)), I get a.grad = [0].

Am I not using the custom autograd interface correctly or is there a bug within MXNet?