Looks like this uses mutation of global/shared state. For the example:
z=x*y+3,
what if there is another function that does:
w=x+2*y
and then both functions do a backward pass (simultaneously, perhaps in different threads or otherwise); then it seems dangerous to collect the results of the backward pass (partial derivatives) in the shared variables x and y and make them accessible through x.get_grad() and y.get_grad(). Imho, in a better design, you'd say z.get_grad(x) and z.get_grad(y), and w.get_grad(x) and w.get_grad(y) to get the partial derivatives.
z=x*y+3,
what if there is another function that does:
w=x+2*y
and then both functions do a backward pass (simultaneously, perhaps in different threads or otherwise); then it seems dangerous to collect the results of the backward pass (partial derivatives) in the shared variables x and y and make them accessible through x.get_grad() and y.get_grad(). Imho, in a better design, you'd say z.get_grad(x) and z.get_grad(y), and w.get_grad(x) and w.get_grad(y) to get the partial derivatives.