class

CubeBackward

CubeBackward

Autograd node for the element-wise cube $y = x^3$ .

Saves the input x so the backward pass can build $3x^2$ in grad_formula and multiply it against the upstream gradient.

Math

\frac{\partial y}{\partial x} = 3x^2, \qquad \frac{\partial \mathcal{L}}{\partial x} = 3x^2\,\frac{\partial \mathcal{L}}{\partial y}.$$

schema_v1OpSchema

Registered as "cube" with AmpPolicy::Promote.

Dispatch: dedicated Accelerate kernel (CPU) / MLX element-wise cube (GPU).

Static methods

static

int dispatch(backend::IBackend & be, const int & a, const int & s, Dtype dt)

Forward — calls IBackend::cube to compute $y = x^3$ .

int grad_formula(const int & g)

Backward — $\partial y/\partial x = 3x^2$ , scaled by grad_out.