Element-wise cube root — returns with a fully wired autograd node.
Defined for all real inputs (positive, zero, and negative); the negative branch is the real cube root, not the complex principal value.
Math
\frac{\partial y_i}{\partial x_i} = \frac{1}{3 y_i^{2}}$$ **Shape** Output shape equals input shape (elementwise). **See Also** `CubeRootBackward`.Parameters
aTensorImplPtrInput tensor (promoted to FP32 by the op schema).
Returns
TensorImplPtrOutput tensor with the same shape as a; gradient is .
Notes
CPU implementation uses Accelerate's vvcbrtf; GPU uses MLX's
power op with exponent . The gradient is singular at zero —
avoid backpropagating through inputs that touch the origin.
Examples
cube_root_op(-8.0) == -2.0 (real cube root of a negative number).
cube_root_op(27.0) == 3.0.