Method gradient

tinygrad/tensor.py:827–850

Computes the gradient of the targets with respect to self. ```python exec="true" source="above" session="tensor" result="python" x = Tensor.eye(3) y = Tensor([[2.0,0,-2.0]]) z = y.matmul(x).sum() dx, dy = z.gradient(x, y) print(dx.tolist()) # dz/dx print(dy.tolist

(self, *targets:Tensor, gradient:Tensor|None=None)

  # ***** toposort and backward pass *****

  def gradient(self, *targets:Tensor, gradient:Tensor|None=None) -> list[Tensor]:
    """
    Computes the gradient of the targets with respect to self.

    ```python exec="true" source="above" session="tensor" result="python"
    x = Tensor.eye(3)
    y = Tensor([[2.0,0,-2.0]])
    z = y.matmul(x).sum()
    dx, dy = z.gradient(x, y)

    print(dx.tolist())  # dz/dx
    print(dy.tolist())  # dz/dy
    ```
    """
    assert gradient is not None or self.shape == tuple(), "when no gradient is provided, backward must be called on a scalar tensor"
    if not (self.is_floating_point() and all(t.is_floating_point() for t in targets)): raise RuntimeError("only float Tensors have gradient")
    if gradient is None: gradient = Tensor(1.0, dtype=self.dtype, device=self.device)
    target_uops = [x.uop for x in targets]
    grads = compute_gradient(self.uop, gradient.uop, set(target_uops))
    ret:list[Tensor] = []
    for x in target_uops:
      if (y:=grads.get(x)) is None: y = x.const_like(0)
      ret.append(Tensor(y))
    return ret

  def backward(self, gradient:Tensor|None=None) -> Tensor:
    """