.deferred()

Create an unrealized tensor that defers computation to .realize().

Used internally by _create_op_result when lazy mode is active. The tensor is a shell — values is None and shape is known — until .realize() walks the graph and calls forward_fn.

Parameters:

• forward_fn (Callable[[], ndarray]) –

  Zero-argument callable that returns a backend ndarray. Captures input tensors by reference, so it stays valid until .realize() populates their .values.

• shape (tuple) –

  Output shape. Available immediately without executing the op.

• dtype (str, default: 'float32' ) –

  Data type string. Defaults to "float32".

• device (str | None, default: None ) –

  Device string. Defaults to the current global default device.

Returns:

• 'Tensor' –

  An unrealized Tensor with values=None.