.realize()

Execute all pending forward computations in the computation graph.

Walks the graph in topological order (inputs before outputs) and executes any stored _forward_fn callables, filling in tensor.values for every unrealized tensor. After this call, self.values and the values of every upstream tensor are guaranteed to be non-None.

In eager mode this is a no-op — all tensors are already realized.

Returns:

• 'Tensor' –

  self, so you can chain: loss = model(x).realize().