Function Grad

oneflow/api/python/autograd/autograd.cpp:117–132

}

Maybe<one::TensorTuple> Grad(const one::TensorTuple& outputs, const one::TensorTuple& inputs,
                             const one::TensorTuple& out_grads, bool retain_graph,
                             bool create_graph, bool allow_unused, bool is_grads_batched) {
  PythonFrameGuard pf;
  BackwardPassScopeGuard backward_guard;
  if (create_graph) { retain_graph = true; }
  if (inputs.empty()) { return Backward(outputs, out_grads, retain_graph, create_graph); }
  CHECK_OR_RETURN(std::all_of(
      inputs.begin(), inputs.end(),
      [](const std::shared_ptr<one::Tensor>& tensor) { return tensor->requires_grad(); }))
      << "All input tensors `.requires_grad` should be true";
  std::shared_ptr<one::TensorTuple> gradients =
      JUST(CheckAndInitOutGrads(outputs, out_grads, is_grads_batched));
  return one::GetThreadLocalAutogradEngine()->RunBackwardAndReturnInputsTensorGradIf(
      outputs, inputs, *gradients, retain_graph, create_graph, allow_unused);
}

namespace py = pybind11;