}
    if (TF_GetCode(status_.get()) != TF_OK) {
      TF_SetStatus(status, TF_GetCode(status_.get()),
                   TF_Message(status_.get()));
      // Reset the member `status_` so future op executions (after recovery from
      // the bad `status`) start with an OK status.
      TF_SetStatus(status_.get(), TF_OK, "");
    }
    cancellation_manager_ = nullptr;

  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future
  // `StartExecute`s will deadlock).
  //
  // If `is_async=false` (constructor argument), `cancellation_manager` must
  // live until `Join` finishes. If `is_async=true` it must live until `Join` is

          "Operation '", op.node.DebugString(), "' was changed by ",
          mutation_type,
          " after it was run by a session. This mutation will have no effect, "
          "and will trigger an error in the future. Either don't modify "
          "nodes after running them or create a new session.");
    }
  }
}

namespace {

// Helper method that creates a shape handle for a shape described by dims.

  //
  // Unlike GradientTape::RecordOperation, Accumulate runs gradient computation
  // immediately. It stores the results, which feed into Accumulate for future
  // operations and may be fetched by calling FetchJVP. ForwardAccumulator
  // maintains a reference to these JVPs: if an `output_tensors` Tensor is
  // deleted, `DeleteGradient` should be called as soon as possible to free the

// message) to `output_graph_def` (allocated by TF_NewBuffer()).
// `output_graph_def`'s underlying buffer will be freed when TF_DeleteBuffer()
// is called.
//
// May fail on very large graphs in the future.
TF_CAPI_EXPORT extern void TF_GraphToGraphDef(TF_Graph* graph,
                                              TF_Buffer* output_graph_def,
                                              TF_Status* status);