//
// This call may not block for execution of ops enqueued concurrently with this
// call.
TF_CAPI_EXPORT extern void TFE_ExecutorWaitForAllPendingNodes(
    TFE_Executor*, TF_Status* status);

// When an error happens, any pending operations are discarded, and newly issued
// ops return an error. This call clears the error state and re-enables
// execution of newly issued ops.
//

  auto tf_dlm_type = GetDlDataType(data_type, status);
  if (!status->status.ok()) {
    return nullptr;
  }

  TensorReference tensor_ref(*tensor);  // This will call buf_->Ref()
  auto* tf_dlm_tensor_ctx = new TfDlManagedTensorCtx(tensor_ref);
  tf_dlm_tensor_ctx->reference = tensor_ref;

  DLManagedTensor* dlm_tensor = &tf_dlm_tensor_ctx->tensor;

      output_gradients, result, /*build_default_zeros_grads*/ false));
  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.
// These APIs are mainly to facilitate testing and are subject to change.
namespace internal {
Status Reset(AbstractOperation* op_, const char* op,

}

// Note: this function looks up a thread local policy. So it should be called in
// the appropriate client thread. In particular, in async mode, it may not be
// safe to call this function from the async EagerExecutor threads.
extern TFE_ContextDevicePlacementPolicy TFE_ContextGetDevicePlacementPolicy(
    TFE_Context* ctx) {
  return static_cast<TFE_ContextDevicePlacementPolicy>(

      new GradientTape<Gradient, BackwardFunction, TapeTensor>(false));
  AccumulatorCallState& call_state = call_state_.top();
  call_state.backward_tape = tape.get();
  auto pop_backward_tape =
      gtl::MakeCleanup([&call_state] { call_state.backward_tape = nullptr; });
  std::vector<Gradient*> forwardprop_aids;
  std::vector<int64_t> sources;
  std::unordered_set<int64_t> sources_set;

  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name) {
  return new TF_AttrBuilder(op_name);
}

    if (!enabled_) {
      return absl::OkStatus();
    }
    if (first_call_) {
      first_call_ = false;
      return absl::OkStatus();
    }
    return tensorflow::errors::Internal("Graph pass runs for more than once!");
  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect

  //   Outputs:
  std::vector<TensorHandlePtr> op_outputs_ TF_GUARDED_BY(execution_mutex_);
  // TF_Status is an incomplete type and so can't be stack allocated. To avoid
  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);