TF_SetAttrString(reduce_desc, "final_op", final_op.c_str(),
                   final_op.length());
  TF_SetAttrIntList(reduce_desc, "subdiv_offsets", nullptr, 0);
  TF_AddInput(reduce_desc, x);
  TF_Operation* reduce_op = TF_FinishOperation(reduce_desc, status);
  if (TF_GetCode(status) != TF_OK) return;
  TF_Operation* operations[]{placeholder_op, reduce_op};
  TF_Output y{reduce_op, 0};
  const char* output_name = "y";

  TF_DeleteTensor(model_out_tensor);

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();
    return absl::OkStatus();
  }

  // Else, reduce sum the output to get a scalar

  // Will sum all dimensions, so get a Tensor containing [0,...,num_dims_out-1].
  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,

      TFE_NewOp(context.get(), "AssertAndCollective", status.get()),
      TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::unique_ptr<ParallelTensor> reduced_values =
      parallel_device.ScalarsFromSequence<float>({1.0, 2.0}, context.get(),
                                                 status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  num_retvals = 1;
  executed = false;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  EXPECT_FALSE(executed);
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);

  // Custom devices can refuse to do type-based dispatch (as hcustom1 is
  // configured to do)
  matmul.reset(MatMulOp(context.get(), hcustom1.get(), hcpu.get()));
  num_retvals = 1;
  executed = false;