// GradientTape which uses ones.
        Gradient* zero = input_tensors[target_index].ZerosLike();
        new_zeros.push_back(zero);
        in_grads.push_back(zero);
      } else {
        in_grads.push_back(nullptr);
      }
    } else {
      in_grads.push_back(current_grad->second);
    }
  }

  // Avoid infinite recursion. Whichever forward function we run, it'll end up

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<TensorHandlePtr> vector_handles;
  vector_handles.reserve(2);
  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =
      ParallelTensor::FromTensorHandles(
          parallel_device, std::move(vector_handles), status.get());

    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();
      status->status = TF_TensorToTensor(input_tensors[i], &input_tensor);
      if (!status->status.ok()) return;
      input_tensors_vector[i] = &input_tensor;
    }
  }

        return result;
      }
      components.emplace_back(TFE_TensorHandleCopySharingTensor(
          absl::get<TFE_TensorHandle*>(inputs[i]), status));
    }
    std::vector<MaybeParallelTensorOwned> result_content;
    result_content.reserve(1);
    result_content.push_back(ParallelTensor::FromTensorHandles(
        parallel_device, std::move(components), status));

    components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {
      components.push_back(std::move(per_device_output_tensors[j][i]));
    }
    if (expected_output_shapes[i].IsFullyDefined()) {
      per_device_outputs.push_back(ParallelTensor::FromTensorHandles(
          *this, std::move(components),
          absl::Span<const int64_t>(expected_output_shapes[i].dim_sizes()),

    for (int i = 0; i < noutputs; ++i) {
      output_names_vec.push_back(string(output_names[i]));
    }
  }

  // Process control output names.
  std::vector<string> control_output_names_vec;
  if (control_output_names) {
    control_output_names_vec.reserve(ncontrol_outputs);
    for (int i = 0; i < ncontrol_outputs; ++i) {
      control_output_names_vec.push_back(string(control_output_names[i]));
    }
  }

                                        TF_Status* status);

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>

  const resp = await github.rest.issues.create({
    owner,
    repo,
    assignees: assignee_logins,
    title: `Issue created for Rollback of PR #${pr_number}: ${pr_title}`,
    body: `Merged PR #${pr_number} is rolled back in ${rollback_commit}.
    Please follow up with the reviewer and close this issue once its resolved.`
  });
  return `Issue created: ${resp.data.number} with Title: ${resp.data.title}`;

  std::vector<TFE_TensorHandle*> tensors() const {
    std::vector<TFE_TensorHandle*> result;
    result.reserve(tensors_.size());
    for (const TensorHandlePtr& tensor : tensors_) {
      result.emplace_back(tensor.get());
    }
    return result;
  }

 private:
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors,

      Multiply(context.get(), second_combined_value.get(),
               second_negative_one.get(), status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Un-pack the parallel tensor to verify that the operation was
  // successful. The resulting structure should be:
  //   second_device{first_device{1. * 3., 2. * 3.}, 3. * 3.}.
  std::array<TensorHandlePtr, 2> second_components;