if (!s.ok()) {
    return nullptr;
  }
  s = op->AddInputList(gradient_tensors);
  if (!s.ok()) {
    return nullptr;
  }

  int num_outputs = 1;
  std::vector<AbstractTensorHandle*> outputs(num_outputs);
  s = op->Execute(absl::Span<AbstractTensorHandle*>(outputs), &num_outputs);
  if (!s.ok()) {
    return nullptr;
  }
  return outputs[0];
}

  const tensorflow::Tensor* tensor;
  status->status = handle->Tensor(&tensor);
  if (!status->status.ok()) {
    return nullptr;
  }

  std::vector<int64_t> dev_dims = TensorShapeAsVector(*handle, &status->status);
  if (!status->status.ok()) {
    return nullptr;
  }
  return new TFE_TensorDebugInfo(dev_dims);
}

TF_CAPI_EXPORT extern void TFE_DeleteTensorDebugInfo(

  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);
    int64_t vals_shape[] = {num_dims_out};
    Range(&vals, 0, num_dims_out);
    AbstractTensorHandle* sum_dims_raw = nullptr;
    TF_RETURN_IF_ERROR(TestTensorHandleWithDims<int32_t, TF_INT32>(
        ctx, vals.data(), vals_shape, 1, &sum_dims_raw));
    sum_dims.reset(sum_dims_raw);
  }

  // Reduce sum the output on all dimensions.

    const char* device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "GPU:0") != nullptr);

    TFE_OpSetDevice(matmul, "CPU:0", status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "CPU:0") != nullptr);
  }

  TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);
  TFE_DeleteContext(ctx);

  TFE_TensorHandle* var_handle0 = TestVariable(ctx, 1.0, dev0_name);
  EXPECT_NE(var_handle0, nullptr);
  TFE_TensorHandle* var_handle1 = TestVariable(ctx, 2.0, dev1_name);
  EXPECT_NE(var_handle1, nullptr);

  TFE_TensorHandle* value_handle = nullptr;
  ReadVariable(ctx, var_handle1, &value_handle);
  CheckTFE_TensorHandleHasFloats(value_handle, {2});
  TFE_DeleteTensorHandle(value_handle);

      TFE_NewOp(context, "CollectiveReduce", status), TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  const char* device = TFE_TensorHandleDeviceName(input, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op.get(), "T", TFE_TensorHandleDataType(input));

  TF_Output input{arg, 0};
  TF_Output output{id, 0};
  TF_Function* fn =
      TF_GraphToFunction(function_graph, "ident", 0, 1, &id, 1, &input, 1,
                         &output, nullptr, nullptr, "test", status);
  ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
  TF_DeleteGraph(function_graph);
  TFE_ContextOptions* opts = TFE_NewContextOptions();

  AbstractTensorHandlePtr x;
  {
    AbstractTensorHandle* x_raw = nullptr;
    absl::Status s =
        TestScalarTensorHandle<float, TF_FLOAT>(ctx_.get(), 2.0f, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    x.reset(x_raw);
  }

  AbstractTensorHandlePtr y;
  {
    AbstractTensorHandle* y_raw = nullptr;
    absl::Status s =

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  auto* arg0 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  auto* arg1 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a first "Add" computing `arg0 + arg1`.
  TF_AbstractTensor* add_output1;
  {

  std::string summary;
  absl::Status cpp_status = parallel_tensor->SummarizeValue(summary);
  if (!cpp_status.ok()) {
    tsl::Set_TF_Status_from_Status(status, cpp_status);
    return nullptr;
  }
  return TF_NewBufferFromString(summary.data(), summary.size());
}

TensorHandlePtr ParallelTensorToTensorHandle(
    const std::string& parallel_device_name, TFE_Context* context,