const DLTensor* dltensor_out = &dlm_out->dl_tensor;
  EXPECT_EQ(dltensor_out->device.device_type, dltensor_in->device.device_type);
  EXPECT_EQ(dltensor_out->device.device_id, dltensor_in->device.device_id);
  EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);
  EXPECT_EQ(dltensor_out->dtype.bits, dltensor_in->dtype.bits);
  EXPECT_EQ(dltensor_out->dtype.lanes, dltensor_in->dtype.lanes);

  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(2, TFE_TensorDebugInfoOnDeviceNumDims(debug_info));
  // Shape is the same for CPU tensors.
  EXPECT_EQ(3, TFE_TensorDebugInfoOnDeviceDim(debug_info, 0));
  EXPECT_EQ(2, TFE_TensorDebugInfoOnDeviceDim(debug_info, 1));

  TFE_DeleteTensorDebugInfo(debug_info);
  TFE_DeleteTensorHandle(h);
  TFE_DeleteContext(ctx);
  TF_DeleteStatus(status);

                  .ok());
  ASSERT_TRUE(worker_server2->Start().ok());

  TFE_ContextSetServerDef(context.get(), 0, serialized.data(),
                          serialized.size(), status.get());
  EXPECT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  BasicTestsForTwoDevices(context.get(),
                          "/job:worker/replica:0/task:1/device:CPU:0",

void TF_DeleteOutputList(TF_OutputList* o) { delete unwrap(o); }
void TF_OutputListSetNumOutputs(TF_OutputList* o, int num_outputs,
                                TF_Status* s) {
  unwrap(o)->expected_num_outputs = num_outputs;
  unwrap(o)->outputs.clear();
  unwrap(o)->outputs.resize(num_outputs);
}
int TF_OutputListNumOutputs(TF_OutputList* o) {
  return unwrap(o)->outputs.size();
}

    const auto tensor_interface =
        tensorflow::down_cast<const tensorflow::TensorInterface*>(src->tensor);

    if (dst->dims() != 0) {
      return InvalidArgument(
          "Malformed TF_RESOURCE tensor: expected a scalar, got a tensor with "
          "shape ",
          dst->shape().DebugString());
    }
    *dst = tensorflow::Tensor(tensorflow::DT_RESOURCE, dst->shape());

  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TF_Tensor* t = TFE_TensorHandleResolve(handle, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  std::unique_ptr<float[]> actual_values(new float[expected_values.size()]);
  EXPECT_EQ(sizeof(float) * expected_values.size(), TF_TensorByteSize(t));

                          "AssertAndCollective", TFE_OpGetAttrs(call_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  EXPECT_NE(TF_GetCode(status.get()), TF_CANCELLED);
  EXPECT_EQ(TF_GetCode(status.get()), TF_INVALID_ARGUMENT);
  EXPECT_THAT(TF_Message(status.get()), HasSubstr("assertion failed"));

  // Note that future collectives with the same context do not work at the

  float product[4] = {0};
  EXPECT_EQ(sizeof(product), TF_TensorByteSize(t));
  memcpy(&product[0], TF_TensorData(t), TF_TensorByteSize(t));
  TF_DeleteTensor(t);
  EXPECT_EQ(7, product[0]);
  EXPECT_EQ(10, product[1]);
  EXPECT_EQ(15, product[2]);
  EXPECT_EQ(22, product[3]);

  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(h1_task0);
  TFE_DeleteTensorHandle(h0_task1);

  TF_Tensor* result_tensor;
  s = GetValue(outputs[0], &result_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto result_value = static_cast<float*>(TF_TensorData(result_tensor));
  EXPECT_EQ(*result_value, 1.0);
  outputs[0]->Unref();
  TF_DeleteTensor(result_tensor);
  result_tensor = nullptr;
}

INSTANTIATE_TEST_SUITE_P(
    CustomGradientTest, CustomGradientTest,

  const auto& shape = reader->GetVariableToShapeMap().at(name);
  int rank = shape.dims();
  if (num_dims != rank) {
    status->status = InvalidArgument("Expected rank is ", num_dims,
                                     " but actual rank is ", rank);
    return;
  }
  for (int i = 0; i < num_dims; i++) {
    dims[i] = shape.dim_size(i);
  }
}