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

  TFE_Context* ctx = TFE_NewContext(opts, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  TFE_TensorHandle* h = TestScalarTensorHandle(ctx, 1.0f);
  TFE_TensorDebugInfo* debug_info = TFE_TensorHandleTensorDebugInfo(h, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(0, TFE_TensorDebugInfoOnDeviceNumDims(debug_info));

                  .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",

  Session* session;
  status->status = NewSession(opt->options, &session);
  if (status->status.ok()) {
    return new TF_DeprecatedSession({session});
  } else {
    DCHECK_EQ(nullptr, session);
    return nullptr;
  }
}

void TF_CloseDeprecatedSession(TF_DeprecatedSession* s, TF_Status* status) {
  status->status = s->session->Close();
}

    status_ = StatusFromTF_Status(status.get());
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();

    {
      AbstractContext* ctx_raw = nullptr;
      status_ =
          BuildImmediateExecutionContext(std::get<1>(GetParam()), &ctx_raw);
      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

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

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  auto* retval_task0 =
      TFE_TensorHandleCopyToDevice(retvals[0], ctx, local_device_name, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

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

  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  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]);

  auto s = GetValue(t, &analytical_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  int64_t num_elem_analytical = 1;
  auto num_dims_analytical = TF_NumDims(analytical_tensor);
  ASSERT_EQ(dims.size(), num_dims_analytical);
  for (int j = 0; j < num_dims_analytical; j++) {
    auto dim_analytical = TF_Dim(analytical_tensor, j);
    ASSERT_EQ(dims[j], dim_analytical);
    num_elem_analytical *= dim_analytical;
  }

    CHECK_EQ(errors::OK, s.code()) << s.message();
  }
};

class PassThroughGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    CHECK_EQ(grad_outputs.size(), 1);
    CHECK_EQ(grad_inputs.size(), 1);

  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  auto outputs = parallel_device.Execute(
      context.get(), {reduced_values.get(), run_collective.get()},
      "AssertAndCollective", TFE_OpGetAttrs(call_op.get()),
      /*expected_max_outputs=*/1, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  ASSERT_EQ(outputs->size(), 1);