ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    Y.reset(Y_raw);
  }

  std::vector<AbstractTensorHandle*> outputs(1);
  status_ = RunModel(ReluGradModel, immediate_execution_ctx_.get(), {Y.get()},
                     absl::MakeSpan(outputs), UseFunction());
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  ASSERT_TF_OK(status_);
  ASSERT_EQ(bucket, "bucket");
  ASSERT_EQ(object, "path/to/object");

  ParseGCSPath("gs://bucket/", true, &bucket, &object, status_);
  ASSERT_TF_OK(status_);
  ASSERT_EQ(bucket, "bucket");

  ParseGCSPath("bucket/path/to/object", false, &bucket, &object, status_);
  ASSERT_EQ(TF_GetCode(status_), TF_INVALID_ARGUMENT);

      TFE_NewOp(context.get(), function_name, status.get()), TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  TFE_OpSetDevice(op.get(), device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  TFE_OpAddInput(op.get(), parallel_value.get(), status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  TFE_Execute(func, &retvals[0], &num_retvals, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  ASSERT_EQ(1, num_retvals);
  TF_Tensor* t = TFE_TensorHandleResolve(retvals[0], status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteTensorHandle(retvals[0]);
  float sum = 0;
  ASSERT_EQ(sizeof(sum), TF_TensorByteSize(t));
  memcpy(&sum, TF_TensorData(t), TF_TensorByteSize(t));

                   /*raw_device_name=*/nullptr, &forward_op);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  if (isa<TracingOperation>(check_numerics_op.get())) {
    s = dyn_cast<TracingOperation>(check_numerics_op.get())
            ->SetOpName("check_numerics");
    ASSERT_EQ(errors::OK, s.code()) << s.message();
  }
  s = AddInput(check_numerics_op.get(), t.get(), &forward_op);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

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

  // Create a variable handle (uninitialized to start) placed on the parallel
  // device.
  std::function<void(Variable*)> variable_deleter = [&](Variable* to_delete) {
    to_delete->Destroy(context, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    ASSERT_EQ(errors::OK, s.code()) << s.message();
    s = dyn_cast<tracing::TracingContext>(ctx.get())->AddParameter(
        DT_FLOAT, shape, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    x.reset(x_raw);
  }

  PartialTensorShape shape;
  Status s = x->Shape(&shape);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  ASSERT_FALSE(shape.unknown_rank());

    delete[] output_names_ptr;
    if (expect_failure) {
      ASSERT_EQ(func_, nullptr);
      return;
    }

    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
    ASSERT_NE(func_, nullptr);
    ASSERT_EQ(std::string(func_name_), std::string(TF_FunctionName(func_)));
    TF_GraphCopyFunction(host_graph_, func_, nullptr, s_);
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
  }

  std::unique_ptr<tensorflow::GrpcServer> worker_server1;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(
                  server_def, tensorflow::Env::Default(), &worker_server1)
                  .ok());
  ASSERT_TRUE(worker_server1->Start().ok());

  server_def.set_task_index(2);
  std::unique_ptr<tensorflow::GrpcServer> worker_server2;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(

                          &numerical_grad_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    numerical_grad.reset(numerical_grad_raw);
  }

  TF_Tensor* numerical_tensor;
  s = GetValue(numerical_grad.get(), &numerical_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);
  ASSERT_EQ(num_elem_numerical, num_grad);