immediate_execution_ctx_.get(), {x.get()}, UseFunction()));
}

TEST_P(CppGradients, TestMatMulGrad) {
  // TODO(vnvo2409): Figure out why `gradient_checker` does not work very
  // well with `MatMul` and remove `TestMatMul*` in
  // `mnist_gradients_test` when done.
  GTEST_SKIP();

  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f};
  int64_t A_dims[] = {3, 3};

  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
  // moment; once canceled, the collective executor requires the program to be
  // restarted / context to be reset.
}

TEST(PARALLEL_DEVICE_LIB, TestDifferentShapes) {

                             serialized_update.size(), status);
  EXPECT_NE(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Even after the prevoiusly failed cluster update, another update and op
  // execution should work fine as long as the provided server_def is valid.
  TFE_ContextUpdateServerDef(ctx, 0, serialized.data(), serialized.size(),
                             status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  tensorflow::mutex execution_mutex_;
  ExecutionState execution_state_ TF_GUARDED_BY(execution_mutex_) =
      ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.
  tensorflow::condition_variable finished_join_;

      ReluModel, ReluGradModel, immediate_execution_ctx_.get(), {X.get()},
      UseFunction()));

  // Mathematically, Relu isn't differentiable at `0`. So `gradient_checker`
  // does not work with it.
  AbstractTensorHandlePtr Y;
  {
    AbstractTensorHandle* Y_raw;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 0.0f, &Y_raw);

#define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)
#define EXPECT_TF_OK(x) EXPECT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)

static const char* content = "abcdefghijklmnopqrstuvwxyz1234567890";
// We will work with content_view instead of content.
static const absl::string_view content_view = content;

namespace gcs = google::cloud::storage;

static std::string InitializeTmpDir() {

    forward_op_.outputs.push_back(retvals[i]);
  }
  // TODO(b/166669239): This is needed to support AttrBuilder::Get for string
  // attributes. Number type attrs and DataType attrs work fine without this.
  // Consider getting rid of this and making the behavior between number types
  // and string consistent.
  forward_op_.attrs.BuildNodeDef();
  // TODO(b/170307493): Populate skip_input_indices here.

}

namespace {

struct SomeThreadData {
  ::tensorflow::mutex mu;
  bool did_work = false;
};

void SomeThreadFunc(void* data) {
  auto* real_data = static_cast<SomeThreadData*>(data);
  ::tensorflow::mutex_lock l(real_data->mu);
  real_data->did_work = true;
}

}  // namespace

TEST(TestEnv, TestThreads) {
  TF_ThreadOptions options;

      // and hence expect the input to be forwarded to the return tensor.
      EXPECT_EQ(orig_ptr, TF_TensorData(t));
    } else {
      // In async mode we expect forwarding to work without releasing the input
      // handle since by the time the kernel is executed we have released the
      // handle in the client code.
      EXPECT_EQ(orig_ptr, TF_TensorData(t));
    }
  } else {

  options->stack_size = 0;
  options->guard_size = 0;
  options->numa_node = -1;
}

TF_Thread* TF_StartThread(const TF_ThreadOptions* options,
                          const char* thread_name, void (*work_func)(void*),
                          void* param) {
  ::tensorflow::ThreadOptions cc_options;
  cc_options.stack_size = options->stack_size;
  cc_options.guard_size = options->guard_size;