}

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};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;

  TestRemoteExecuteUpdateServerDefResourceAccess(true);
}

void TestRemoteExecuteUpdateServerDefWithFailures(bool async) {
  // Fail fast on GetStatus requests so we can get errors instead of timeout
  // when updating cluster with non-exsitent worker
  tensorflow::setenv("GRPC_FAIL_FAST", "TRUE", /*overwrite=*/1);

  tensorflow::ServerDef server_def = GetServerDef(2);
  // This server def has the task index set to 0.

// function. Otherwise return nullptr in result and error Status from the
// function.
//
// If `library_filename` has already been loaded, we return a cached handle.
// Device and Kernels/Ops are registered as globals when a library is loaded
// for the first time.
TF_Library* TF_LoadPluggableDeviceLibrary(const char* library_filename,
                                          TF_Status* status) {

#include "tensorflow/core/platform/logging.h"

namespace tensorflow {

namespace {

// Managing context for the DLManagedTensor, will manage the lifetime of
// DLManagedTensor. When calling DLManagedTensor::deleter, it will notify the
// original framework of destruction, and this context will be deleted also.
struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;

                      &FloatDeallocator, nullptr);
}

// All the *Helper methods are used as a workaround for the restrictions that
// one cannot call ASSERT_* methods in non-void-returning functions (when
// exceptions are disabled during compilation)
void PlaceholderHelper(TF_Graph* graph, TF_Status* s, const char* name,
                       TF_DataType dtype, const std::vector<int64_t>& dims,

  // Now load a block of a different file "b" at timestamp `now` + 1
  env->SetNowSeconds(now + 1);
  TF_EXPECT_OK(ReadCache(&cache, "b", 0, 1, &out));
  // Now load a different block of file "a" at timestamp `now` + 1. When the
  // first block of "a" expires, this block should also be removed because it
  // also belongs to file "a".
  TF_EXPECT_OK(ReadCache(&cache, "a", 8, 1, &out));

using tensorflow::TF_StatusPtr;

namespace tensorflow {
namespace {

// The tests are parameterized on:
// - a string representing the tracing implementation: "mlir" or "graphdef".
// - a boolean that when true enables TFRT as the execution engine.
class UnifiedCAPI
    : public ::testing::TestWithParam<std::tuple<const char*, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());

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

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is
  // placed on task1, Handle types are LOCAL, REMOTE, LOCAL on task1.
  int num_replicas = 3;

      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such
        // nodes must have a single output.
        if (node->num_outputs() != 1) {
          return InvalidArgument(
              "When `num_opers` is set to -1, nodes referenced in `inputs` "
              "must have a single output. Node ",

    std::unique_ptr<ParallelTensor> t, TF_Status* status) {
  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;