TF_GUARDED_BY(mu);

  // The keys of this map are all the active sessions using this graph. Each
  // value records whether the graph has been mutated since the corresponding
  // session has been run (this is detected in RecordMutation function). If the
  // string is empty, no mutation has occurred. Otherwise the string is a
  // description of the mutation suitable for returning to the user.
  //

                                   bool heavy_load_on_streaming_rpc,
                                   bool remote_func_outputs = false,
                                   bool has_packed_input = false);

package(
    # copybara:uncomment default_applicable_licenses = ["//tensorflow:license"],
    licenses = ["notice"],
)

# Currently pybind extension shared objects must use only C API headers since
# the C API has static initializers duplicated in the Python bindings. So we
# need a second rule that omits .cc files, in
# tensorflow/python:_pywrap_parallel_device.
filegroup(
    name = "lib_headers",
    srcs = ["parallel_device_lib.h"],

      immediate_execution_ctx_.get(), {X.get(), Y.get()}, UseFunction()));
}

TEST_P(CppGradients, TestBiasAddGrad) {
  if (UseFunction() && UseMlir()) {
    GTEST_SKIP() << "SetAttrString has not been implemented yet.\n";
  }

  // A
  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f};
  int64_t A_dims[] = {2, 2};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;

# Build TensorFlow v2
build --define=tf_api_version=2 --action_env=TF2_BEHAVIOR=1

# Prevent double-compilation of some TF code, ref. b/183279666 (internal)
# > TF's gen_api_init_files has a genrule to run the core TensorFlow code
# > on the host machine. If we don't have --distinct_host_configuration=false,
# > the core TensorFlow code will be built once for the host and once for the
# > target platform.

  };
  const uint32 block_count = 256;
  tf_gcs_filesystem::RamFileBlockCache cache(
      block_size, block_count * block_size, 0, fetcher);
  std::vector<char> out;
  out.resize(block_count, 0);
  // The cache has space for `block_count` blocks. The loop with i = 0 should
  // fill the cache, and the loop with i = 1 should be all cache hits. The
  // fetcher checks that it is called once and only once for each offset (to

  const int num_devices = TF_DeviceListCount(devices);
  bool has_gpu0 = false;
  bool has_gpu1 = false;
  for (int i = 0; i < num_devices; ++i) {
    const char* dev = TF_DeviceListName(devices, i, status.get());
    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
    string device_name(dev);
    if (device_name.find("GPU:0") != string::npos) {
      has_gpu0 = true;
    }

  EXPECT_EQ(attr_value.i(), 2);

  // Placeholder oper now has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{feed, 0}));
  TF_Input feed_port;
  EXPECT_EQ(1, TF_OperationOutputConsumers(TF_Output{feed, 0}, &feed_port, 1));
  EXPECT_EQ(add, feed_port.oper);
  EXPECT_EQ(0, feed_port.index);

  // The scalar const oper also has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{three, 0}));

// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleBackingDeviceName(

                                       bool remote_func_outputs = false,
                                       bool has_packed_input = false) {
  return TestRemoteExecuteSilentCopies(async, remote, /*func=*/true,
                                       heavy_load_on_streaming_rpc,
                                       remote_func_outputs, has_packed_input);
}

TEST(CAPI, RemoteExecuteSilentCopiesAsyncFunc) {