(*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the

          mutation_type,
          " after it was run by a session. This mutation will have no effect, "
          "and will trigger an error in the future. Either don't modify "
          "nodes after running them or create a new session.");
    }
  }
}

namespace {

// Helper method that creates a shape handle for a shape described by dims.
tensorflow::shape_inference::ShapeHandle ShapeHandleFromDims(

                                        std::move(mixed_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> size_op(

                                    /*remote_func_outputs=*/true);
}
TEST(CAPI, RemoteExecuteSilentCopiesLocalAsyncFuncOrdering) {
  // A remote input may be not ready when we start running a function. Test that
  // the function execution should wait until the remote input is ready.
  TestRemoteExecuteSilentCopiesFunc(/*async=*/true, /*remote=*/false,

  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,
               std::vector<TFE_TensorHandle*> inputs,

  env->SetNowSeconds(now + 2);
  // The pruning thread periodically compares env->NowSeconds() with the oldest
  // block's timestamp to see if it should evict any files. At the current fake
  // timestamp of `now` + 2, file "a" is stale because its first block is stale,
  // but file "b" is not stale yet. Thus, once the pruning thread wakes up (in

  return true;
}

size_t RamFileBlockCache::CacheSize() const {
  absl::MutexLock lock(&mu_);
  return cache_size_;
}

void RamFileBlockCache::Prune() {
  while (!stop_pruning_thread_.WaitForNotificationWithTimeout(
      absl::Microseconds(1000000))) {
    absl::MutexLock lock(&mu_);
    uint64_t now = timer_seconds_();
    while (!lra_list_.empty()) {

    const TP_OptimizerConfigs tp_configs, const char* device_type) {
  ConfigList configs;
  // disable_model_pruning is turned off by default.
  if (tp_configs.disable_model_pruning == TF_TriState_On)
    configs.disable_model_pruning = true;
  else
    configs.disable_model_pruning = false;
  // The other configs are turned on by default.
  CONFIG_TOGGLE(implementation_selector);

  return nullptr;
}

// For TFE_CustomDevice::copy_tensor_from_device in the parallel device
// registration.
//
// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either

bool GraphErrorInjectionPass::enabled_ = false;

// Test to ensure that a registered graph optimization pass is only executed
// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops