// 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
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback

                       const char* dst, TF_Status* status) {
  // If target is a directory return TF_FAILED_PRECONDITION.
  // Target might be missing, so don't error in that case.
  struct stat st;
  if (stat(dst, &st) != 0) {
    if (errno != ENOENT) {
      TF_SetStatusFromIOError(status, errno, dst);
      return;
    }
  } else if (S_ISDIR(st.st_mode)) {

  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;
  for (int i = 0; i < target_tensor_ids.size(); ++i) {
    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];
      sources_that_are_targets.insert(
          std::make_pair(target_id, TapeTensor(tensor)));
    }
  }

  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;
    // Note that dimensions with size=1 can have any stride.
    if (shape_arr[i] != 1 && stride_arr[i] != expected_stride) {
      valid = false;
    }
    expected_stride *= shape_arr[i];
  }
  return valid;
}

==============================================================================*/
// This file extends/implements core graph optimizer base classes in terms of
// the C API defined in grappler.h. A class "CSomething" represents a
// "Something" that can be manipulated via calls in the C interface and a C
// struct called "TP_Something".

#include "tensorflow/c/experimental/grappler/grappler.h"

#include <algorithm>
#include <cstddef>
#include <cstring>

      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_;

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

                                    fn_name, "'");
    output_tensors->emplace_back(node, idx);
  }
  return absl::OkStatus();
}

// Populates `body_nodes` with the nodes that will become function's body.
// Performs various checks.
Status ComputeBodyNodes(
    const TF_Graph* fn_body, const char* fn_name, int num_opers,
    const TF_Operation* const* opers,

namespace gcs = google::cloud::storage;

// The environment variable that overrides the block size for aligned reads from
// GCS. Specified in MB (e.g. "16" = 16 x 1024 x 1024 = 16777216 bytes).
constexpr char kBlockSize[] = "GCS_READ_CACHE_BLOCK_SIZE_MB";
constexpr size_t kDefaultBlockSize = 64 * 1024 * 1024;
// The environment variable that overrides the max size of the LRU cache of
// blocks read from GCS. Specified in MB.

  (*sess_options.config.mutable_device_count())["CPU"] = num_cpus;
  status->status =
      tensorflow::DeviceFactory::AddCpuDevices(sess_options, prefix, &devices);

  // Remove the device that has the host device name since host device is alreay
  // in an initialized context.
  for (auto d = devices.begin(); d != devices.end();) {
    if (absl::StrContains(d->get()->name(), "CPU:0")) {
      d = devices.erase(d);