// ExtendSessionGraphHelper manually.
  std::atomic<bool> extend_before_run;
};

struct TF_ImportGraphDefOptions {
  tensorflow::ImportGraphDefOptions opts;

  // Backing memory for TensorId fields in opts.
  // TODO(skyewm): it'd be better if ImportGraphDefOptions owned this.
  std::vector<tensorflow::string> tensor_id_data;
};

struct TF_ImportGraphDefResults {

    // .backing_device of shape is CPU since the tensor is backed by CPU
    backing_device_name =
        TFE_TensorHandleBackingDeviceName(retvals[0], status.get());
    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
    ASSERT_TRUE(absl::StrContains(backing_device_name, "CPU:0"))
        << backing_device_name;

    TFE_DeleteOp(shape_op);

  // The cache should now be empty.
  EXPECT_EQ(cache.CacheSize(), 0);
}

TEST(RamFileBlockCacheTest, ParallelReads) {
  // This fetcher won't respond until either `callers` threads are calling it
  // concurrently (at which point it will respond with success to all callers),
  // or 10 seconds have elapsed (at which point it will respond with an error).
  const int callers = 4;
  BlockingCounter counter(callers);

TFCI_BAZEL_TARGET_SELECTING_CONFIG_PREFIX=linux_arm64
# Note: this is not set to "--cpu", because that changes the package name
# to tensorflow_cpu. These ARM builds are supposed to have the name "tensorflow"
# despite lacking Nvidia CUDA support.
TFCI_BUILD_PIP_PACKAGE_ARGS="--repo_env=WHEEL_NAME=tensorflow"
TFCI_DOCKER_ENABLE=1
TFCI_DOCKER_IMAGE=gcr.io/tensorflow-sigs/build-arm64:tf-2-16-multi-python
TFCI_DOCKER_PULL_ENABLE=1

// Some TF ops need a step container to be set to limit the lifetime of some
// resources (mostly TensorArray and Stack, used in while loop gradients in
// graph mode). Calling this on a context tells it to start a step.
TF_CAPI_EXPORT extern void TFE_ContextStartStep(TFE_Context* ctx);

// Ends a step. When there is no active step (that is, every started step has

//   materialized in the address space of the calling process.
// * int is used as an index into arrays.
// * Deletion functions are safe to call on nullptr.
//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation

  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {
    // Special-cased operation for packing per-device tensors into one parallel
    // tensor.
    if (inputs.size() != parallel_device.num_underlying_devices()) {
      std::string message(absl::StrCat(

               tensorflow::uint64 send_device_incarnation);

// Return a 1-D INT32 tensor containing a single value 1.
TFE_TensorHandle* TestAxisTensorHandle(TFE_Context* ctx);

// Return an op taking minimum of `input` long `axis` dimension.
TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis);

// If there is a device of type `device_type`, returns true

  static Variable* Create(TFE_Context* context, TF_DataType type,
                          const int64_t* dims, const int num_dims,
                          const char* device, TF_Status* status);
  // Dereferences the backing buffer for the variable. Note that since this can
  // fail (it runs operations), it must be called explicitly and the resulting
  // `status` checked.
  void Destroy(TFE_Context* context, TF_Status* status);

  if (TF_GetCode(status) != TF_OK) return nullptr;
  return new Variable(var_handle, type);
}

void Variable::Destroy(TFE_Context* context, TF_Status* status) {
  // Free the backing buffer for the variable.
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
      TFE_NewOp(context, "DestroyResourceOp", status), &TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return;