// 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

      TF_DeleteTensor);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_EQ(111., *static_cast<float*>(TF_TensorData(resolved_value.get())));

  // Free the backing buffer for the variable.
  op.reset(TFE_NewOp(context.get(), "DestroyResourceOp", status.get()));
  TFE_OpAddInput(op.get(), var_handle, status.get());
  TFE_OpSetDevice(op.get(), name, status.get());

// TFE_ExecutorWaitForAllPendingNodes before calling this API if you want to
// make sure all nodes are finished.
TF_CAPI_EXPORT extern void TFE_DeleteExecutor(TFE_Executor*);

// Returns true if the executor is in async mode.
TF_CAPI_EXPORT extern bool TFE_ExecutorIsAsync(TFE_Executor*);

// Causes the calling thread to block till all ops dispatched in this executor

#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;

// Delete NewFunctionLibraryDefinition.
TF_CAPI_EXPORT extern void TF_DeleteFunctionLibraryDefinition(
    TF_FunctionLibraryDefinition* fn_lib);

// Shorthand for calling LookUp to get the OpDef from FunctionLibraryDefinition
// given op name. The returned OpDef is represented by TF_Buffer.
TF_CAPI_EXPORT extern void TF_LookUpOpDef(TF_FunctionLibraryDefinition* fn_lib,

    ASSERT_TRUE(remote_arg->HasLocalMirror(nullptr));
  }

  if (remote_func_outputs) {
    const string backing_device =
        TFE_TensorHandleBackingDeviceName(retvals[0], status);
    ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
    EXPECT_EQ(backing_device, task2_name);
  }

  auto* retval_task0 = TFE_TensorHandleCopyToDevice(