if (n == 0) {
    TF_SetStatus(status, TF_OK, "");
    return 0;
  }
  if (!IsCacheEnabled() || (n > max_bytes_)) {
    // The cache is effectively disabled, so we pass the read through to the
    // fetcher without breaking it up into blocks.
    return block_fetcher_(filename, offset, n, buffer, status);
  }
  // Calculate the block-aligned start and end of the read.

      pruning_thread_.reset(
          TF_StartThread(&thread_options, "TF_prune_FBC", PruneThread, this));
    }
    TF_VLog(1, "GCS file block cache is %s.\n",
            (IsCacheEnabled() ? "enabled" : "disabled"));
  }

  ~RamFileBlockCache() {
    if (pruning_thread_) {
      stop_pruning_thread_.Notify();
      // Destroying pruning_thread_ will block until Prune() receives the above
      // notification and returns.

    ```bash
    bazel test ${flags} //tensorflow/python/saved_model:load_test --test_filter=*LoadTest.test_capture_variables*
    ```

    **Note:** You can add `--test_sharding_strategy=disabled` to the `flags` to
    disable the sharding so that all the test outputs are in one file. However,
    it may slow down the tests for not running in parallel and may cause the

}

// All the *Helper methods are used as a workaround for the restrictions that
// one cannot call ASSERT_* methods in non-void-returning functions (when
// exceptions are disabled during compilation)
void PlaceholderHelper(TF_Graph* graph, TF_Status* s, const char* name,
                       TF_DataType dtype, const std::vector<int64_t>& dims,
                       TF_Operation** op) {

  int device_id = TFE_TensorHandleDeviceID(hcpu, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  ASSERT_EQ(0, device_id) << device_id;

  // Disable the test if no GPU is present.
  string gpu_device_name;
  if (GetDeviceName(ctx, &gpu_device_name, "GPU")) {
    TFE_TensorHandle* hgpu = TFE_TensorHandleCopyToDevice(

// By default the entire file is reuploaded.
constexpr char kAppendMode[] = "GCS_APPEND_MODE";
// If GCS_APPEND_MODE=compose then instead the new data is uploaded to a
// temporary object and composed with the original object. This is disabled by
// default as the multiple API calls required add a risk of stranding temporary
// objects.
constexpr char kComposeAppend[] = "compose";

// We can cast `google::cloud::StatusCode` to `TF_Code` because they have the

TEST(CAPI, DistributedFunctionNoError) {
  TestDistributedFunctionCancellation(false);
}

// TODO(b/170399182): Update test once an alternative to using the function
// optimization hook is in place.
TEST(CAPI, DISABLED_DistributedFunctionCancelledOnError) {
  TestDistributedFunctionCancellation(true);
}

void TestRemoteExecuteDeleteContextWithOutstandingRPC(bool async) {
  tensorflow::ServerDef server_def = GetServerDef(2);

// Enables tracing of RunMetadata on the ops executed from this context.
TF_CAPI_EXPORT extern void TFE_ContextEnableRunMetadata(TFE_Context* ctx);

// Disables tracing of RunMetadata on the ops executed from this context.
TF_CAPI_EXPORT extern void TFE_ContextDisableRunMetadata(TFE_Context* ctx);

// Populates the passed-in buffer with a serialized RunMetadata protocol buffer

// Enables only graph collection in RunMetadata on the functions executed from
// this context.
TF_CAPI_EXPORT extern void TFE_ContextEnableGraphCollection(TFE_Context* ctx);

// Disables only graph collection in RunMetadata on the functions executed from
// this context.
TF_CAPI_EXPORT extern void TFE_ContextDisableGraphCollection(TFE_Context* ctx);

}

unsigned char TF_GetXlaConstantFoldingDisabled() {
  return static_cast<unsigned char>(
      tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding);
}

void TF_SetXlaConstantFoldingDisabled(unsigned char should_enable) {
  tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding =
      static_cast<bool>(should_enable);
}

void TF_SetXlaMinClusterSize(int size) {