std::vector<char> out;
  TF_EXPECT_OK(ReadCache(&cache1, want_filename, want_offset, want_n, &out));
  EXPECT_EQ(calls, 1);
  TF_EXPECT_OK(ReadCache(&cache2, want_filename, want_offset, want_n, &out));
  EXPECT_EQ(calls, 2);
  TF_EXPECT_OK(ReadCache(&cache3, want_filename, want_offset, want_n, &out));
  EXPECT_EQ(calls, 3);
  TF_EXPECT_OK(ReadCache(&cache4, want_filename, want_offset, want_n, &out));
  EXPECT_EQ(calls, 4);
}

        "--quiet", "--wait", "--nocache", \
        "--add", "Microsoft.VisualStudio.Workload.VCTools", \
        "--add", "Microsoft.VisualStudio.Component.VC.Tools.x86.x64", \
        "--add", "Microsoft.VisualStudio.Component.Windows10SDK.19041" -Wait; \
    Start-Process -FilePath C:/TEMP/vs_buildtools.exe -ArgumentList "--installPath", "C:/VS", \
        "--quiet", "--wait", "--nocache", "--includeOptional", \

  StatusPtr first_bad_status(nullptr);

  for (const auto& dt : device_threads_) {
    StatusPtr async_wait_status(TF_NewStatus());
    dt->AsyncWait(async_wait_status.get());
    // Prefer non cancelled errors to uncover real failures.
    if (TF_GetCode(async_wait_status.get()) != TF_OK &&
        (first_bad_status == nullptr ||
         TF_GetCode(first_bad_status.get()) == TF_CANCELLED)) {

  //
  // 1. Create a variable on `remote_device`, using `ctx_0`.
  TFE_TensorHandle* handle_0 =
      CreateVariable(ctx_0, 1.2, remote_device, /*variable_name=*/"var");

  // 2. Wait for `var` to be created and initialized on the worker.
  TF_Status* status = TF_NewStatus();
  TFE_ContextAsyncWait(ctx_0, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);

    uint64_t timestamp;
    /// Mutex to guard state variable
    absl::Mutex mu;
    /// The state of the block.
    FetchState state ABSL_GUARDED_BY(mu) = FetchState::CREATED;
    /// Wait on cond_var if state is FETCHING.
    absl::CondVar cond_var;
  };

  /// \brief The block map type for the file block cache.
  ///
  /// The block map is an ordered map from Key to Block.

  // functions (and hence the tensors they keep alive). Instead, everything
  // is deleted in ~GradientTape. Persistent GradientTapes are useful when
  // users want to compute multiple gradients over the same tape.
  explicit GradientTape(bool persistent) : persistent_(persistent) {}
  ~GradientTape() {
    for (const auto& pair : op_tape_) {

// `shape_inference::InferenceContext` constructor. Note the following:
//   - The inputs of the `op` are not used for shape inference. So, it is
//     OK to not have the inputs properly set in `op`. See `input_tensors`
//     if you want shape inference to consider the input tensors of the
//     op for shape inference.
//   - The types need not be set in `input_shapes` as it is not used.